refactor(3d): Remove duplicated Hunyuan3D code and use dynamic loading

- Removes the vendored hunyuan3d_repo, hy3dpaint, and hy3dshape directories to rely on a single source of truth downloaded from Hugging Face.
- Modifies the 3D model generator to dynamically load the Hunyuan3D modules using importlib, making the integration more robust and maintainable.
- Adds a test script for verifying the 3D generation pipeline.

hy3dpaint/DifferentiableRenderer/MeshRender.py

@@ -1,1414 +0,0 @@
-# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT
-# except for the third-party components listed below.
-# Hunyuan 3D does not impose any additional limitations beyond what is outlined
-# in the repsective licenses of these third-party components.
-# Users must comply with all terms and conditions of original licenses of these third-party
-# components and must ensure that the usage of the third party components adheres to
-# all relevant laws and regulations.
-
-# For avoidance of doubts, Hunyuan 3D means the large language models and
-# their software and algorithms, including trained model weights, parameters (including
-# optimizer states), machine-learning model code, inference-enabling code, training-enabling code,
-# fine-tuning enabling code and other elements of the foregoing made publicly available
-# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT.
-
-import cv2
-import torch
-import trimesh
-import numpy as np
-from PIL import Image
-import torch.nn.functional as F
-from typing import Union, Optional, Tuple, List, Any, Callable
-from dataclasses import dataclass
-from enum import Enum
-from .camera_utils import (
-    transform_pos,
-    get_mv_matrix,
-    get_orthographic_projection_matrix,
-    get_perspective_projection_matrix,
-)
-
-try:
-    from .mesh_utils import load_mesh, save_mesh
-except:
-    print("Bpy IO CAN NOT BE Imported!!!")
-
-try:
-    from .mesh_inpaint_processor import meshVerticeInpaint  # , meshVerticeColor
-except:
-    print("InPaint Function CAN NOT BE Imported!!!")
-
-
-class RenderMode(Enum):
-    """Rendering mode enumeration."""
-    NORMAL = "normal"
-    POSITION = "position"
-    ALPHA = "alpha"
-    UV_POS = "uvpos"
-
-
-class ReturnType(Enum):
-    """Return type enumeration."""
-    TENSOR = "th"
-    NUMPY = "np"
-    PIL = "pl"
-
-
-class TextureType(Enum):
-    """Texture type enumeration."""
-    DIFFUSE = "diffuse"
-    METALLIC_ROUGHNESS = "mr"
-    NORMAL = "normal"
-
-
-@dataclass
-class RenderConfig:
-    """Unified rendering configuration."""
-    elev: float = 0
-    azim: float = 0
-    camera_distance: Optional[float] = None
-    center: Optional[List[float]] = None
-    resolution: Optional[Union[int, Tuple[int, int]]] = None
-    bg_color: List[float] = None
-    return_type: str = "th"
-    
-    def __post_init__(self):
-        if self.bg_color is None:
-            self.bg_color = [1, 1, 1]
-
-
-@dataclass
-class ViewState:
-    """Camera view state for rendering pipeline."""
-    proj_mat: torch.Tensor
-    mv_mat: torch.Tensor
-    pos_camera: torch.Tensor
-    pos_clip: torch.Tensor
-    resolution: Tuple[int, int]
-
-
-def stride_from_shape(shape):
-    """
-    Calculate stride values from a given shape for multi-dimensional indexing.
-    
-    Args:
-        shape: Tuple or list representing tensor dimensions
-        
-    Returns:
-        List of stride values for each dimension
-    """
-    stride = [1]
-    for x in reversed(shape[1:]):
-        stride.append(stride[-1] * x)
-    return list(reversed(stride))
-
-
-def scatter_add_nd_with_count(input, count, indices, values, weights=None):
-    """
-    Perform scatter-add operation on N-dimensional tensors with counting.
-    
-    Args:
-        input: Input tensor [..., C] with D dimensions + C channels
-        count: Count tensor [..., 1] with D dimensions  
-        indices: Index tensor [N, D] of type long
-        values: Value tensor [N, C] to scatter
-        weights: Optional weight tensor [N, C], defaults to ones if None
-        
-    Returns:
-        Tuple of (updated_input, updated_count) tensors
-    """
-    # input: [..., C], D dimension + C channel
-    # count: [..., 1], D dimension
-    # indices: [N, D], long
-    # values: [N, C]
-
-    D = indices.shape[-1]
-    C = input.shape[-1]
-    size = input.shape[:-1]
-    stride = stride_from_shape(size)
-
-    assert len(size) == D
-
-    input = input.view(-1, C)  # [HW, C]
-    count = count.view(-1, 1)
-
-    flatten_indices = (indices * torch.tensor(stride, dtype=torch.long, device=indices.device)).sum(-1)  # [N]
-
-    if weights is None:
-        weights = torch.ones_like(values[..., :1])
-
-    input.scatter_add_(0, flatten_indices.unsqueeze(1).repeat(1, C), values)
-    count.scatter_add_(0, flatten_indices.unsqueeze(1), weights)
-
-    return input.view(*size, C), count.view(*size, 1)
-
-
-def linear_grid_put_2d(H, W, coords, values, return_count=False):
-    """
-    Place values on a 2D grid using bilinear interpolation.
-    
-    Args:
-        H: Grid height
-        W: Grid width  
-        coords: Coordinate tensor [N, 2] with values in range [0, 1]
-        values: Value tensor [N, C] to place on grid
-        return_count: Whether to return count information
-        
-    Returns:
-        2D grid tensor [H, W, C] with interpolated values, optionally with count tensor
-    """
-    # coords: [N, 2], float in [0, 1]
-    # values: [N, C]
-
-    C = values.shape[-1]
-
-    indices = coords * torch.tensor([H - 1, W - 1], dtype=torch.float32, device=coords.device)
-    indices_00 = indices.floor().long()  # [N, 2]
-    indices_00[:, 0].clamp_(0, H - 2)
-    indices_00[:, 1].clamp_(0, W - 2)
-    indices_01 = indices_00 + torch.tensor([0, 1], dtype=torch.long, device=indices.device)
-    indices_10 = indices_00 + torch.tensor([1, 0], dtype=torch.long, device=indices.device)
-    indices_11 = indices_00 + torch.tensor([1, 1], dtype=torch.long, device=indices.device)
-
-    h = indices[..., 0] - indices_00[..., 0].float()
-    w = indices[..., 1] - indices_00[..., 1].float()
-    w_00 = (1 - h) * (1 - w)
-    w_01 = (1 - h) * w
-    w_10 = h * (1 - w)
-    w_11 = h * w
-
-    result = torch.zeros(H, W, C, device=values.device, dtype=values.dtype)  # [H, W, C]
-    count = torch.zeros(H, W, 1, device=values.device, dtype=values.dtype)  # [H, W, 1]
-    weights = torch.ones_like(values[..., :1])  # [N, 1]
-
-    result, count = scatter_add_nd_with_count(
-        result, count, indices_00, values * w_00.unsqueeze(1), weights * w_00.unsqueeze(1)
-    )
-    result, count = scatter_add_nd_with_count(
-        result, count, indices_01, values * w_01.unsqueeze(1), weights * w_01.unsqueeze(1)
-    )
-    result, count = scatter_add_nd_with_count(
-        result, count, indices_10, values * w_10.unsqueeze(1), weights * w_10.unsqueeze(1)
-    )
-    result, count = scatter_add_nd_with_count(
-        result, count, indices_11, values * w_11.unsqueeze(1), weights * w_11.unsqueeze(1)
-    )
-
-    if return_count:
-        return result, count
-
-    mask = count.squeeze(-1) > 0
-    result[mask] = result[mask] / count[mask].repeat(1, C)
-
-    return result
-
-
-def mipmap_linear_grid_put_2d(H, W, coords, values, min_resolution=128, return_count=False):
-    """
-    Place values on 2D grid using mipmap-based multiresolution interpolation to fill holes.
-    
-    Args:
-        H: Grid height
-        W: Grid width
-        coords: Coordinate tensor [N, 2] with values in range [0, 1] 
-        values: Value tensor [N, C] to place on grid
-        min_resolution: Minimum resolution for mipmap levels
-        return_count: Whether to return count information
-        
-    Returns:
-        2D grid tensor [H, W, C] with filled values, optionally with count tensor
-    """
-    # coords: [N, 2], float in [0, 1]
-    # values: [N, C]
-
-    C = values.shape[-1]
-
-    result = torch.zeros(H, W, C, device=values.device, dtype=values.dtype)  # [H, W, C]
-    count = torch.zeros(H, W, 1, device=values.device, dtype=values.dtype)  # [H, W, 1]
-
-    cur_H, cur_W = H, W
-
-    while min(cur_H, cur_W) > min_resolution:
-
-        # try to fill the holes
-        mask = count.squeeze(-1) == 0
-        if not mask.any():
-            break
-
-        cur_result, cur_count = linear_grid_put_2d(cur_H, cur_W, coords, values, return_count=True)
-        result[mask] = (
-            result[mask]
-            + F.interpolate(
-                cur_result.permute(2, 0, 1).unsqueeze(0).contiguous(), (H, W), mode="bilinear", align_corners=False
-            )
-            .squeeze(0)
-            .permute(1, 2, 0)
-            .contiguous()[mask]
-        )
-        count[mask] = (
-            count[mask]
-            + F.interpolate(cur_count.view(1, 1, cur_H, cur_W), (H, W), mode="bilinear", align_corners=False).view(
-                H, W, 1
-            )[mask]
-        )
-        cur_H //= 2
-        cur_W //= 2
-
-    if return_count:
-        return result, count
-
-    mask = count.squeeze(-1) > 0
-    result[mask] = result[mask] / count[mask].repeat(1, C)
-
-    return result
-
-
-# ============ Core utility functions for reducing duplication ============
-
-def _normalize_image_input(image: Union[np.ndarray, torch.Tensor, Image.Image]) -> Union[np.ndarray, torch.Tensor]:
-    """Normalize image input to consistent format."""
-    if isinstance(image, Image.Image):
-        return np.array(image) / 255.0
-    elif isinstance(image, torch.Tensor):
-        return image.cpu().numpy() if image.is_cuda else image
-    return image
-
-
-def _convert_texture_format(tex: Union[np.ndarray, torch.Tensor, Image.Image], 
-                          texture_size: Tuple[int, int], device: str, force_set: bool = False) -> torch.Tensor:
-    """Unified texture format conversion logic."""
-    if not force_set:
-        if isinstance(tex, np.ndarray):
-            tex = Image.fromarray((tex * 255).astype(np.uint8))
-        elif isinstance(tex, torch.Tensor):
-            tex_np = tex.cpu().numpy()
-            tex = Image.fromarray((tex_np * 255).astype(np.uint8))
-        
-        tex = tex.resize(texture_size).convert("RGB")
-        tex = np.array(tex) / 255.0
-        return torch.from_numpy(tex).to(device).float()
-    else:
-        if isinstance(tex, np.ndarray):
-            tex = torch.from_numpy(tex)
-        return tex.to(device).float()
-
-
-def _format_output(image: torch.Tensor, return_type: str) -> Union[torch.Tensor, np.ndarray, Image.Image]:
-    """Convert output to requested format."""
-    if return_type == ReturnType.NUMPY.value:
-        return image.cpu().numpy()
-    elif return_type == ReturnType.PIL.value:
-        img_np = image.cpu().numpy() * 255
-        return Image.fromarray(img_np.astype(np.uint8))
-    return image
-
-
-def _ensure_resolution_format(resolution: Optional[Union[int, Tuple[int, int]]], 
-                             default: Tuple[int, int]) -> Tuple[int, int]:
-    """Ensure resolution is in (height, width) format."""
-    if resolution is None:
-        return default
-    if isinstance(resolution, (int, float)):
-        return (int(resolution), int(resolution))
-    return tuple(resolution)
-
-
-def _apply_background_mask(content: torch.Tensor, visible_mask: torch.Tensor, 
-                          bg_color: List[float], device: str) -> torch.Tensor:
-    """Apply background color to masked regions."""
-    bg_tensor = torch.tensor(bg_color, dtype=torch.float32, device=device)
-    return content * visible_mask + bg_tensor * (1 - visible_mask)
-
-
-class MeshRender:
-    def __init__(
-        self,
-        camera_distance=1.45,
-        camera_type="orth",
-        default_resolution=1024,
-        texture_size=1024,
-        use_antialias=True,
-        max_mip_level=None,
-        filter_mode="linear-mipmap-linear",
-        bake_mode="back_sample",
-        raster_mode="cr",
-        shader_type="face",
-        use_opengl=False,
-        device="cuda",
-    ):
-        """
-        Initialize mesh renderer with configurable parameters.
-        
-        Args:
-            camera_distance: Distance from camera to object center
-            camera_type: Type of camera projection ("orth" or "perspective") 
-            default_resolution: Default rendering resolution
-            texture_size: Size of texture maps
-            use_antialias: Whether to use antialiasing
-            max_mip_level: Maximum mipmap level for texture filtering
-            filter_mode: Texture filtering mode
-            bake_mode: Texture baking method ("back_sample", "linear", "mip-map")
-            raster_mode: Rasterization backend ("cr" for custom rasterizer)
-            shader_type: Shading type ("face" or "vertex")
-            use_opengl: Whether to use OpenGL backend (deprecated)
-            device: Computing device ("cuda" or "cpu")
-        """
-
-        self.device = device
-
-        self.set_default_render_resolution(default_resolution)
-        self.set_default_texture_resolution(texture_size)
-
-        self.camera_distance = camera_distance
-        self.use_antialias = use_antialias
-        self.max_mip_level = max_mip_level
-        self.filter_mode = filter_mode
-        self.bake_angle_thres = 75
-        self.set_boundary_unreliable_scale(2)
-        self.bake_mode = bake_mode
-        self.shader_type = shader_type
-
-        self.raster_mode = raster_mode
-        if self.raster_mode == "cr":
-            import custom_rasterizer as cr
-
-            self.raster = cr
-        else:
-            raise f"No raster named {self.raster_mode}"
-
-        if camera_type == "orth":
-            self.set_orth_scale(1.2)
-        elif camera_type == "perspective":
-            self.camera_proj_mat = get_perspective_projection_matrix(
-                49.13, self.default_resolution[1] / self.default_resolution[0], 0.01, 100.0
-            )
-        else:
-            raise f"No camera type {camera_type}"
-
-        # Removed multiprocessing components for single-threaded version
-
-    def _create_view_state(self, config: RenderConfig) -> ViewState:
-        """Create unified view state for rendering pipeline."""
-        proj = self.camera_proj_mat
-        r_mv = get_mv_matrix(
-            elev=config.elev,
-            azim=config.azim,
-            camera_distance=self.camera_distance if config.camera_distance is None else config.camera_distance,
-            center=config.center,
-        )
-        
-        pos_camera = transform_pos(r_mv, self.vtx_pos, keepdim=True)
-        pos_clip = transform_pos(proj, pos_camera)
-        resolution = _ensure_resolution_format(config.resolution, self.default_resolution)
-        
-        return ViewState(proj, r_mv, pos_camera, pos_clip, resolution)
-
-    def _compute_face_normals(self, triangles: torch.Tensor) -> torch.Tensor:
-        """Compute face normals from triangle vertices."""
-        return F.normalize(
-            torch.cross(
-                triangles[:, 1, :] - triangles[:, 0, :],
-                triangles[:, 2, :] - triangles[:, 0, :],
-                dim=-1,
-            ),
-            dim=-1,
-        )
-
-    def _get_normals_for_shading(self, view_state: ViewState, use_abs_coor: bool = False) -> torch.Tensor:
-        """Get normals based on shader type and coordinate system."""
-        if use_abs_coor:
-            mesh_triangles = self.vtx_pos[self.pos_idx[:, :3], :]
-        else:
-            pos_camera = view_state.pos_camera[:, :3] / view_state.pos_camera[:, 3:4]
-            mesh_triangles = pos_camera[self.pos_idx[:, :3], :]
-        
-        face_normals = self._compute_face_normals(mesh_triangles)
-        
-        # Common rasterization
-        rast_out, _ = self.raster_rasterize(view_state.pos_clip, self.pos_idx, resolution=view_state.resolution)
-        
-        if self.shader_type == "vertex":
-            vertex_normals = trimesh.geometry.mean_vertex_normals(
-                vertex_count=self.vtx_pos.shape[0],
-                faces=self.pos_idx.cpu(),
-                face_normals=face_normals.cpu(),
-            )
-            vertex_normals = torch.from_numpy(vertex_normals).float().to(self.device).contiguous()
-            normal, _ = self.raster_interpolate(vertex_normals[None, ...], rast_out, self.pos_idx)
-        
-        elif self.shader_type == "face":
-            tri_ids = rast_out[..., 3]
-            tri_ids_mask = tri_ids > 0
-            tri_ids = ((tri_ids - 1) * tri_ids_mask).long()
-            normal = torch.zeros(rast_out.shape[0], rast_out.shape[1], rast_out.shape[2], 3).to(rast_out)
-            normal.reshape(-1, 3)[tri_ids_mask.view(-1)] = face_normals.reshape(-1, 3)[tri_ids[tri_ids_mask].view(-1)]
-        
-        return normal, rast_out
-
-    def _unified_render_pipeline(self, config: RenderConfig, mode: RenderMode, **kwargs) -> torch.Tensor:
-        """Unified rendering pipeline for all render modes."""
-        view_state = self._create_view_state(config)
-        
-        if mode == RenderMode.ALPHA:
-            rast_out, _ = self.raster_rasterize(view_state.pos_clip, self.pos_idx, resolution=view_state.resolution)
-            return rast_out[..., -1:].long()
-        
-        elif mode == RenderMode.UV_POS:
-            return self.uv_feature_map(self.vtx_pos * 0.5 + 0.5)
-        
-        elif mode == RenderMode.NORMAL:
-            use_abs_coor = kwargs.get('use_abs_coor', False)
-            normalize_rgb = kwargs.get('normalize_rgb', True)
-            
-            normal, rast_out = self._get_normals_for_shading(view_state, use_abs_coor)
-            visible_mask = torch.clamp(rast_out[..., -1:], 0, 1)
-            
-            result = _apply_background_mask(normal, visible_mask, config.bg_color, self.device)
-            
-            if normalize_rgb:
-                result = (result + 1) * 0.5
-            
-            if self.use_antialias:
-                result = self.raster_antialias(result, rast_out, view_state.pos_clip, self.pos_idx)
-            
-            return result[0, ...]
-        
-        elif mode == RenderMode.POSITION:
-            rast_out, _ = self.raster_rasterize(view_state.pos_clip, self.pos_idx, resolution=view_state.resolution)
-            
-            tex_position = 0.5 - self.vtx_pos[:, :3] / self.scale_factor
-            tex_position = tex_position.contiguous()
-            
-            position, _ = self.raster_interpolate(tex_position[None, ...], rast_out, self.pos_idx)
-            visible_mask = torch.clamp(rast_out[..., -1:], 0, 1)
-            
-            result = _apply_background_mask(position, visible_mask, config.bg_color, self.device)
-            
-            if self.use_antialias:
-                result = self.raster_antialias(result, rast_out, view_state.pos_clip, self.pos_idx)
-            
-            return result[0, ...]
-
-    def set_orth_scale(self, ortho_scale):
-        """
-        Set the orthographic projection scale and update camera projection matrix.
-        
-        Args:
-            ortho_scale: Scale factor for orthographic projection
-        """
-        self.ortho_scale = ortho_scale
-        self.camera_proj_mat = get_orthographic_projection_matrix(
-            left=-self.ortho_scale * 0.5,
-            right=self.ortho_scale * 0.5,
-            bottom=-self.ortho_scale * 0.5,
-            top=self.ortho_scale * 0.5,
-            near=0.1,
-            far=100,
-        )
-
-    def raster_rasterize(self, pos, tri, resolution, ranges=None, grad_db=True):
-        """
-        Rasterize triangular mesh using the configured rasterization backend.
-        
-        Args:
-            pos: Vertex positions in clip space
-            tri: Triangle indices
-            resolution: Rendering resolution [height, width]
-            ranges: Optional rendering ranges (unused in current implementation)
-            grad_db: Whether to compute gradients (unused in current implementation)
-            
-        Returns:
-            Tuple of (rasterization_output, gradient_info)
-        """
-
-        if self.raster_mode == "cr":
-            rast_out_db = None
-            if pos.dim() == 2:
-                pos = pos.unsqueeze(0)
-
-            # 确保pos是float32类型
-            if pos.dtype == torch.float64:
-                pos = pos.to(torch.float32)
-
-            # 确保tri是int32类型
-            if tri.dtype == torch.int64:
-                tri = tri.to(torch.int32)
-
-            findices, barycentric = self.raster.rasterize(pos, tri, resolution)
-            rast_out = torch.cat((barycentric, findices.unsqueeze(-1)), dim=-1)
-            rast_out = rast_out.unsqueeze(0)
-        else:
-            raise f"No raster named {self.raster_mode}"
-
-        return rast_out, rast_out_db
-
-    def raster_interpolate(self, uv, rast_out, uv_idx):
-        """
-        Interpolate texture coordinates or vertex attributes across rasterized triangles.
-        
-        Args:
-            uv: UV coordinates or vertex attributes to interpolate
-            rast_out: Rasterization output containing barycentric coordinates
-            uv_idx: UV or vertex indices for triangles
-            
-        Returns:
-            Tuple of (interpolated_values, gradient_info)
-        """
-
-        if self.raster_mode == "cr":
-            textd = None
-            barycentric = rast_out[0, ..., :-1]
-            findices = rast_out[0, ..., -1]
-            if uv.dim() == 2:
-                uv = uv.unsqueeze(0)
-            textc = self.raster.interpolate(uv, findices, barycentric, uv_idx)
-        else:
-            raise f"No raster named {self.raster_mode}"
-
-        return textc, textd
-
-    def raster_antialias(self, color, rast, pos, tri, topology_hash=None, pos_gradient_boost=1.0):
-        """
-        Apply antialiasing to rendered colors (currently returns input unchanged).
-        
-        Args:
-            color: Input color values
-            rast: Rasterization output
-            pos: Vertex positions
-            tri: Triangle indices  
-            topology_hash: Optional topology hash for optimization
-            pos_gradient_boost: Gradient boosting factor
-            
-        Returns:
-            Antialiased color values
-        """
-
-        if self.raster_mode == "cr":
-            color = color
-        else:
-            raise f"No raster named {self.raster_mode}"
-
-        return color
-
-    def set_boundary_unreliable_scale(self, scale):
-        """
-        Set the kernel size for boundary unreliable region detection during texture baking.
-        
-        Args:
-            scale: Scale factor relative to 512 resolution baseline
-        """
-        self.bake_unreliable_kernel_size = int(
-            (scale / 512) * max(self.default_resolution[0], self.default_resolution[1])
-        )
-
-    def load_mesh(
-        self,
-        mesh,
-        scale_factor=1.15,
-        auto_center=True,
-    ):
-        """
-        Load mesh from file and set up rendering data structures.
-        
-        Args:
-            mesh: Path to mesh file or mesh object
-            scale_factor: Scaling factor for mesh normalization
-            auto_center: Whether to automatically center the mesh
-        """
-        vtx_pos, pos_idx, vtx_uv, uv_idx, texture_data = load_mesh(mesh)
-        self.set_mesh(
-            vtx_pos, pos_idx, vtx_uv=vtx_uv, uv_idx=uv_idx, scale_factor=scale_factor, auto_center=auto_center
-        )
-        if texture_data is not None:
-            self.set_texture(texture_data)
-
-    def save_mesh(self, mesh_path, downsample=False):
-        """
-        Save current mesh with textures to file.
-        
-        Args:
-            mesh_path: Output file path
-            downsample: Whether to downsample textures by half
-        """
-
-        vtx_pos, pos_idx, vtx_uv, uv_idx = self.get_mesh(normalize=False)
-        texture_data = self.get_texture()
-        texture_metallic, texture_roughness = self.get_texture_mr()
-        texture_normal = self.get_texture_normal()
-        if downsample:
-            texture_data = cv2.resize(texture_data, (texture_data.shape[1] // 2, texture_data.shape[0] // 2))
-            if texture_metallic is not None:
-                texture_metallic = cv2.resize(
-                    texture_metallic, (texture_metallic.shape[1] // 2, texture_metallic.shape[0] // 2)
-                )
-            if texture_roughness is not None:
-                texture_roughness = cv2.resize(
-                    texture_roughness, (texture_roughness.shape[1] // 2, texture_roughness.shape[0] // 2)
-                )
-            if texture_normal is not None:
-                texture_normal = cv2.resize(
-                    texture_normal, (texture_normal.shape[1] // 2, texture_normal.shape[0] // 2)
-                )
-
-        save_mesh(
-            mesh_path,
-            vtx_pos,
-            pos_idx,
-            vtx_uv,
-            uv_idx,
-            texture_data,
-            metallic=texture_metallic,
-            roughness=texture_roughness,
-            normal=texture_normal,
-        )
-
-    def set_mesh(self, vtx_pos, pos_idx, vtx_uv=None, uv_idx=None, scale_factor=1.15, auto_center=True):
-        """
-        Set mesh geometry data and perform coordinate transformations.
-        
-        Args:
-            vtx_pos: Vertex positions [N, 3]
-            pos_idx: Triangle vertex indices [F, 3]
-            vtx_uv: UV coordinates [N, 2], optional
-            uv_idx: Triangle UV indices [F, 3], optional  
-            scale_factor: Scaling factor for mesh normalization
-            auto_center: Whether to automatically center and scale the mesh
-        """
-        self.vtx_pos = torch.from_numpy(vtx_pos).to(self.device)
-        self.pos_idx = torch.from_numpy(pos_idx).to(self.device)
-
-        # 确保顶点位置是float32类型
-        if self.vtx_pos.dtype == torch.float64:
-            self.vtx_pos = self.vtx_pos.to(torch.float32)
-
-        # 确保索引类型为int32
-        if self.pos_idx.dtype == torch.int64:
-            self.pos_idx = self.pos_idx.to(torch.int32)
-
-        if (vtx_uv is not None) and (uv_idx is not None):
-            self.vtx_uv = torch.from_numpy(vtx_uv).to(self.device)
-            self.uv_idx = torch.from_numpy(uv_idx).to(self.device)
-
-            # 确保UV坐标是float32类型
-            if self.vtx_uv.dtype == torch.float64:
-                self.vtx_uv = self.vtx_uv.to(torch.float32)
-
-            # 确保UV索引类型为int32
-            if self.uv_idx.dtype == torch.int64:
-                self.uv_idx = self.uv_idx.to(torch.int32)
-        else:
-            self.vtx_uv = None
-            self.uv_idx = None
-
-        self.vtx_pos[:, [0, 1]] = -self.vtx_pos[:, [0, 1]]
-        self.vtx_pos[:, [1, 2]] = self.vtx_pos[:, [2, 1]]
-        if (vtx_uv is not None) and (uv_idx is not None):
-            self.vtx_uv[:, 1] = 1.0 - self.vtx_uv[:, 1]
-            pass
-
-        if auto_center:
-            max_bb = (self.vtx_pos - 0).max(0)[0]
-            min_bb = (self.vtx_pos - 0).min(0)[0]
-            center = (max_bb + min_bb) / 2
-            scale = torch.norm(self.vtx_pos - center, dim=1).max() * 2.0
-            self.vtx_pos = (self.vtx_pos - center) * (scale_factor / float(scale))
-            self.scale_factor = scale_factor
-            self.mesh_normalize_scale_factor = scale_factor / float(scale)
-            self.mesh_normalize_scale_center = center.unsqueeze(0).cpu().numpy()
-        else:
-            self.scale_factor = 1.0
-            self.mesh_normalize_scale_factor = 1.0
-            self.mesh_normalize_scale_center = np.array([[0, 0, 0]])
-
-        if uv_idx is not None:
-            self.extract_textiles()
-
-    def _set_texture_unified(self, tex: Union[np.ndarray, torch.Tensor, Image.Image], 
-                           texture_type: TextureType, force_set: bool = False):
-        """Unified texture setting method."""
-        converted_tex = _convert_texture_format(tex, self.texture_size, self.device, force_set)
-        
-        if texture_type == TextureType.DIFFUSE:
-            self.tex = converted_tex
-        elif texture_type == TextureType.METALLIC_ROUGHNESS:
-            self.tex_mr = converted_tex
-        elif texture_type == TextureType.NORMAL:
-            self.tex_normalMap = converted_tex
-
-    def set_texture(self, tex, force_set=False):
-        """Set the main diffuse texture for the mesh."""
-        self._set_texture_unified(tex, TextureType.DIFFUSE, force_set)
-
-    def set_texture_mr(self, mr, force_set=False):
-        """Set metallic-roughness texture for PBR rendering."""
-        self._set_texture_unified(mr, TextureType.METALLIC_ROUGHNESS, force_set)
-
-    def set_texture_normal(self, normal, force_set=False):
-        """Set normal map texture for surface detail."""
-        self._set_texture_unified(normal, TextureType.NORMAL, force_set)
-
-    def set_default_render_resolution(self, default_resolution):
-        """
-        Set the default resolution for rendering operations.
-        
-        Args:
-            default_resolution: Resolution as int (square) or tuple (height, width)
-        """
-        if isinstance(default_resolution, int):
-            default_resolution = (default_resolution, default_resolution)
-        self.default_resolution = default_resolution
-
-    def set_default_texture_resolution(self, texture_size):
-        """
-        Set the default texture resolution for UV mapping operations.
-        
-        Args:
-            texture_size: Texture size as int (square) or tuple (height, width)
-        """
-        if isinstance(texture_size, int):
-            texture_size = (texture_size, texture_size)
-        self.texture_size = texture_size
-
-    def get_face_num(self):
-        """
-        Get the number of triangular faces in the mesh.
-        
-        Returns:
-            Number of faces as integer
-        """
-        return self.pos_idx.shape[0]
-
-    def get_vertex_num(self):
-        """
-        Get the number of vertices in the mesh.
-        
-        Returns:
-            Number of vertices as integer
-        """
-        return self.vtx_pos.shape[0]
-
-    def get_face_areas(self, from_one_index=False):
-        """
-        Calculate the area of each triangular face in the mesh.
-        
-        Args:
-            from_one_index: If True, insert zero at beginning for 1-indexed face IDs
-            
-        Returns:
-            Numpy array of face areas
-        """
-        v0 = self.vtx_pos[self.pos_idx[:, 0], :]
-        v1 = self.vtx_pos[self.pos_idx[:, 1], :]
-        v2 = self.vtx_pos[self.pos_idx[:, 2], :]
-
-        # 计算两个边向量
-        edge1 = v1 - v0
-        edge2 = v2 - v0
-
-        # 计算叉积的模长的一半即为面积
-        areas = torch.norm(torch.cross(edge1, edge2, dim=-1), dim=-1) * 0.5
-
-        areas = areas.cpu().numpy()
-
-        if from_one_index:
-            # 在数组前面插入一个0,��为三角片索引是从1开始的
-            areas = np.insert(areas, 0, 0)
-
-        return areas
-
-    def get_mesh(self, normalize=True):
-        """
-        Get mesh geometry with optional coordinate denormalization.
-        
-        Args:
-            normalize: Whether to keep normalized coordinates (True) or restore original scale (False)
-            
-        Returns:
-            Tuple of (vertex_positions, face_indices, uv_coordinates, uv_indices)
-        """
-        vtx_pos = self.vtx_pos.cpu().numpy()
-        pos_idx = self.pos_idx.cpu().numpy()
-        vtx_uv = self.vtx_uv.cpu().numpy()
-        uv_idx = self.uv_idx.cpu().numpy()
-
-        # 坐标变换的逆变换
-        if not normalize:
-            vtx_pos = vtx_pos / self.mesh_normalize_scale_factor
-            vtx_pos = vtx_pos + self.mesh_normalize_scale_center
-        vtx_pos[:, [1, 2]] = vtx_pos[:, [2, 1]]
-        vtx_pos[:, [0, 1]] = -vtx_pos[:, [0, 1]]
-
-        vtx_uv[:, 1] = 1.0 - vtx_uv[:, 1]
-        return vtx_pos, pos_idx, vtx_uv, uv_idx
-
-    def get_texture(self):
-        """
-        Get the current diffuse texture as numpy array.
-        
-        Returns:
-            Texture as numpy array in range [0, 1]
-        """
-        return self.tex.cpu().numpy()
-
-    def get_texture_mr(self):
-        """
-        Get metallic and roughness textures as separate channels.
-        
-        Returns:
-            Tuple of (metallic_texture, roughness_texture) as numpy arrays, or (None, None) if not set
-        """
-        metallic, roughness = None, None
-        if hasattr(self, "tex_mr"):
-            mr = self.tex_mr.cpu().numpy()
-            metallic = np.repeat(mr[:, :, 0:1], repeats=3, axis=2)
-            roughness = np.repeat(mr[:, :, 1:2], repeats=3, axis=2)
-        return metallic, roughness
-
-    def get_texture_normal(self):
-        """
-        Get the normal map texture as numpy array.
-        
-        Returns:
-            Normal map as numpy array, or None if not set
-        """
-        normal = None
-        if hasattr(self, "tex_normalMap"):
-            normal = self.tex_normalMap.cpu().numpy()
-        return normal
-
-    def to(self, device):
-        """
-        Move all tensor attributes to the specified device.
-        
-        Args:
-            device: Target device ("cuda", "cpu", etc.)
-        """
-        self.device = device
-
-        for attr_name in dir(self):
-            attr_value = getattr(self, attr_name)
-            if isinstance(attr_value, torch.Tensor):
-                setattr(self, attr_name, attr_value.to(self.device))
-
-    def color_rgb_to_srgb(self, image):
-        """
-        Convert RGB color values to sRGB color space using gamma correction.
-        
-        Args:
-            image: Input image as PIL Image, numpy array, or torch tensor
-            
-        Returns:
-            sRGB corrected image in same format as input
-        """
-        if isinstance(image, Image.Image):
-            image_rgb = torch.tesnor(np.array(image) / 255.0).float().to(self.device)
-        elif isinstance(image, np.ndarray):
-            image_rgb = torch.tensor(image).float()
-        else:
-            image_rgb = image.to(self.device)
-
-        image_srgb = torch.where(
-            image_rgb <= 0.0031308, 12.92 * image_rgb, 1.055 * torch.pow(image_rgb, 1 / 2.4) - 0.055
-        )
-
-        if isinstance(image, Image.Image):
-            image_srgb = Image.fromarray((image_srgb.cpu().numpy() * 255).astype(np.uint8))
-        elif isinstance(image, np.ndarray):
-            image_srgb = image_srgb.cpu().numpy()
-        else:
-            image_srgb = image_srgb.to(image.device)
-
-        return image_srgb
-
-    def extract_textiles(self):
-        """
-        Extract texture-space position and normal information by rasterizing 
-        the mesh in UV coordinate space. Creates texture-space geometry mappings.
-        """
-
-        vnum = self.vtx_uv.shape[0]
-        vtx_uv = torch.cat(
-            (self.vtx_uv, torch.zeros_like(self.vtx_uv[:, 0:1]), torch.ones_like(self.vtx_uv[:, 0:1])), axis=1
-        )
-        vtx_uv = vtx_uv.view(1, vnum, 4) * 2 - 1
-
-        rast_out, rast_out_db = self.raster_rasterize(vtx_uv, self.uv_idx, resolution=self.texture_size)
-        position, _ = self.raster_interpolate(self.vtx_pos, rast_out, self.pos_idx)
-
-        v0 = self.vtx_pos[self.pos_idx[:, 0], :]
-        v1 = self.vtx_pos[self.pos_idx[:, 1], :]
-        v2 = self.vtx_pos[self.pos_idx[:, 2], :]
-        face_normals = F.normalize(torch.cross(v1 - v0, v2 - v0, dim=-1), dim=-1)
-        vertex_normals = trimesh.geometry.mean_vertex_normals(
-            vertex_count=self.vtx_pos.shape[0],
-            faces=self.pos_idx.cpu(),
-            face_normals=face_normals.cpu(),
-        )
-        vertex_normals = torch.from_numpy(vertex_normals).to(self.vtx_pos).contiguous()
-        position_normal, _ = self.raster_interpolate(vertex_normals[None, ...], rast_out, self.pos_idx)
-        visible_mask = torch.clamp(rast_out[..., -1:], 0, 1)[0, ..., 0]
-        position = position[0]
-        position_normal = position_normal[0]
-        tri_ids = rast_out[0, ..., 3]
-        tri_ids_mask = tri_ids > 0
-        tri_ids = ((tri_ids - 1) * tri_ids_mask).long()
-        position_normal.reshape(-1, 3)[tri_ids_mask.view(-1)] = face_normals.reshape(-1, 3)[
-            tri_ids[tri_ids_mask].view(-1)
-        ]
-
-        row = torch.arange(position.shape[0]).to(visible_mask.device)
-        col = torch.arange(position.shape[1]).to(visible_mask.device)
-        grid_i, grid_j = torch.meshgrid(row, col, indexing="ij")
-
-        mask = visible_mask.reshape(-1) > 0
-        position = position.reshape(-1, 3)[mask]
-        position_normal = position_normal.reshape(-1, 3)[mask]
-        position = torch.cat((position, torch.ones_like(position[:, :1])), axis=-1)
-        grid = torch.stack((grid_i, grid_j), -1).reshape(-1, 2)[mask]
-
-        texture_indices = (
-            torch.ones(self.texture_size[0], self.texture_size[1], device=self.device, dtype=torch.long) * -1
-        )
-        texture_indices.view(-1)[grid[:, 0] * self.texture_size[1] + grid[:, 1]] = torch.arange(grid.shape[0]).to(
-            device=self.device, dtype=torch.long
-        )
-
-        self.tex_position = position
-        self.tex_normal = position_normal
-        self.tex_grid = grid
-        self.texture_indices = texture_indices
-
-    def render_normal(self, elev, azim, camera_distance=None, center=None, resolution=None,
-                     bg_color=[1, 1, 1], use_abs_coor=False, normalize_rgb=True, return_type="th"):
-        """Render surface normals of the mesh from specified viewpoint."""
-        config = RenderConfig(elev, azim, camera_distance, center, resolution, bg_color, return_type)
-        image = self._unified_render_pipeline(config, RenderMode.NORMAL, 
-                                            use_abs_coor=use_abs_coor, normalize_rgb=normalize_rgb)
-        return _format_output(image, return_type)
-
-    def convert_normal_map(self, image):
-        """
-        Convert normal map from standard format to renderer's coordinate system.
-        Applies coordinate transformations for proper normal interpretation.
-        
-        Args:
-            image: Input normal map as PIL Image or numpy array
-            
-        Returns:
-            Converted normal map as PIL Image
-        """
-        # blue is front, red is left, green is top
-        if isinstance(image, Image.Image):
-            image = np.array(image)
-        mask = (image == [255, 255, 255]).all(axis=-1)
-
-        image = (image / 255.0) * 2.0 - 1.0
-
-        image[..., [1]] = -image[..., [1]]
-        image[..., [1, 2]] = image[..., [2, 1]]
-        image[..., [0]] = -image[..., [0]]
-
-        image = (image + 1.0) * 0.5
-
-        image = (image * 255).astype(np.uint8)
-        image[mask] = [127, 127, 255]
-
-        return Image.fromarray(image)
-
-    def render_position(self, elev, azim, camera_distance=None, center=None, resolution=None, 
-                       bg_color=[1, 1, 1], return_type="th"):
-        """Render world-space positions of visible mesh surface points."""
-        config = RenderConfig(elev, azim, camera_distance, center, resolution, bg_color, return_type)
-        image = self._unified_render_pipeline(config, RenderMode.POSITION)
-        
-        if return_type == ReturnType.PIL.value:
-            image = image.squeeze(-1).cpu().numpy() * 255
-            return Image.fromarray(image.astype(np.uint8))
-        return _format_output(image, return_type)
-
-    def render_uvpos(self, return_type="th"):
-        """Render vertex positions mapped to UV texture space."""
-        config = RenderConfig(return_type=return_type)
-        image = self._unified_render_pipeline(config, RenderMode.UV_POS)
-        return _format_output(image, return_type)
-
-    def render_alpha(self, elev, azim, camera_distance=None, center=None, resolution=None, return_type="th"):
-        """Render binary alpha mask indicating visible mesh regions."""
-        config = RenderConfig(elev, azim, camera_distance, center, resolution, return_type=return_type)
-        image = self._unified_render_pipeline(config, RenderMode.ALPHA)
-        
-        if return_type == ReturnType.PIL.value:
-            raise Exception("PIL format not supported for alpha rendering")
-        return _format_output(image, return_type)
-
-    def uv_feature_map(self, vert_feat, bg=None):
-        """
-        Map per-vertex features to UV texture space using mesh topology.
-        
-        Args:
-            vert_feat: Per-vertex feature tensor [N, C]
-            bg: Background value for unmapped regions (optional)
-            
-        Returns:
-            Feature map in UV texture space [H, W, C]
-        """
-        vtx_uv = self.vtx_uv * 2 - 1.0
-        vtx_uv = torch.cat([vtx_uv, torch.zeros_like(self.vtx_uv)], dim=1).unsqueeze(0)
-        vtx_uv[..., -1] = 1
-        uv_idx = self.uv_idx
-        rast_out, rast_out_db = self.raster_rasterize(vtx_uv, uv_idx, resolution=self.texture_size)
-        feat_map, _ = self.raster_interpolate(vert_feat[None, ...], rast_out, uv_idx)
-        feat_map = feat_map[0, ...]
-        if bg is not None:
-            visible_mask = torch.clamp(rast_out[..., -1:], 0, 1)[0, ...]
-            feat_map[visible_mask == 0] = bg
-        return feat_map
-
-    def render_sketch_from_geometry(self, normal_image, depth_image):
-        """
-        Generate sketch-style edge image from rendered normal and depth maps.
-        
-        Args:
-            normal_image: Rendered normal map tensor
-            depth_image: Rendered depth map tensor
-            
-        Returns:
-            Binary edge sketch image as tensor
-        """
-        normal_image_np = normal_image.cpu().numpy()
-        depth_image_np = depth_image.cpu().numpy()
-
-        normal_image_np = (normal_image_np * 255).astype(np.uint8)
-        depth_image_np = (depth_image_np * 255).astype(np.uint8)
-        normal_image_np = cv2.cvtColor(normal_image_np, cv2.COLOR_RGB2GRAY)
-
-        normal_edges = cv2.Canny(normal_image_np, 80, 150)
-        depth_edges = cv2.Canny(depth_image_np, 30, 80)
-
-        combined_edges = np.maximum(normal_edges, depth_edges)
-
-        sketch_image = torch.from_numpy(combined_edges).to(normal_image.device).float() / 255.0
-        sketch_image = sketch_image.unsqueeze(-1)
-
-        return sketch_image
-
-    def render_sketch_from_depth(self, depth_image):
-        """
-        Generate sketch-style edge image from depth map using edge detection.
-        
-        Args:
-            depth_image: Input depth map tensor
-            
-        Returns:
-            Binary edge sketch image as tensor
-        """
-        depth_image_np = depth_image.cpu().numpy()
-        depth_image_np = (depth_image_np * 255).astype(np.uint8)
-        depth_edges = cv2.Canny(depth_image_np, 30, 80)
-        combined_edges = depth_edges
-        sketch_image = torch.from_numpy(combined_edges).to(depth_image.device).float() / 255.0
-        sketch_image = sketch_image.unsqueeze(-1)
-        return sketch_image
-
-    def back_project(self, image, elev, azim, camera_distance=None, center=None, method=None):
-        """
-        Back-project a rendered image onto the mesh's UV texture space.
-        Handles visibility, viewing angle, and boundary detection for texture baking.
-        
-        Args:
-            image: Input image to back-project (PIL Image, numpy array, or tensor)
-            elev: Camera elevation angle in degrees used for rendering
-            azim: Camera azimuth angle in degrees used for rendering
-            camera_distance: Camera distance (uses default if None)
-            center: Camera focus center (uses origin if None) 
-            method: Back-projection method ("linear", "mip-map", "back_sample", uses default if None)
-            
-        Returns:
-            Tuple of (texture, cosine_map, boundary_map) tensors in UV space
-        """
-
-        if isinstance(image, Image.Image):
-            image = torch.tensor(np.array(image) / 255.0)
-        elif isinstance(image, np.ndarray):
-            image = torch.tensor(image)
-        if image.dim() == 2:
-            image = image.unsqueeze(-1)
-        image = image.float().to(self.device)
-        resolution = image.shape[:2]
-        channel = image.shape[-1]
-        texture = torch.zeros(self.texture_size + (channel,)).to(self.device)
-        cos_map = torch.zeros(self.texture_size + (1,)).to(self.device)
-
-        proj = self.camera_proj_mat
-        r_mv = get_mv_matrix(
-            elev=elev,
-            azim=azim,
-            camera_distance=self.camera_distance if camera_distance is None else camera_distance,
-            center=center,
-        )
-        pos_camera = transform_pos(r_mv, self.vtx_pos, keepdim=True)
-        pos_clip = transform_pos(proj, pos_camera)
-        pos_camera = pos_camera[:, :3] / pos_camera[:, 3:4]
-
-        v0 = pos_camera[self.pos_idx[:, 0], :]
-        v1 = pos_camera[self.pos_idx[:, 1], :]
-        v2 = pos_camera[self.pos_idx[:, 2], :]
-        face_normals = F.normalize(torch.cross(v1 - v0, v2 - v0, dim=-1), dim=-1)
-
-        tex_depth = pos_camera[:, 2].reshape(1, -1, 1).contiguous()
-        rast_out, rast_out_db = self.raster_rasterize(pos_clip, self.pos_idx, resolution=resolution)
-        visible_mask = torch.clamp(rast_out[..., -1:], 0, 1)[0, ...]
-
-        if self.shader_type == "vertex":
-            vertex_normals = trimesh.geometry.mean_vertex_normals(
-                vertex_count=self.vtx_pos.shape[0],
-                faces=self.pos_idx.cpu(),
-                face_normals=face_normals.cpu(),
-            )
-            vertex_normals = torch.from_numpy(vertex_normals).float().to(self.device).contiguous()
-            normal, _ = self.raster_interpolate(vertex_normals[None, ...], rast_out, self.pos_idx)
-        elif self.shader_type == "face":
-            tri_ids = rast_out[..., 3]
-            tri_ids_mask = tri_ids > 0
-            tri_ids = ((tri_ids - 1) * tri_ids_mask).long()
-            normal = torch.zeros(rast_out.shape[0], rast_out.shape[1], rast_out.shape[2], 3).to(rast_out)
-            normal.reshape(-1, 3)[tri_ids_mask.view(-1)] = face_normals.reshape(-1, 3)[tri_ids[tri_ids_mask].view(-1)]
-
-        normal = normal[0, ...]
-        uv, _ = self.raster_interpolate(self.vtx_uv[None, ...], rast_out, self.uv_idx)
-        depth, _ = self.raster_interpolate(tex_depth, rast_out, self.pos_idx)
-        depth = depth[0, ...]
-
-        depth_max, depth_min = depth[visible_mask > 0].max(), depth[visible_mask > 0].min()
-        depth_normalized = (depth - depth_min) / (depth_max - depth_min)
-        depth_image = depth_normalized * visible_mask  # Mask out background.
-
-        sketch_image = self.render_sketch_from_depth(depth_image)
-
-        lookat = torch.tensor([[0, 0, -1]], device=self.device)
-        cos_image = torch.nn.functional.cosine_similarity(lookat, normal.view(-1, 3))
-        cos_image = cos_image.view(normal.shape[0], normal.shape[1], 1)
-
-        cos_thres = np.cos(self.bake_angle_thres / 180 * np.pi)
-        cos_image[cos_image < cos_thres] = 0
-
-        # shrink
-        if self.bake_unreliable_kernel_size > 0:
-            kernel_size = self.bake_unreliable_kernel_size * 2 + 1
-            kernel = torch.ones((1, 1, kernel_size, kernel_size), dtype=torch.float32).to(sketch_image.device)
-
-            visible_mask = visible_mask.permute(2, 0, 1).unsqueeze(0).float()
-            visible_mask = F.conv2d(1.0 - visible_mask, kernel, padding=kernel_size // 2)
-            visible_mask = 1.0 - (visible_mask > 0).float()  # 二值化
-            visible_mask = visible_mask.squeeze(0).permute(1, 2, 0)
-
-            sketch_image = sketch_image.permute(2, 0, 1).unsqueeze(0)
-            sketch_image = F.conv2d(sketch_image, kernel, padding=kernel_size // 2)
-            sketch_image = (sketch_image > 0).float()  # 二值化
-            sketch_image = sketch_image.squeeze(0).permute(1, 2, 0)
-            visible_mask = visible_mask * (sketch_image < 0.5)
-
-        cos_image[visible_mask == 0] = 0
-
-        method = self.bake_mode if method is None else method
-
-        if method == "linear":
-            proj_mask = (visible_mask != 0).view(-1)
-            uv = uv.squeeze(0).contiguous().view(-1, 2)[proj_mask]
-            image = image.squeeze(0).contiguous().view(-1, channel)[proj_mask]
-            cos_image = cos_image.contiguous().view(-1, 1)[proj_mask]
-            sketch_image = sketch_image.contiguous().view(-1, 1)[proj_mask]
-
-            texture = linear_grid_put_2d(self.texture_size[1], self.texture_size[0], uv[..., [1, 0]], image)
-            cos_map = linear_grid_put_2d(self.texture_size[1], self.texture_size[0], uv[..., [1, 0]], cos_image)
-            boundary_map = linear_grid_put_2d(self.texture_size[1], self.texture_size[0], uv[..., [1, 0]], sketch_image)
-        elif method == "mip-map":
-            proj_mask = (visible_mask != 0).view(-1)
-            uv = uv.squeeze(0).contiguous().view(-1, 2)[proj_mask]
-            image = image.squeeze(0).contiguous().view(-1, channel)[proj_mask]
-            cos_image = cos_image.contiguous().view(-1, 1)[proj_mask]
-
-            texture = mipmap_linear_grid_put_2d(
-                self.texture_size[1], self.texture_size[0], uv[..., [1, 0]], image, min_resolution=128
-            )
-            cos_map = mipmap_linear_grid_put_2d(
-                self.texture_size[1], self.texture_size[0], uv[..., [1, 0]], cos_image, min_resolution=256
-            )
-
-            if self.vtx_map is not None:
-                vertex_normals = vertex_normals[self.vtx_map, :]
-            normal_map = self.uv_feature_map(vertex_normals)
-            cos_map_uv = torch.nn.functional.cosine_similarity(lookat, normal_map.view(-1, 3))  # .abs()
-            cos_map_uv = cos_map_uv.view(1, 1, normal_map.shape[0], normal_map.shape[1])
-            cos_map_uv = torch.nn.functional.max_pool2d(cos_map_uv, kernel_size=3, stride=1, padding=1)
-            cos_map_uv = cos_map_uv.reshape(self.texture_size[0], self.texture_size[1], 1)
-            cos_map_uv[cos_map_uv < cos_thres] = 0
-            # cos_map = torch.min(cos_map, cos_map_uv)
-            cos_map[cos_map_uv < cos_thres] = 0
-        elif method == "back_sample":
-
-            img_proj = torch.from_numpy(
-                np.array(((proj[0, 0], 0, 0, 0), (0, proj[1, 1], 0, 0), (0, 0, 1, 0), (0, 0, 0, 1)))
-            ).to(self.tex_position)
-            w2c = torch.from_numpy(r_mv).to(self.tex_position)
-            v_proj = self.tex_position @ w2c.T @ img_proj
-            inner_mask = (v_proj[:, 0] <= 1.0) & (v_proj[:, 0] >= -1.0) & (v_proj[:, 1] <= 1.0) & (v_proj[:, 1] >= -1.0)
-            inner_valid_idx = torch.where(inner_mask)[0].long()
-            img_x = torch.clamp(
-                ((v_proj[:, 0].clamp(-1, 1) * 0.5 + 0.5) * (resolution[0])).long(), 0, resolution[0] - 1
-            )
-            img_y = torch.clamp(
-                ((v_proj[:, 1].clamp(-1, 1) * 0.5 + 0.5) * (resolution[1])).long(), 0, resolution[1] - 1
-            )
-
-            indices = img_y * resolution[0] + img_x
-            sampled_z = depth.reshape(-1)[indices]
-            sampled_m = visible_mask.reshape(-1)[indices]
-            v_z = v_proj[:, 2]
-
-            sampled_w = cos_image.reshape(-1)[indices]
-            depth_thres = 3e-3
-
-            # valid_idx = torch.where((torch.abs(v_z - sampled_z) < depth_thres) * (sampled_m*sampled_w>0))[0]
-            valid_idx = torch.where((torch.abs(v_z - sampled_z) < depth_thres) & (sampled_m * sampled_w > 0))[0]
-
-            intersection_mask = torch.isin(valid_idx, inner_valid_idx)
-            valid_idx = valid_idx[intersection_mask].to(inner_valid_idx)
-
-            indices = indices[valid_idx]
-            sampled_b = sketch_image.reshape(-1)[indices]
-            sampled_w = sampled_w[valid_idx]
-
-            # bilinear sampling rgb
-            wx = ((v_proj[:, 0] * 0.5 + 0.5) * resolution[0] - img_x)[valid_idx].reshape(-1, 1)
-            wy = ((v_proj[:, 1] * 0.5 + 0.5) * resolution[1] - img_y)[valid_idx].reshape(-1, 1)
-            img_x = img_x[valid_idx]
-            img_y = img_y[valid_idx]
-            img_x_r = torch.clamp(img_x + 1, 0, resolution[0] - 1)
-            img_y_r = torch.clamp(img_y + 1, 0, resolution[1] - 1)
-            indices_lr = img_y * resolution[0] + img_x_r
-            indices_rl = img_y_r * resolution[0] + img_x
-            indices_rr = img_y_r * resolution[0] + img_x_r
-            rgb = image.reshape(-1, channel)
-            sampled_rgb = (rgb[indices] * (1 - wx) + rgb[indices_lr] * wx) * (1 - wy) + (
-                rgb[indices_rl] * (1 - wx) + rgb[indices_rr] * wx
-            ) * wy
-
-            # return sampled_rgb, sampled_w, sampled_b, valid_idx
-            texture = torch.zeros(self.texture_size[0], self.texture_size[1], channel, device=self.device).reshape(
-                -1, channel
-            )
-            cos_map = torch.zeros(self.texture_size[0], self.texture_size[1], 1, device=self.device).reshape(-1)
-            boundary_map = torch.zeros(self.texture_size[0], self.texture_size[1], 1, device=self.device).reshape(-1)
-
-            valid_tex_indices = self.tex_grid[valid_idx, 0] * self.texture_size[1] + self.tex_grid[valid_idx, 1]
-            texture[valid_tex_indices, :] = sampled_rgb
-            cos_map[valid_tex_indices] = sampled_w
-            boundary_map[valid_tex_indices] = sampled_b
-
-            texture = texture.view(self.texture_size[0], self.texture_size[1], channel)
-            cos_map = cos_map.view(self.texture_size[0], self.texture_size[1], 1)
-            # texture = torch.clamp(texture,0,1)
-
-        else:
-            raise f"No bake mode {method}"
-        return texture, cos_map, boundary_map
-
-    def bake_texture(self, colors, elevs, azims, camera_distance=None, center=None, exp=6, weights=None):
-        """
-        Bake multiple view images into a single UV texture using weighted blending.
-        
-        Args:
-            colors: List of input images (tensors, numpy arrays, or PIL Images)
-            elevs: List of elevation angles for each view
-            azims: List of azimuth angles for each view
-            camera_distance: Camera distance (uses default if None)
-            center: Camera focus center (uses origin if None)
-            exp: Exponent for cosine weighting (higher values favor front-facing views)
-            weights: Optional per-view weights (defaults to 1.0 for all views)
-            
-        Returns:
-            Tuple of (merged_texture, trust_map) tensors in UV space
-        """
-        if isinstance(colors, torch.Tensor):
-            colors = [colors[i, ...].float().permute(1, 2, 0) for i in range(colors.shape[0])]
-        else:
-            for i in range(len(colors)):
-                if isinstance(colors[i], Image.Image):
-                    colors[i] = torch.tensor(np.array(colors[i]) / 255.0, device=self.device).float()
-        if weights is None:
-            weights = [1.0 for _ in range(len(colors))]
-        textures = []
-        cos_maps = []
-        for color, elev, azim, weight in zip(colors, elevs, azims, weights):
-            texture, cos_map, _ = self.back_project(color, elev, azim, camera_distance, center)
-            cos_map = weight * (cos_map**exp)
-            textures.append(texture)
-            cos_maps.append(cos_map)
-
-        texture_merge, trust_map_merge = self.fast_bake_texture(textures, cos_maps)
-        return texture_merge, trust_map_merge
-
-    @torch.no_grad()
-    def fast_bake_texture(self, textures, cos_maps):
-        """
-        Efficiently merge multiple textures using cosine-weighted blending.
-        Optimizes by skipping views that don't contribute new information.
-        
-        Args:
-            textures: List of texture tensors to merge
-            cos_maps: List of corresponding cosine weight maps
-            
-        Returns:
-            Tuple of (merged_texture, valid_mask) tensors
-        """
-
-        channel = textures[0].shape[-1]
-        texture_merge = torch.zeros(self.texture_size + (channel,)).to(self.device)
-        trust_map_merge = torch.zeros(self.texture_size + (1,)).to(self.device)
-        for texture, cos_map in zip(textures, cos_maps):
-            view_sum = (cos_map > 0).sum()
-            painted_sum = ((cos_map > 0) * (trust_map_merge > 0)).sum()
-            if painted_sum / view_sum > 0.99:
-                continue
-            texture_merge += texture * cos_map
-            trust_map_merge += cos_map
-        texture_merge = texture_merge / torch.clamp(trust_map_merge, min=1e-8)
-
-        return texture_merge, trust_map_merge > 1e-8
-
-    @torch.no_grad()
-    def uv_inpaint(self, texture, mask, vertex_inpaint=True, method="NS", return_float=False):
-        """
-        Inpaint missing regions in UV texture using mesh-aware and traditional methods.
-        
-        Args:
-            texture: Input texture as tensor, numpy array, or PIL Image
-            mask: Binary mask indicating regions to inpaint (1 = keep, 0 = inpaint)
-            vertex_inpaint: Whether to use mesh vertex connectivity for inpainting
-            method: Inpainting method ("NS" for Navier-Stokes)
-            return_float: Whether to return float values (False returns uint8)
-            
-        Returns:
-            Inpainted texture as numpy array
-        """
-
-        if isinstance(texture, torch.Tensor):
-            texture_np = texture.cpu().numpy()
-        elif isinstance(texture, np.ndarray):
-            texture_np = texture
-        elif isinstance(texture, Image.Image):
-            texture_np = np.array(texture) / 255.0
-
-        if isinstance(mask, torch.Tensor):
-            mask = (mask.squeeze(-1).cpu().numpy() * 255).astype(np.uint8)
-
-        if vertex_inpaint:
-            vtx_pos, pos_idx, vtx_uv, uv_idx = self.get_mesh()
-            texture_np, mask = meshVerticeInpaint(texture_np, mask, vtx_pos, vtx_uv, pos_idx, uv_idx)
-
-        if method == "NS":
-            texture_np = cv2.inpaint((texture_np * 255).astype(np.uint8), 255 - mask, 3, cv2.INPAINT_NS)
-            assert return_float == False
-
-        return texture_np

hy3dpaint/DifferentiableRenderer/camera_utils.py

@@ -1,107 +0,0 @@
-# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT
-# except for the third-party components listed below.
-# Hunyuan 3D does not impose any additional limitations beyond what is outlined
-# in the repsective licenses of these third-party components.
-# Users must comply with all terms and conditions of original licenses of these third-party
-# components and must ensure that the usage of the third party components adheres to
-# all relevant laws and regulations.
-
-# For avoidance of doubts, Hunyuan 3D means the large language models and
-# their software and algorithms, including trained model weights, parameters (including
-# optimizer states), machine-learning model code, inference-enabling code, training-enabling code,
-# fine-tuning enabling code and other elements of the foregoing made publicly available
-# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT.
-
-import math
-
-import numpy as np
-import torch
-
-
-def transform_pos(mtx, pos, keepdim=False):
-    t_mtx = torch.from_numpy(mtx).to(pos.device) if isinstance(mtx, np.ndarray) else mtx
-    if pos.shape[-1] == 3:
-        posw = torch.cat([pos, torch.ones([pos.shape[0], 1]).to(pos.device)], axis=1)
-    else:
-        posw = pos
-
-    if keepdim:
-        return torch.matmul(posw, t_mtx.t())[...]
-    else:
-        return torch.matmul(posw, t_mtx.t())[None, ...]
-
-
-def get_mv_matrix(elev, azim, camera_distance, center=None):
-    elev = -elev
-    azim += 90
-
-    elev_rad = math.radians(elev)
-    azim_rad = math.radians(azim)
-
-    camera_position = np.array(
-        [
-            camera_distance * math.cos(elev_rad) * math.cos(azim_rad),
-            camera_distance * math.cos(elev_rad) * math.sin(azim_rad),
-            camera_distance * math.sin(elev_rad),
-        ]
-    )
-
-    if center is None:
-        center = np.array([0, 0, 0])
-    else:
-        center = np.array(center)
-
-    lookat = center - camera_position
-    lookat = lookat / np.linalg.norm(lookat)
-
-    up = np.array([0, 0, 1.0])
-    right = np.cross(lookat, up)
-    right = right / np.linalg.norm(right)
-    up = np.cross(right, lookat)
-    up = up / np.linalg.norm(up)
-
-    c2w = np.concatenate([np.stack([right, up, -lookat], axis=-1), camera_position[:, None]], axis=-1)
-
-    w2c = np.zeros((4, 4))
-    w2c[:3, :3] = np.transpose(c2w[:3, :3], (1, 0))
-    w2c[:3, 3:] = -np.matmul(np.transpose(c2w[:3, :3], (1, 0)), c2w[:3, 3:])
-    w2c[3, 3] = 1.0
-
-    return w2c.astype(np.float32)
-
-
-def get_orthographic_projection_matrix(left=-1, right=1, bottom=-1, top=1, near=0, far=2):
-    """
-    计算正交投影矩阵。
-
-    参数:
-        left (float): 投影区域左侧边界。
-        right (float): 投影区域右侧边界。
-        bottom (float): 投影区域底部边界。
-        top (float): 投影区域顶部边界。
-        near (float): 投影区域近裁剪面距离。
-        far (float): 投影区域远裁剪面距离。
-
-    返回:
-        numpy.ndarray: 正交投影矩阵。
-    """
-    ortho_matrix = np.eye(4, dtype=np.float32)
-    ortho_matrix[0, 0] = 2 / (right - left)
-    ortho_matrix[1, 1] = 2 / (top - bottom)
-    ortho_matrix[2, 2] = -2 / (far - near)
-    ortho_matrix[0, 3] = -(right + left) / (right - left)
-    ortho_matrix[1, 3] = -(top + bottom) / (top - bottom)
-    ortho_matrix[2, 3] = -(far + near) / (far - near)
-    return ortho_matrix
-
-
-def get_perspective_projection_matrix(fovy, aspect_wh, near, far):
-    fovy_rad = math.radians(fovy)
-    return np.array(
-        [
-            [1.0 / (math.tan(fovy_rad / 2.0) * aspect_wh), 0, 0, 0],
-            [0, 1.0 / math.tan(fovy_rad / 2.0), 0, 0],
-            [0, 0, -(far + near) / (far - near), -2.0 * far * near / (far - near)],
-            [0, 0, -1, 0],
-        ]
-    ).astype(np.float32)

hy3dpaint/DifferentiableRenderer/compile_mesh_painter.sh

@@ -1 +0,0 @@
-c++ -O3 -Wall -shared -std=c++11 -fPIC `python -m pybind11 --includes` mesh_inpaint_processor.cpp -o mesh_inpaint_processor`python3-config --extension-suffix`

hy3dpaint/DifferentiableRenderer/mesh_inpaint_processor.cpp

@@ -1,395 +0,0 @@
-#include <pybind11/numpy.h>
-#include <pybind11/pybind11.h>
-#include <pybind11/stl.h>
-
-#include <algorithm>
-#include <cmath>
-#include <queue>
-#include <vector>
-#include <functional>
-
-namespace py = pybind11;
-using namespace std;
-
-namespace {
-// 内部数据结构，避免重复的buffer获取和指针设置
-struct MeshData {
-    int texture_height, texture_width, texture_channel;
-    int vtx_num;
-    float* texture_ptr;
-    uint8_t* mask_ptr;
-    float* vtx_pos_ptr;
-    float* vtx_uv_ptr;
-    int* pos_idx_ptr;
-    int* uv_idx_ptr;
-    
-    // 存储buffer以防止被销毁
-    py::buffer_info texture_buf, mask_buf, vtx_pos_buf, vtx_uv_buf, pos_idx_buf, uv_idx_buf;
-    
-    MeshData(py::array_t<float>& texture, py::array_t<uint8_t>& mask, 
-             py::array_t<float>& vtx_pos, py::array_t<float>& vtx_uv,
-             py::array_t<int>& pos_idx, py::array_t<int>& uv_idx) {
-        
-        texture_buf = texture.request();
-        mask_buf = mask.request();
-        vtx_pos_buf = vtx_pos.request();
-        vtx_uv_buf = vtx_uv.request();
-        pos_idx_buf = pos_idx.request();
-        uv_idx_buf = uv_idx.request();
-
-        texture_height = texture_buf.shape[0];
-        texture_width = texture_buf.shape[1];
-        texture_channel = texture_buf.shape[2];
-        texture_ptr = static_cast<float*>(texture_buf.ptr);
-        mask_ptr = static_cast<uint8_t*>(mask_buf.ptr);
-
-        vtx_num = vtx_pos_buf.shape[0];
-        vtx_pos_ptr = static_cast<float*>(vtx_pos_buf.ptr);
-        vtx_uv_ptr = static_cast<float*>(vtx_uv_buf.ptr);
-        pos_idx_ptr = static_cast<int*>(pos_idx_buf.ptr);
-        uv_idx_ptr = static_cast<int*>(uv_idx_buf.ptr);
-    }
-};
-
-// 公共函数：计算UV坐标
-pair<int, int> calculateUVCoordinates(int vtx_uv_idx, const MeshData& data) {
-    int uv_v = round(data.vtx_uv_ptr[vtx_uv_idx * 2] * (data.texture_width - 1));
-    int uv_u = round((1.0 - data.vtx_uv_ptr[vtx_uv_idx * 2 + 1]) * (data.texture_height - 1));
-    return make_pair(uv_u, uv_v);
-}
-
-// 公共函数：计算距离权重
-float calculateDistanceWeight(const array<float, 3>& vtx_0, const array<float, 3>& vtx1) {
-    float dist_weight = 1.0f / max(
-        sqrt(
-            pow(vtx_0[0] - vtx1[0], 2) + 
-            pow(vtx_0[1] - vtx1[1], 2) + 
-            pow(vtx_0[2] - vtx1[2], 2)
-        ), 1E-4);
-    return dist_weight * dist_weight;
-}
-
-// 公共函数：获取顶点位置
-array<float, 3> getVertexPosition(int vtx_idx, const MeshData& data) {
-    return {data.vtx_pos_ptr[vtx_idx * 3], 
-            data.vtx_pos_ptr[vtx_idx * 3 + 1], 
-            data.vtx_pos_ptr[vtx_idx * 3 + 2]};
-}
-
-// 公共函数：构建图结构
-void buildGraph(vector<vector<int>>& G, const MeshData& data) {
-    G.resize(data.vtx_num);
-    for(int i = 0; i < data.uv_idx_buf.shape[0]; ++i) {
-        for(int k = 0; k < 3; ++k) {
-            G[data.pos_idx_ptr[i * 3 + k]].push_back(data.pos_idx_ptr[i * 3 + (k + 1) % 3]);
-        }
-    }
-}
-
-// 通用初始化函数：处理两种掩码类型（float和int）
-template<typename MaskType>
-void initializeVertexDataGeneric(const MeshData& data, vector<MaskType>& vtx_mask, 
-                                vector<vector<float>>& vtx_color, vector<int>* uncolored_vtxs = nullptr,
-                                MaskType mask_value = static_cast<MaskType>(1)) {
-    vtx_mask.assign(data.vtx_num, static_cast<MaskType>(0));
-    vtx_color.assign(data.vtx_num, vector<float>(data.texture_channel, 0.0f));
-    
-    if(uncolored_vtxs) {
-        uncolored_vtxs->clear();
-    }
-
-    for(int i = 0; i < data.uv_idx_buf.shape[0]; ++i) {
-        for(int k = 0; k < 3; ++k) {
-            int vtx_uv_idx = data.uv_idx_ptr[i * 3 + k];
-            int vtx_idx = data.pos_idx_ptr[i * 3 + k];
-            auto uv_coords = calculateUVCoordinates(vtx_uv_idx, data);
-
-            if(data.mask_ptr[uv_coords.first * data.texture_width + uv_coords.second] > 0) {
-                vtx_mask[vtx_idx] = mask_value;
-                for(int c = 0; c < data.texture_channel; ++c) {
-                    vtx_color[vtx_idx][c] = data.texture_ptr[(uv_coords.first * data.texture_width + 
-                                                            uv_coords.second) * data.texture_channel + c];
-                }
-            } else if(uncolored_vtxs) {
-                uncolored_vtxs->push_back(vtx_idx);
-            }
-        }
-    }
-}
-
-// 通用平滑算法：支持不同的掩码类型和检查函数
-template<typename MaskType>
-void performSmoothingAlgorithm(const MeshData& data, const vector<vector<int>>& G,
-                              vector<MaskType>& vtx_mask, vector<vector<float>>& vtx_color, 
-                              const vector<int>& uncolored_vtxs,
-                              function<bool(MaskType)> is_colored_func,
-                              function<void(MaskType&)> set_colored_func) {
-    int smooth_count = 2;
-    int last_uncolored_vtx_count = 0;
-    
-    while(smooth_count > 0) {
-        int uncolored_vtx_count = 0;
-
-        for(int vtx_idx : uncolored_vtxs) {
-            vector<float> sum_color(data.texture_channel, 0.0f);
-            float total_weight = 0.0f;
-
-            array<float, 3> vtx_0 = getVertexPosition(vtx_idx, data);
-            
-            for(int connected_idx : G[vtx_idx]) {
-                if(is_colored_func(vtx_mask[connected_idx])) {
-                    array<float, 3> vtx1 = getVertexPosition(connected_idx, data);
-                    float dist_weight = calculateDistanceWeight(vtx_0, vtx1);
-                    
-                    for(int c = 0; c < data.texture_channel; ++c) {
-                        sum_color[c] += vtx_color[connected_idx][c] * dist_weight;
-                    }
-                    total_weight += dist_weight;
-                }
-            }
-
-            if(total_weight > 0.0f) {
-                for(int c = 0; c < data.texture_channel; ++c) {
-                    vtx_color[vtx_idx][c] = sum_color[c] / total_weight;
-                }
-                set_colored_func(vtx_mask[vtx_idx]);
-            } else {
-                uncolored_vtx_count++;
-            }
-        }
-
-        if(last_uncolored_vtx_count == uncolored_vtx_count) {
-            smooth_count--;
-        } else {
-            smooth_count++;
-        }
-        last_uncolored_vtx_count = uncolored_vtx_count;
-    }
-}
-
-// 前向传播算法的通用实现
-void performForwardPropagation(const MeshData& data, const vector<vector<int>>& G,
-                              vector<float>& vtx_mask, vector<vector<float>>& vtx_color,
-                              queue<int>& active_vtxs) {
-    while(!active_vtxs.empty()) {
-        queue<int> pending_active_vtxs;
-        
-        while(!active_vtxs.empty()) {
-            int vtx_idx = active_vtxs.front();
-            active_vtxs.pop();
-            array<float, 3> vtx_0 = getVertexPosition(vtx_idx, data);
-            
-            for(int connected_idx : G[vtx_idx]) {
-                if(vtx_mask[connected_idx] > 0) continue;
-                
-                array<float, 3> vtx1 = getVertexPosition(connected_idx, data);
-                float dist_weight = calculateDistanceWeight(vtx_0, vtx1);
-                
-                for(int c = 0; c < data.texture_channel; ++c) {
-                    vtx_color[connected_idx][c] += vtx_color[vtx_idx][c] * dist_weight;
-                }
-                
-                if(vtx_mask[connected_idx] == 0) {
-                    pending_active_vtxs.push(connected_idx);
-                }
-                vtx_mask[connected_idx] -= dist_weight;
-            }
-        }
-
-        while(!pending_active_vtxs.empty()) {
-            int vtx_idx = pending_active_vtxs.front();
-            pending_active_vtxs.pop();
-            
-            for(int c = 0; c < data.texture_channel; ++c) {
-                vtx_color[vtx_idx][c] /= -vtx_mask[vtx_idx];
-            }
-            vtx_mask[vtx_idx] = 1.0f;
-            active_vtxs.push(vtx_idx);
-        }
-    }
-}
-
-// 公共函数：创建输出数组
-pair<py::array_t<float>, py::array_t<uint8_t>> createOutputArrays(
-    const MeshData& data, const vector<float>& vtx_mask, 
-    const vector<vector<float>>& vtx_color) {
-    
-    py::array_t<float> new_texture(data.texture_buf.size);
-    py::array_t<uint8_t> new_mask(data.mask_buf.size);
-
-    auto new_texture_buf = new_texture.request();
-    auto new_mask_buf = new_mask.request();
-
-    float* new_texture_ptr = static_cast<float*>(new_texture_buf.ptr);
-    uint8_t* new_mask_ptr = static_cast<uint8_t*>(new_mask_buf.ptr);
-    
-    // Copy original texture and mask to new arrays
-    copy(data.texture_ptr, data.texture_ptr + data.texture_buf.size, new_texture_ptr);
-    copy(data.mask_ptr, data.mask_ptr + data.mask_buf.size, new_mask_ptr);
-
-    for(int face_idx = 0; face_idx < data.uv_idx_buf.shape[0]; ++face_idx) {
-        for(int k = 0; k < 3; ++k) {
-            int vtx_uv_idx = data.uv_idx_ptr[face_idx * 3 + k];
-            int vtx_idx = data.pos_idx_ptr[face_idx * 3 + k];
-
-            if(vtx_mask[vtx_idx] == 1.0f) {
-                auto uv_coords = calculateUVCoordinates(vtx_uv_idx, data);
-                
-                for(int c = 0; c < data.texture_channel; ++c) {
-                    new_texture_ptr[
-                        (uv_coords.first * data.texture_width + uv_coords.second) * 
-                        data.texture_channel + c
-                    ] = vtx_color[vtx_idx][c];
-                }
-                new_mask_ptr[uv_coords.first * data.texture_width + uv_coords.second] = 255;
-            }
-        }
-    }
-
-    // Reshape the new arrays to match the original texture and mask shapes
-    new_texture.resize({data.texture_height, data.texture_width, 3});
-    new_mask.resize({data.texture_height, data.texture_width});
-
-    return make_pair(new_texture, new_mask);
-}
-
-// 创建顶点颜色输出数组的专用函数
-pair<py::array_t<float>, py::array_t<uint8_t>> createVertexColorOutput(
-    const MeshData& data, const vector<int>& vtx_mask, 
-    const vector<vector<float>>& vtx_color) {
-    
-    py::array_t<float> py_vtx_color({data.vtx_num, data.texture_channel});
-    py::array_t<uint8_t> py_vtx_mask({data.vtx_num});
-
-    auto py_vtx_color_buf = py_vtx_color.request();
-    auto py_vtx_mask_buf = py_vtx_mask.request();
-
-    float* py_vtx_color_ptr = static_cast<float*>(py_vtx_color_buf.ptr);
-    uint8_t* py_vtx_mask_ptr = static_cast<uint8_t*>(py_vtx_mask_buf.ptr);
-
-    for(int i = 0; i < data.vtx_num; ++i) {
-        py_vtx_mask_ptr[i] = vtx_mask[i];
-        for(int c = 0; c < data.texture_channel; ++c) {
-            py_vtx_color_ptr[i * data.texture_channel + c] = vtx_color[i][c];
-        }
-    }
-
-    return make_pair(py_vtx_color, py_vtx_mask);
-}
-
-} // anonymous namespace
-
-// 重构后的 meshVerticeInpaint_smooth 函数
-pair<py::array_t<float>, py::array_t<uint8_t>> meshVerticeInpaint_smooth(
-    py::array_t<float> texture, py::array_t<uint8_t> mask, py::array_t<float> vtx_pos, py::array_t<float> vtx_uv,
-    py::array_t<int> pos_idx, py::array_t<int> uv_idx) {
-    
-    MeshData data(texture, mask, vtx_pos, vtx_uv, pos_idx, uv_idx);
-    
-    vector<float> vtx_mask;
-    vector<vector<float>> vtx_color;
-    vector<int> uncolored_vtxs;
-    vector<vector<int>> G;
-
-    initializeVertexDataGeneric(data, vtx_mask, vtx_color, &uncolored_vtxs, 1.0f);
-    buildGraph(G, data);
-
-    // 使用通用平滑算法
-    performSmoothingAlgorithm<float>(data, G, vtx_mask, vtx_color, uncolored_vtxs,
-        [](float mask_val) { return mask_val > 0; },  // 检查是否着色
-        [](float& mask_val) { mask_val = 1.0f; }      // 设置为已着色
-    );
-
-    return createOutputArrays(data, vtx_mask, vtx_color);
-}
-
-// 重构后的 meshVerticeInpaint_forward 函数
-pair<py::array_t<float>, py::array_t<uint8_t>> meshVerticeInpaint_forward(
-    py::array_t<float> texture, py::array_t<uint8_t> mask, py::array_t<float> vtx_pos, py::array_t<float> vtx_uv,
-    py::array_t<int> pos_idx, py::array_t<int> uv_idx) {
-    
-    MeshData data(texture, mask, vtx_pos, vtx_uv, pos_idx, uv_idx);
-    
-    vector<float> vtx_mask;
-    vector<vector<float>> vtx_color;
-    vector<vector<int>> G;
-    queue<int> active_vtxs;
-
-    // 使用通用初始化（不需要 uncolored_vtxs）
-    initializeVertexDataGeneric(data, vtx_mask, vtx_color, nullptr, 1.0f);
-    buildGraph(G, data);
-
-    // 收集活跃顶点
-    for(int i = 0; i < data.vtx_num; ++i) {
-        if(vtx_mask[i] == 1.0f) {
-            active_vtxs.push(i);
-        }
-    }
-
-    // 使用通用前向传播算法
-    performForwardPropagation(data, G, vtx_mask, vtx_color, active_vtxs);
-
-    return createOutputArrays(data, vtx_mask, vtx_color);
-}
-
-// 主接口函数
-pair<py::array_t<float>, py::array_t<uint8_t>> meshVerticeInpaint(
-    py::array_t<float> texture, py::array_t<uint8_t> mask, py::array_t<float> vtx_pos, py::array_t<float> vtx_uv,
-    py::array_t<int> pos_idx, py::array_t<int> uv_idx, const string& method = "smooth") {
-    
-    if(method == "smooth") {
-        return meshVerticeInpaint_smooth(texture, mask, vtx_pos, vtx_uv, pos_idx, uv_idx);
-    } else if(method == "forward") {
-        return meshVerticeInpaint_forward(texture, mask, vtx_pos, vtx_uv, pos_idx, uv_idx);
-    } else {
-        throw invalid_argument("Invalid method. Use 'smooth' or 'forward'.");
-    }
-}
-
-//============================
-
-// 重构后的 meshVerticeColor_smooth 函数
-pair<py::array_t<float>, py::array_t<uint8_t>> meshVerticeColor_smooth(
-    py::array_t<float> texture, py::array_t<uint8_t> mask, py::array_t<float> vtx_pos, py::array_t<float> vtx_uv,
-    py::array_t<int> pos_idx, py::array_t<int> uv_idx) {
-    
-    MeshData data(texture, mask, vtx_pos, vtx_uv, pos_idx, uv_idx);
-    
-    vector<int> vtx_mask;
-    vector<vector<float>> vtx_color;
-    vector<int> uncolored_vtxs;
-    vector<vector<int>> G;
-
-    initializeVertexDataGeneric(data, vtx_mask, vtx_color, &uncolored_vtxs, 1);
-    buildGraph(G, data);
-
-    // 使用通用平滑算法
-    performSmoothingAlgorithm<int>(data, G, vtx_mask, vtx_color, uncolored_vtxs,
-        [](int mask_val) { return mask_val > 0; },    // 检查是否着色
-        [](int& mask_val) { mask_val = 2; }           // 设置为已着色（值为2）
-    );
-
-    return createVertexColorOutput(data, vtx_mask, vtx_color);
-}
-
-// meshVerticeColor 主接口函数
-pair<py::array_t<float>, py::array_t<uint8_t>> meshVerticeColor(
-    py::array_t<float> texture, py::array_t<uint8_t> mask, py::array_t<float> vtx_pos, py::array_t<float> vtx_uv,
-    py::array_t<int> pos_idx, py::array_t<int> uv_idx, const string& method = "smooth") {
-    
-    if(method == "smooth") {
-        return meshVerticeColor_smooth(texture, mask, vtx_pos, vtx_uv, pos_idx, uv_idx);
-    } else {
-        throw invalid_argument("Invalid method. Use 'smooth' or 'forward'.");
-    }
-}
-
-// Python绑定
-PYBIND11_MODULE(mesh_inpaint_processor, m) {
-    m.def("meshVerticeInpaint", &meshVerticeInpaint, "A function to process mesh", 
-          py::arg("texture"), py::arg("mask"), py::arg("vtx_pos"), py::arg("vtx_uv"), 
-          py::arg("pos_idx"), py::arg("uv_idx"), py::arg("method") = "smooth");
-    m.def("meshVerticeColor", &meshVerticeColor, "A function to process mesh", 
-          py::arg("texture"), py::arg("mask"), py::arg("vtx_pos"), py::arg("vtx_uv"), 
-          py::arg("pos_idx"), py::arg("uv_idx"), py::arg("method") = "smooth");
-} 

hy3dpaint/DifferentiableRenderer/mesh_utils.py

@@ -1,284 +0,0 @@
-# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT
-# except for the third-party components listed below.
-# Hunyuan 3D does not impose any additional limitations beyond what is outlined
-# in the repsective licenses of these third-party components.
-# Users must comply with all terms and conditions of original licenses of these third-party
-# components and must ensure that the usage of the third party components adheres to
-# all relevant laws and regulations.
-
-# For avoidance of doubts, Hunyuan 3D means the large language models and
-# their software and algorithms, including trained model weights, parameters (including
-# optimizer states), machine-learning model code, inference-enabling code, training-enabling code,
-# fine-tuning enabling code and other elements of the foregoing made publicly available
-# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT.
-
-import os
-import cv2
-import bpy
-import math
-import numpy as np
-from io import StringIO
-from typing import Optional, Tuple, Dict, Any
-
-
-def _safe_extract_attribute(obj: Any, attr_path: str, default: Any = None) -> Any:
-    """Extract nested attribute safely from object."""
-    try:
-        for attr in attr_path.split("."):
-            obj = getattr(obj, attr)
-        return obj
-    except AttributeError:
-        return default
-
-
-def _convert_to_numpy(data: Any, dtype: np.dtype) -> Optional[np.ndarray]:
-    """Convert data to numpy array with specified dtype, handling None values."""
-    if data is None:
-        return None
-    return np.asarray(data, dtype=dtype)
-
-
-def load_mesh(mesh):
-    """Load mesh data including vertices, faces, UV coordinates and texture."""
-    # Extract vertex positions and face indices
-    vtx_pos = _safe_extract_attribute(mesh, "vertices")
-    pos_idx = _safe_extract_attribute(mesh, "faces")
-
-    # Extract UV coordinates (reusing face indices for UV indices)
-    vtx_uv = _safe_extract_attribute(mesh, "visual.uv")
-    uv_idx = pos_idx  # Reuse face indices for UV mapping
-
-    # Convert to numpy arrays with appropriate dtypes
-    vtx_pos = _convert_to_numpy(vtx_pos, np.float32)
-    pos_idx = _convert_to_numpy(pos_idx, np.int32)
-    vtx_uv = _convert_to_numpy(vtx_uv, np.float32)
-    uv_idx = _convert_to_numpy(uv_idx, np.int32)
-
-    texture_data = None
-    return vtx_pos, pos_idx, vtx_uv, uv_idx, texture_data
-
-
-def _get_base_path_and_name(mesh_path: str) -> Tuple[str, str]:
-    """Get base path without extension and mesh name."""
-    base_path = os.path.splitext(mesh_path)[0]
-    name = os.path.basename(base_path)
-    return base_path, name
-
-
-def _save_texture_map(
-    texture: np.ndarray,
-    base_path: str,
-    suffix: str = "",
-    image_format: str = ".jpg",
-    color_convert: Optional[int] = None,
-) -> str:
-    """Save texture map with optional color conversion."""
-    path = f"{base_path}{suffix}{image_format}"
-    processed_texture = (texture * 255).astype(np.uint8)
-
-    if color_convert is not None:
-        processed_texture = cv2.cvtColor(processed_texture, color_convert)
-        cv2.imwrite(path, processed_texture)
-    else:
-        cv2.imwrite(path, processed_texture[..., ::-1])  # RGB to BGR
-
-    return os.path.basename(path)
-
-
-def _write_mtl_properties(f, properties: Dict[str, Any]):
-    """Write material properties to MTL file."""
-    for key, value in properties.items():
-        if isinstance(value, (list, tuple)):
-            f.write(f"{key} {' '.join(map(str, value))}\n")
-        else:
-            f.write(f"{key} {value}\n")
-
-
-def _create_obj_content(
-    vtx_pos: np.ndarray, vtx_uv: np.ndarray, pos_idx: np.ndarray, uv_idx: np.ndarray, name: str
-) -> str:
-    """Create OBJ file content."""
-    buffer = StringIO()
-
-    # Write header and vertices
-    buffer.write(f"mtllib {name}.mtl\no {name}\n")
-    np.savetxt(buffer, vtx_pos, fmt="v %.6f %.6f %.6f")
-    np.savetxt(buffer, vtx_uv, fmt="vt %.6f %.6f")
-    buffer.write("s 0\nusemtl Material\n")
-
-    # Write faces
-    pos_idx_plus1 = pos_idx + 1
-    uv_idx_plus1 = uv_idx + 1
-    face_format = np.frompyfunc(lambda *x: f"{int(x[0])}/{int(x[1])}", 2, 1)
-    faces = face_format(pos_idx_plus1, uv_idx_plus1)
-    face_strings = [f"f {' '.join(face)}" for face in faces]
-    buffer.write("\n".join(face_strings) + "\n")
-
-    return buffer.getvalue()
-
-
-def save_obj_mesh(mesh_path, vtx_pos, pos_idx, vtx_uv, uv_idx, texture, metallic=None, roughness=None, normal=None):
-    """Save mesh as OBJ file with textures and material."""
-    # Convert inputs to numpy arrays
-    vtx_pos = _convert_to_numpy(vtx_pos, np.float32)
-    vtx_uv = _convert_to_numpy(vtx_uv, np.float32)
-    pos_idx = _convert_to_numpy(pos_idx, np.int32)
-    uv_idx = _convert_to_numpy(uv_idx, np.int32)
-
-    base_path, name = _get_base_path_and_name(mesh_path)
-
-    # Create and save OBJ content
-    obj_content = _create_obj_content(vtx_pos, vtx_uv, pos_idx, uv_idx, name)
-    with open(mesh_path, "w") as obj_file:
-        obj_file.write(obj_content)
-
-    # Save texture maps
-    texture_maps = {}
-    texture_maps["diffuse"] = _save_texture_map(texture, base_path)
-
-    if metallic is not None:
-        texture_maps["metallic"] = _save_texture_map(metallic, base_path, "_metallic", color_convert=cv2.COLOR_RGB2GRAY)
-    if roughness is not None:
-        texture_maps["roughness"] = _save_texture_map(
-            roughness, base_path, "_roughness", color_convert=cv2.COLOR_RGB2GRAY
-        )
-    if normal is not None:
-        texture_maps["normal"] = _save_texture_map(normal, base_path, "_normal")
-
-    # Create MTL file
-    _create_mtl_file(base_path, texture_maps, metallic is not None)
-
-
-def _create_mtl_file(base_path: str, texture_maps: Dict[str, str], is_pbr: bool):
-    """Create MTL material file."""
-    mtl_path = f"{base_path}.mtl"
-
-    with open(mtl_path, "w") as f:
-        f.write("newmtl Material\n")
-
-        if is_pbr:
-            # PBR material properties
-            properties = {
-                "Kd": [0.800, 0.800, 0.800],
-                "Ke": [0.000, 0.000, 0.000],  # 鐜鍏夐伄钄�
-                "Ni": 1.500,  # 鎶樺皠绯绘暟
-                "d": 1.0,  # 閫忔槑搴�
-                "illum": 2,  # 鍏夌収妯″瀷
-                "map_Kd": texture_maps["diffuse"],
-            }
-            _write_mtl_properties(f, properties)
-
-            # Additional PBR maps
-            map_configs = [("metallic", "map_Pm"), ("roughness", "map_Pr"), ("normal", "map_Bump -bm 1.0")]
-
-            for texture_key, mtl_key in map_configs:
-                if texture_key in texture_maps:
-                    f.write(f"{mtl_key} {texture_maps[texture_key]}\n")
-        else:
-            # Standard material properties
-            properties = {
-                "Ns": 250.000000,
-                "Ka": [0.200, 0.200, 0.200],
-                "Kd": [0.800, 0.800, 0.800],
-                "Ks": [0.500, 0.500, 0.500],
-                "Ke": [0.000, 0.000, 0.000],
-                "Ni": 1.500,
-                "d": 1.0,
-                "illum": 3,
-                "map_Kd": texture_maps["diffuse"],
-            }
-            _write_mtl_properties(f, properties)
-
-
-def save_mesh(mesh_path, vtx_pos, pos_idx, vtx_uv, uv_idx, texture, metallic=None, roughness=None, normal=None):
-    """Save mesh using OBJ format."""
-    save_obj_mesh(
-        mesh_path, vtx_pos, pos_idx, vtx_uv, uv_idx, texture, metallic=metallic, roughness=roughness, normal=normal
-    )
-
-
-def _setup_blender_scene():
-    """Setup Blender scene for conversion."""
-    if "convert" not in bpy.data.scenes:
-        bpy.data.scenes.new("convert")
-    bpy.context.window.scene = bpy.data.scenes["convert"]
-
-
-def _clear_scene_objects():
-    """Clear all objects from current Blender scene."""
-    for obj in bpy.context.scene.objects:
-        obj.select_set(True)
-        bpy.data.objects.remove(obj, do_unlink=True)
-
-
-def _select_mesh_objects():
-    """Select all mesh objects in scene."""
-    bpy.ops.object.select_all(action="DESELECT")
-    for obj in bpy.context.scene.objects:
-        if obj.type == "MESH":
-            obj.select_set(True)
-
-
-def _merge_vertices_if_needed(merge_vertices: bool):
-    """Merge duplicate vertices if requested."""
-    if not merge_vertices:
-        return
-
-    for obj in bpy.context.selected_objects:
-        if obj.type == "MESH":
-            bpy.context.view_layer.objects.active = obj
-            bpy.ops.object.mode_set(mode="EDIT")
-            bpy.ops.mesh.select_all(action="SELECT")
-            bpy.ops.mesh.remove_doubles()
-            bpy.ops.object.mode_set(mode="OBJECT")
-
-
-def _apply_shading(shade_type: str, auto_smooth_angle: float):
-    """Apply shading to selected objects."""
-    shading_ops = {
-        "SMOOTH": lambda: bpy.ops.object.shade_smooth(),
-        "FLAT": lambda: bpy.ops.object.shade_flat(),
-        "AUTO_SMOOTH": lambda: _apply_auto_smooth(auto_smooth_angle),
-    }
-
-    if shade_type in shading_ops:
-        shading_ops[shade_type]()
-
-
-def _apply_auto_smooth(auto_smooth_angle: float):
-    """Apply auto smooth based on Blender version."""
-    angle_rad = math.radians(auto_smooth_angle)
-
-    if bpy.app.version < (4, 1, 0):
-        bpy.ops.object.shade_smooth(use_auto_smooth=True, auto_smooth_angle=angle_rad)
-    elif bpy.app.version < (4, 2, 0):
-        bpy.ops.object.shade_smooth_by_angle(angle=angle_rad)
-    else:
-        bpy.ops.object.shade_auto_smooth(angle=angle_rad)
-
-
-def convert_obj_to_glb(
-    obj_path: str,
-    glb_path: str,
-    shade_type: str = "SMOOTH",
-    auto_smooth_angle: float = 60,
-    merge_vertices: bool = False,
-) -> bool:
-    """Convert OBJ file to GLB format using Blender."""
-    try:
-        _setup_blender_scene()
-        _clear_scene_objects()
-
-        # Import OBJ file
-        bpy.ops.wm.obj_import(filepath=obj_path)
-        _select_mesh_objects()
-
-        # Process meshes
-        _merge_vertices_if_needed(merge_vertices)
-        _apply_shading(shade_type, auto_smooth_angle)
-
-        # Export to GLB
-        bpy.ops.export_scene.gltf(filepath=glb_path, use_active_scene=True)
-        return True
-    except Exception:
-        return False

hy3dpaint/LICENSE

@@ -1,81 +0,0 @@
-TENCENT HUNYUAN 3D 2.1 COMMUNITY LICENSE AGREEMENT
-Tencent Hunyuan 3D 2.1 Release Date: June 13, 2025
-THIS LICENSE AGREEMENT DOES NOT APPLY IN THE EUROPEAN UNION, UNITED KINGDOM AND SOUTH KOREA AND IS EXPRESSLY LIMITED TO THE TERRITORY, AS DEFINED BELOW.
-By clicking to agree or by using, reproducing, modifying, distributing, performing or displaying any portion or element of the Tencent Hunyuan 3D 2.1 Works, including via any Hosted Service, You will be deemed to have recognized and accepted the content of this Agreement, which is effective immediately.
-1.	DEFINITIONS.
-a.	“Acceptable Use Policy” shall mean the policy made available by Tencent as set forth in the Exhibit A.
-b.	“Agreement” shall mean the terms and conditions for use, reproduction, distribution, modification, performance and displaying of Tencent Hunyuan 3D 2.1 Works or any portion or element thereof set forth herein.
-c.	“Documentation” shall mean the specifications, manuals and documentation for Tencent Hunyuan 3D 2.1 made publicly available by Tencent.
-d.	“Hosted Service” shall mean a hosted service offered via an application programming interface (API), web access, or any other electronic or remote means.
-e.	“Licensee,” “You” or “Your” shall mean a natural person or legal entity exercising the rights granted by this Agreement and/or using the Tencent Hunyuan 3D 2.1 Works for any purpose and in any field of use.
-f.	“Materials” shall mean, collectively, Tencent’s proprietary Tencent Hunyuan 3D 2.1 and Documentation (and any portion thereof) as made available by Tencent under this Agreement.
-g.	“Model Derivatives” shall mean all: (i) modifications to Tencent Hunyuan 3D 2.1 or any Model Derivative of Tencent Hunyuan 3D 2.1; (ii) works based on Tencent Hunyuan 3D 2.1 or any Model Derivative of Tencent Hunyuan 3D 2.1; or (iii) any other machine learning model which is created by transfer of patterns of the weights, parameters, operations, or Output of Tencent Hunyuan 3D 2.1 or any Model Derivative of Tencent Hunyuan 3D 2.1, to that model in order to cause that model to perform similarly to Tencent Hunyuan 3D 2.1 or a Model Derivative of Tencent Hunyuan 3D 2.1, including distillation methods, methods that use intermediate data representations, or methods based on the generation of synthetic data Outputs by Tencent Hunyuan 3D 2.1 or a Model Derivative of Tencent Hunyuan 3D 2.1 for training that model. For clarity, Outputs by themselves are not deemed Model Derivatives.
-h.	“Output” shall mean the information and/or content output of Tencent Hunyuan 3D 2.1 or a Model Derivative that results from operating or otherwise using Tencent Hunyuan 3D 2.1 or a Model Derivative, including via a Hosted Service.
-i.	“Tencent,” “We” or “Us” shall mean THL Q Limited.
-j.	“Tencent Hunyuan 3D 2.1” shall mean the 3D generation models and their software and algorithms, including trained model weights, parameters (including optimizer states), machine-learning model code, inference-enabling code, training-enabling code, fine-tuning enabling code and other elements of the foregoing made publicly available by Us at [ https://github.com/Tencent-Hunyuan/Hunyuan3D-2.1].
-k.	“Tencent Hunyuan 3D 2.1 Works” shall mean: (i) the Materials; (ii) Model Derivatives; and (iii) all derivative works thereof.
-l.	“Territory” shall mean the worldwide territory, excluding the territory of the European Union, United Kingdom and South Korea. 
-m.	“Third Party” or “Third Parties” shall mean individuals or legal entities that are not under common control with Us or You.
-n.	“including” shall mean including but not limited to.
-2.	GRANT OF RIGHTS.
-We grant You, for the Territory only, a non-exclusive, non-transferable and royalty-free limited license under Tencent’s intellectual property or other rights owned by Us embodied in or utilized by the Materials to use, reproduce, distribute, create derivative works of (including Model Derivatives), and make modifications to the Materials, only in accordance with the terms of this Agreement and the Acceptable Use Policy, and You must not violate (or encourage or permit anyone else to violate) any term of this Agreement or the Acceptable Use Policy.
-3.	DISTRIBUTION.
-You may, subject to Your compliance with this Agreement, distribute or make available to Third Parties the Tencent Hunyuan 3D 2.1 Works, exclusively in the Territory, provided that You meet all of the following conditions:
-a.	You must provide all such Third Party recipients of the Tencent Hunyuan 3D 2.1 Works or products or services using them a copy of this Agreement;
-b.	You must cause any modified files to carry prominent notices stating that You changed the files;
-c.	You are encouraged to: (i) publish at least one technology introduction blogpost or one public statement expressing Your experience of using the Tencent Hunyuan 3D 2.1 Works; and (ii) mark the products or services developed by using the Tencent Hunyuan 3D 2.1 Works to indicate that the product/service is “Powered by Tencent Hunyuan”; and
-d.	All distributions to Third Parties (other than through a Hosted Service) must be accompanied by a “Notice” text file that contains the following notice: “Tencent Hunyuan 3D 2.1 is licensed under the Tencent Hunyuan 3D 2.1 Community License Agreement, Copyright © 2025 Tencent. All Rights Reserved. The trademark rights of “Tencent Hunyuan” are owned by Tencent or its affiliate.”
-You may add Your own copyright statement to Your modifications and, except as set forth in this Section and in Section 5, may provide additional or different license terms and conditions for use, reproduction, or distribution of Your modifications, or for any such Model Derivatives as a whole, provided Your use, reproduction, modification, distribution, performance and display of the work otherwise complies with the terms and conditions of this Agreement (including as regards the Territory). If You receive Tencent Hunyuan 3D 2.1 Works from a Licensee as part of an integrated end user product, then this Section 3 of this Agreement will not apply to You.
-4.	ADDITIONAL COMMERCIAL TERMS.
-If, on the Tencent Hunyuan 3D 2.1 version release date, the monthly active users of all products or services made available by or for Licensee is greater than 1 million monthly active users in the preceding calendar month, You must request a license from Tencent, which Tencent may grant to You in its sole discretion, and You are not authorized to exercise any of the rights under this Agreement unless or until Tencent otherwise expressly grants You such rights. 
-Subject to Tencent's written approval, you may request a license for the use of Tencent Hunyuan 3D 2.1 by submitting the following information to [email protected]:
-a.	Your company’s name and associated business sector that plans to use Tencent Hunyuan 3D 2.1.
-b.	Your intended use case and the purpose of using Tencent Hunyuan 3D 2.1.
-c.	Your plans to modify Tencent Hunyuan 3D 2.1 or create Model Derivatives.
-5.	RULES OF USE.
-a.	Your use of the Tencent Hunyuan 3D 2.1 Works must comply with applicable laws and regulations (including trade compliance laws and regulations) and adhere to the Acceptable Use Policy for the Tencent Hunyuan 3D 2.1 Works, which is hereby incorporated by reference into this Agreement. You must include the use restrictions referenced in these Sections 5(a) and 5(b) as an enforceable provision in any agreement (e.g., license agreement, terms of use, etc.) governing the use and/or distribution of Tencent Hunyuan 3D 2.1 Works and You must provide notice to subsequent users to whom You distribute that Tencent Hunyuan 3D 2.1 Works are subject to the use restrictions in these Sections 5(a) and 5(b).
-b.	You must not use the Tencent Hunyuan 3D 2.1 Works or any Output or results of the Tencent Hunyuan 3D 2.1 Works to improve any other AI model (other than Tencent Hunyuan 3D 2.1 or Model Derivatives thereof).
-c.	You must not use, reproduce, modify, distribute, or display the Tencent Hunyuan 3D 2.1 Works, Output or results of the Tencent Hunyuan 3D 2.1 Works outside the Territory. Any such use outside the Territory is unlicensed and unauthorized under this Agreement.
-6.	INTELLECTUAL PROPERTY.
-a.	Subject to Tencent’s ownership of Tencent Hunyuan 3D 2.1 Works made by or for Tencent and intellectual property rights therein, conditioned upon Your compliance with the terms and conditions of this Agreement, as between You and Tencent, You will be the owner of any derivative works and modifications of the Materials and any Model Derivatives that are made by or for You.
-b.	No trademark licenses are granted under this Agreement, and in connection with the Tencent Hunyuan 3D 2.1 Works, Licensee may not use any name or mark owned by or associated with Tencent or any of its affiliates, except as required for reasonable and customary use in describing and distributing the Tencent Hunyuan 3D 2.1 Works. Tencent hereby grants You a license to use “Tencent Hunyuan” (the “Mark”) in the Territory solely as required to comply with the provisions of Section 3(c), provided that You comply with any applicable laws related to trademark protection. All goodwill arising out of Your use of the Mark will inure to the benefit of Tencent.
-c.	If You commence a lawsuit or other proceedings (including a cross-claim or counterclaim in a lawsuit) against Us or any person or entity alleging that the Materials or any Output, or any portion of any of the foregoing, infringe any intellectual property or other right owned or licensable by You, then all licenses granted to You under this Agreement shall terminate as of the date such lawsuit or other proceeding is filed. You will defend, indemnify and hold harmless Us from and against any claim by any Third Party arising out of or related to Your or the Third Party’s use or distribution of the Tencent Hunyuan 3D 2.1 Works.
-d.	Tencent claims no rights in Outputs You generate. You and Your users are solely responsible for Outputs and their subsequent uses.
-7.	DISCLAIMERS OF WARRANTY AND LIMITATIONS OF LIABILITY.
-a.	We are not obligated to support, update, provide training for, or develop any further version of the Tencent Hunyuan 3D 2.1 Works or to grant any license thereto.
-b.	UNLESS AND ONLY TO THE EXTENT REQUIRED BY APPLICABLE LAW, THE TENCENT HUNYUAN 3D 2.1 WORKS AND ANY OUTPUT AND RESULTS THEREFROM ARE PROVIDED “AS IS” WITHOUT ANY EXPRESS OR IMPLIED WARRANTIES OF ANY KIND INCLUDING ANY WARRANTIES OF TITLE, MERCHANTABILITY, NONINFRINGEMENT, COURSE OF DEALING, USAGE OF TRADE, OR FITNESS FOR A PARTICULAR PURPOSE. YOU ARE SOLELY RESPONSIBLE FOR DETERMINING THE APPROPRIATENESS OF USING, REPRODUCING, MODIFYING, PERFORMING, DISPLAYING OR DISTRIBUTING ANY OF THE TENCENT HUNYUAN 3D 2.1 WORKS OR OUTPUTS AND ASSUME ANY AND ALL RISKS ASSOCIATED WITH YOUR OR A THIRD PARTY’S USE OR DISTRIBUTION OF ANY OF THE TENCENT HUNYUAN 3D 2.1 WORKS OR OUTPUTS AND YOUR EXERCISE OF RIGHTS AND PERMISSIONS UNDER THIS AGREEMENT.
-c.	TO THE FULLEST EXTENT PERMITTED BY APPLICABLE LAW, IN NO EVENT SHALL TENCENT OR ITS AFFILIATES BE LIABLE UNDER ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, TORT, NEGLIGENCE, PRODUCTS LIABILITY, OR OTHERWISE, FOR ANY DAMAGES, INCLUDING ANY DIRECT, INDIRECT, SPECIAL, INCIDENTAL, EXEMPLARY, CONSEQUENTIAL OR PUNITIVE DAMAGES, OR LOST PROFITS OF ANY KIND ARISING FROM THIS AGREEMENT OR RELATED TO ANY OF THE TENCENT HUNYUAN 3D 2.1 WORKS OR OUTPUTS, EVEN IF TENCENT OR ITS AFFILIATES HAVE BEEN ADVISED OF THE POSSIBILITY OF ANY OF THE FOREGOING.
-8.	SURVIVAL AND TERMINATION.
-a.	The term of this Agreement shall commence upon Your acceptance of this Agreement or access to the Materials and will continue in full force and effect until terminated in accordance with the terms and conditions herein.
-b.	We may terminate this Agreement if You breach any of the terms or conditions of this Agreement. Upon termination of this Agreement, You must promptly delete and cease use of the Tencent Hunyuan 3D 2.1 Works. Sections 6(a), 6(c), 7 and 9 shall survive the termination of this Agreement.
-9.	GOVERNING LAW AND JURISDICTION.
-a.	This Agreement and any dispute arising out of or relating to it will be governed by the laws of the Hong Kong Special Administrative Region of the People’s Republic of China, without regard to conflict of law principles, and the UN Convention on Contracts for the International Sale of Goods does not apply to this Agreement.
-b.	Exclusive jurisdiction and venue for any dispute arising out of or relating to this Agreement will be a court of competent jurisdiction in the Hong Kong Special Administrative Region of the People’s Republic of China, and Tencent and Licensee consent to the exclusive jurisdiction of such court with respect to any such dispute.
- 
-EXHIBIT A
-ACCEPTABLE USE POLICY
-
-Tencent reserves the right to update this Acceptable Use Policy from time to time.
-Last modified: November 5, 2024
-
-Tencent endeavors to promote safe and fair use of its tools and features, including Tencent Hunyuan 3D 2.1. You agree not to use Tencent Hunyuan 3D 2.1 or Model Derivatives:
-1.	Outside the Territory;
-2.	In any way that violates any applicable national, federal, state, local, international or any other law or regulation;
-3.	To harm Yourself or others;
-4.	To repurpose or distribute output from Tencent Hunyuan 3D 2.1 or any Model Derivatives to harm Yourself or others; 
-5.	To override or circumvent the safety guardrails and safeguards We have put in place;
-6.	For the purpose of exploiting, harming or attempting to exploit or harm minors in any way;
-7.	To generate or disseminate verifiably false information and/or content with the purpose of harming others or influencing elections;
-8.	To generate or facilitate false online engagement, including fake reviews and other means of fake online engagement;
-9.	To intentionally defame, disparage or otherwise harass others;
-10.	To generate and/or disseminate malware (including ransomware) or any other content to be used for the purpose of harming electronic systems;
-11.	To generate or disseminate personal identifiable information with the purpose of harming others;
-12.	To generate or disseminate information (including images, code, posts, articles), and place the information in any public context (including –through the use of bot generated tweets), without expressly and conspicuously identifying that the information and/or content is machine generated;
-13.	To impersonate another individual without consent, authorization, or legal right;
-14.	To make high-stakes automated decisions in domains that affect an individual’s safety, rights or wellbeing (e.g., law enforcement, migration, medicine/health, management of critical infrastructure, safety components of products, essential services, credit, employment, housing, education, social scoring, or insurance);
-15.	In a manner that violates or disrespects the social ethics and moral standards of other countries or regions;
-16.	To perform, facilitate, threaten, incite, plan, promote or encourage violent extremism or terrorism;
-17.	For any use intended to discriminate against or harm individuals or groups based on protected characteristics or categories, online or offline social behavior or known or predicted personal or personality characteristics;
-18.	To intentionally exploit any of the vulnerabilities of a specific group of persons based on their age, social, physical or mental characteristics, in order to materially distort the behavior of a person pertaining to that group in a manner that causes or is likely to cause that person or another person physical or psychological harm;
-19.	For military purposes;
-20.	To engage in the unauthorized or unlicensed practice of any profession including, but not limited to, financial, legal, medical/health, or other professional practices.

hy3dpaint/README.md

@@ -1,96 +0,0 @@
-# Hunyuan3D-Paint 2.1
-
-Hunyuan3D-Paint 2.1 is a high quality PBR texture generation model for 3D meshes, powered by [RomanTex](https://github.com/oakshy/RomanTex) and [MaterialMVP](https://github.com/ZebinHe/MaterialMVP/).
-
-
-## Quick Inference
-You need to manually download the RealESRGAN weight to the `ckpt` folder using the following command:
-```bash
-wget https://github.com/xinntao/Real-ESRGAN/releases/download/v0.1.0/RealESRGAN_x4plus.pth -P ckpt
-```
-
-Given a 3D mesh `mesh.glb` and a reference image `image.png`, you can run inference using the following code. The result will be saved as `textured_mesh.glb`.
-
-```bash
-python3 demo.py
-```
-**Optional arguments in `demo.py`:**
-
-- `max_num_view` : Maximum number of views, adaptively selected by the model (integer between 6 to 9)
-
-- `resolution` : Resolution for generated PBR textures (512 or 768)
-
-**Memory Recommendation:** For `max_num_view=6` and `resolution=512`, we recommend using a GPU with at least **21GB VRAM**.
-
-## Training
-
-### Data Prepare
-We provide a piece of data in `train_examples` for the overfitting training test. The data structure should be organized as follows:
-
-```
-train_examples/
-├── examples.json
-└── 001/
-    ├── render_tex/                 # Rendered generated PBR images
-    │   ├── 000.png                 # Rendered views (RGB images)
-    │   ├── 000_albedo.png          # Albedo maps for each view
-    │   ├── 000_mr.png              # Metallic-Roughness maps for each view, R and G channels
-    │   ├── 000_normal.png          # Normal maps
-    │   ├── 000_normal.png          # Normal maps
-    │   ├── 000_pos.png             # Position maps
-    │   ├── 000_pos.png             # Position maps
-    │   ├── 001.png                 # Additional views...
-    │   ├── 001_albedo.png
-    │   ├── 001_mr.png
-    │   ├── 001_normal.png
-    │   ├── 001_pos.png
-    │   └── ...                     # More views (002, 003, 004, 005, ...)
-    └── render_cond/                # Rendered reference images (at least two light conditions should be rendered to facilitate consistency loss)
-        ├── 000_light_AL.png        # Light condition 1 (Area Light)
-        ├── 000_light_ENVMAP.png    # Light condition 2 (Environment map)
-        ├── 000_light_PL.png        # Light condition 3 (Point lighting)
-        ├── 001_light_AL.png        
-        ├── 001_light_ENVMAP.png
-        ├── 001_light_PL.png
-        └── ...                      # More lighting conditions (002-005, ...)
-```
-
-Each training example contains:
-- **render_tex/**: Multi-view renderings with PBR material properties
-  - Main RGB images (`XXX.png`)
-  - Albedo maps (`XXX_albedo.png`)
-  - Metallic-Roughness maps (`XXX_mr.png`)
-  - Normal maps (`XXX_normal.png/jpg`)
-  - Position maps (`XXX_pos.png/jpg`)
-  - Camera transforms (`transforms.json`)
-- **render_cond/**: Lighting condition maps for each view
-  - Ambient lighting (`XXX_light_AL.png`)
-  - Environment map lighting (`XXX_light_ENVMAP.png`)
-  - Point lighting (`XXX_light_PL.png`)
-
-### Launch Training
-
-
-```bash
-python3 train.py --base 'cfgs/hunyuan-paint-pbr.yaml' --name overfit --logdir logs/
-```
-
-## BibTeX
-
-If you found Hunyuan3D-Paint 2.1 helpful, please cite our papers:
-
-```bibtex
-@article{feng2025romantex,
-  title={RomanTex: Decoupling 3D-aware Rotary Positional Embedded Multi-Attention Network for Texture Synthesis},
-  author={Feng, Yifei and Yang, Mingxin and Yang, Shuhui and Zhang, Sheng and Yu, Jiaao and Zhao, Zibo and Liu, Yuhong and Jiang, Jie and Guo, Chunchao},
-  journal={arXiv preprint arXiv:2503.19011},
-  year={2025}
-}
-
-@article{he2025materialmvp,
-  title={MaterialMVP: Illumination-Invariant Material Generation via Multi-view PBR Diffusion},
-  author={He, Zebin and Yang, Mingxin and Yang, Shuhui and Tang, Yixuan and Wang, Tao and Zhang, Kaihao and Chen, Guanying and Liu, Yuhong and Jiang, Jie and Guo, Chunchao and Luo, Wenhan},
-  journal={arXiv preprint arXiv:2503.10289},
-  year={2025}
-}
-```

hy3dpaint/cfgs/hunyuan-paint-pbr.yaml

@@ -1,52 +0,0 @@
-model:
-  base_learning_rate: 5.0e-05
-  target: hunyuanpaintpbr.model.HunyuanPaint
-  params:
-    num_view: 6
-    view_size: 512
-    drop_cond_prob: 0.1
-    
-    noise_in_channels: 12
-
-    stable_diffusion_config:
-      pretrained_model_name_or_path: stabilityai/stable-diffusion-2-1
-      custom_pipeline: ./hunyuanpaintpbr
-      
-
-data:
-  target: src.data.objaverse_hunyuan.DataModuleFromConfig
-  params:
-    batch_size: 1
-    num_workers: 4
-    train:
-      - 
-        target: src.data.dataloader.objaverse_loader_forTexturePBR.TextureDataset
-        params:
-          num_view: 6
-          json_path: train_examples/examples.json
-    validation:
-      -
-        target: src.data.dataloader.objaverse_loader_forTexturePBR.TextureDataset
-        params:
-          num_view: 6
-          json_path: train_examples/examples.json
-
-lightning:
-  modelcheckpoint:
-    params:
-      every_n_train_steps: 10000
-      save_top_k: -1
-      save_last: true
-  callbacks: {}
-
-  trainer:
-    benchmark: true
-    max_epochs: -1
-    gradient_clip_val: 1.0
-    val_check_interval: 1000
-    num_sanity_val_steps: 0
-    accumulate_grad_batches: 1
-    check_val_every_n_epoch: null   # if not set this, validation does not run
-
-init_control_from:
-resume_from: 

hy3dpaint/convert_utils.py

@@ -1,140 +0,0 @@
-import trimesh
-import pygltflib
-import numpy as np
-from PIL import Image
-import base64
-import io
-
-
-def combine_metallic_roughness(metallic_path, roughness_path, output_path):
-    """
-    将metallic和roughness贴图合并为一张贴图
-    GLB格式要求metallic在B通道，roughness在G通道
-    """
-    # 加载贴图
-    metallic_img = Image.open(metallic_path).convert("L")  # 转为灰度
-    roughness_img = Image.open(roughness_path).convert("L")  # 转为灰度
-
-    # 确保尺寸一致
-    if metallic_img.size != roughness_img.size:
-        roughness_img = roughness_img.resize(metallic_img.size)
-
-    # 创建RGB图像
-    width, height = metallic_img.size
-    combined = Image.new("RGB", (width, height))
-
-    # 转为numpy数组便于操作
-    metallic_array = np.array(metallic_img)
-    roughness_array = np.array(roughness_img)
-
-    # 创建合并的数组 (R, G, B) = (AO, Roughness, Metallic)
-    combined_array = np.zeros((height, width, 3), dtype=np.uint8)
-    combined_array[:, :, 0] = 255  # R通道：AO (如果没有AO贴图，设为白色)
-    combined_array[:, :, 1] = roughness_array  # G通道：Roughness
-    combined_array[:, :, 2] = metallic_array  # B通道：Metallic
-
-    # 转回PIL图像并保存
-    combined = Image.fromarray(combined_array)
-    combined.save(output_path)
-    return output_path
-
-
-def create_glb_with_pbr_materials(obj_path, textures_dict, output_path):
-    """
-    使用pygltflib创建包含完整PBR材质的GLB文件
-
-    textures_dict = {
-        'albedo': 'path/to/albedo.png',
-        'metallic': 'path/to/metallic.png',
-        'roughness': 'path/to/roughness.png',
-        'normal': 'path/to/normal.png',  # 可选
-        'ao': 'path/to/ao.png'  # 可选
-    }
-    """
-    # 1. 加载OBJ文件
-    mesh = trimesh.load(obj_path)
-
-    # 2. 先导出为临时GLB
-    temp_glb = "temp.glb"
-    mesh.export(temp_glb)
-
-    # 3. 加载GLB文件进行材质编辑
-    gltf = pygltflib.GLTF2().load(temp_glb)
-
-    # 4. 准备纹理数据
-    def image_to_data_uri(image_path):
-        """将图像转换为data URI"""
-        with open(image_path, "rb") as f:
-            image_data = f.read()
-        encoded = base64.b64encode(image_data).decode()
-        return f"data:image/png;base64,{encoded}"
-
-    # 5. 合并metallic和roughness
-    if "metallic" in textures_dict and "roughness" in textures_dict:
-        mr_combined_path = "mr_combined.png"
-        combine_metallic_roughness(textures_dict["metallic"], textures_dict["roughness"], mr_combined_path)
-        textures_dict["metallicRoughness"] = mr_combined_path
-
-    # 6. 添加图像到GLTF
-    images = []
-    textures = []
-
-    texture_mapping = {
-        "albedo": "baseColorTexture",
-        "metallicRoughness": "metallicRoughnessTexture",
-        "normal": "normalTexture",
-        "ao": "occlusionTexture",
-    }
-
-    for tex_type, tex_path in textures_dict.items():
-        if tex_type in texture_mapping and tex_path:
-            # 添加图像
-            image = pygltflib.Image(uri=image_to_data_uri(tex_path))
-            images.append(image)
-
-            # 添加纹理
-            texture = pygltflib.Texture(source=len(images) - 1)
-            textures.append(texture)
-
-    # 7. 创建PBR材质
-    pbr_metallic_roughness = pygltflib.PbrMetallicRoughness(
-        baseColorFactor=[1.0, 1.0, 1.0, 1.0], metallicFactor=1.0, roughnessFactor=1.0
-    )
-
-    # 设置纹理索引
-    texture_index = 0
-    if "albedo" in textures_dict:
-        pbr_metallic_roughness.baseColorTexture = pygltflib.TextureInfo(index=texture_index)
-        texture_index += 1
-
-    if "metallicRoughness" in textures_dict:
-        pbr_metallic_roughness.metallicRoughnessTexture = pygltflib.TextureInfo(index=texture_index)
-        texture_index += 1
-
-    # 创建材质
-    material = pygltflib.Material(name="PBR_Material", pbrMetallicRoughness=pbr_metallic_roughness)
-
-    # 添加法线贴图
-    if "normal" in textures_dict:
-        material.normalTexture = pygltflib.NormalTextureInfo(index=texture_index)
-        texture_index += 1
-
-    # 添加AO贴图
-    if "ao" in textures_dict:
-        material.occlusionTexture = pygltflib.OcclusionTextureInfo(index=texture_index)
-
-    # 8. 更新GLTF
-    gltf.images = images
-    gltf.textures = textures
-    gltf.materials = [material]
-
-    # 确保mesh使用材质
-    if gltf.meshes:
-        for primitive in gltf.meshes[0].primitives:
-            primitive.material = 0
-
-    # 9. 保存最终GLB
-    gltf.save(output_path)
-    print(f"PBR GLB文件已保存: {output_path}")
-
-

hy3dpaint/demo.py

@@ -1,35 +0,0 @@
-# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT
-# except for the third-party components listed below.
-# Hunyuan 3D does not impose any additional limitations beyond what is outlined
-# in the repsective licenses of these third-party components.
-# Users must comply with all terms and conditions of original licenses of these third-party
-# components and must ensure that the usage of the third party components adheres to
-# all relevant laws and regulations.
-
-# For avoidance of doubts, Hunyuan 3D means the large language models and
-# their software and algorithms, including trained model weights, parameters (including
-# optimizer states), machine-learning model code, inference-enabling code, training-enabling code,
-# fine-tuning enabling code and other elements of the foregoing made publicly available
-# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT.
-
-from textureGenPipeline import Hunyuan3DPaintPipeline, Hunyuan3DPaintConfig
-
-try:
-    from utils.torchvision_fix import apply_fix
-
-    apply_fix()
-except ImportError:
-    print("Warning: torchvision_fix module not found, proceeding without compatibility fix")
-except Exception as e:
-    print(f"Warning: Failed to apply torchvision fix: {e}")
-
-
-if __name__ == "__main__":
-
-    max_num_view = 6  # can be 6 to 9
-    resolution = 512  # can be 768 or 512
-
-    conf = Hunyuan3DPaintConfig(max_num_view, resolution)
-    paint_pipeline = Hunyuan3DPaintPipeline(conf)
-    output_mesh_path = paint_pipeline(mesh_path="./assets/case_1/mesh.glb", image_path="./assets/case_1/image.png")
-    print(f"Output mesh path: {output_mesh_path}")

hy3dpaint/hunyuanpaintpbr/__init__.py

@@ -1,39 +0,0 @@
-# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT
-# except for the third-party components listed below.
-# Hunyuan 3D does not impose any additional limitations beyond what is outlined
-# in the repsective licenses of these third-party components.
-# Users must comply with all terms and conditions of original licenses of these third-party
-# components and must ensure that the usage of the third party components adheres to
-# all relevant laws and regulations.
-
-# For avoidance of doubts, Hunyuan 3D means the large language models and
-# their software and algorithms, including trained model weights, parameters (including
-# optimizer states), machine-learning model code, inference-enabling code, training-enabling code,
-# fine-tuning enabling code and other elements of the foregoing made publicly available
-# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT.
-
-from .pipeline import HunyuanPaintPipeline
-from .unet.model import HunyuanPaint
-from .unet.modules import (
-    Dino_v2,
-    Basic2p5DTransformerBlock,
-    ImageProjModel,
-    UNet2p5DConditionModel,
-)
-from .unet.attn_processor import (
-    PoseRoPEAttnProcessor2_0,
-    SelfAttnProcessor2_0,
-    RefAttnProcessor2_0,
-)
-
-__all__ = [
-    'HunyuanPaintPipeline',
-    'HunyuanPaint',
-    'Dino_v2',
-    'Basic2p5DTransformerBlock',
-    'ImageProjModel',
-    'UNet2p5DConditionModel',
-    'PoseRoPEAttnProcessor2_0',
-    'SelfAttnProcessor2_0',
-    'RefAttnProcessor2_0',
-]

hy3dpaint/hunyuanpaintpbr/pipeline.py

@@ -1,736 +0,0 @@
-# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT
-# except for the third-party components listed below.
-# Hunyuan 3D does not impose any additional limitations beyond what is outlined
-# in the repsective licenses of these third-party components.
-# Users must comply with all terms and conditions of original licenses of these third-party
-# components and must ensure that the usage of the third party components adheres to
-# all relevant laws and regulations.
-
-# For avoidance of doubts, Hunyuan 3D means the large language models and
-# their software and algorithms, including trained model weights, parameters (including
-# optimizer states), machine-learning model code, inference-enabling code, training-enabling code,
-# fine-tuning enabling code and other elements of the foregoing made publicly available
-# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT.
-
-from typing import Any, Dict, Optional
-from diffusers.models import AutoencoderKL, UNet2DConditionModel
-from diffusers.schedulers import KarrasDiffusionSchedulers
-
-import numpy
-import torch
-import torch.utils.checkpoint
-import torch.distributed
-import numpy as np
-import transformers
-from PIL import Image
-from einops import rearrange
-from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
-from typing import Any, Callable, Dict, List, Optional, Union, Tuple
-
-import diffusers
-from diffusers import (
-    AutoencoderKL,
-    DiffusionPipeline,
-    UNet2DConditionModel,
-)
-from diffusers.image_processor import VaeImageProcessor
-
-from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion import (
-    StableDiffusionPipeline,
-    retrieve_timesteps,
-    rescale_noise_cfg,
-)
-
-from diffusers.utils import deprecate
-from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback
-from diffusers.image_processor import PipelineImageInput
-from diffusers.pipelines.stable_diffusion.pipeline_output import StableDiffusionPipelineOutput
-from .unet.modules import UNet2p5DConditionModel
-from .unet.attn_processor import SelfAttnProcessor2_0, RefAttnProcessor2_0, PoseRoPEAttnProcessor2_0
-
-__all__ = [
-    "HunyuanPaintPipeline",
-    "UNet2p5DConditionModel",
-    "SelfAttnProcessor2_0",
-    "RefAttnProcessor2_0",
-    "PoseRoPEAttnProcessor2_0",
-]
-
-
-def to_rgb_image(maybe_rgba: Image.Image):
-    if maybe_rgba.mode == "RGB":
-        return maybe_rgba
-    elif maybe_rgba.mode == "RGBA":
-        rgba = maybe_rgba
-        img = numpy.random.randint(127, 128, size=[rgba.size[1], rgba.size[0], 3], dtype=numpy.uint8)
-        img = Image.fromarray(img, "RGB")
-        img.paste(rgba, mask=rgba.getchannel("A"))
-        return img
-    else:
-        raise ValueError("Unsupported image type.", maybe_rgba.mode)
-
-
-class HunyuanPaintPipeline(StableDiffusionPipeline):
-
-    """Custom pipeline for multiview PBR texture generation.
-    
-    Extends Stable Diffusion with:
-    - Material-specific conditioning
-    - Multiview processing
-    - Position-aware attention
-    - 2.5D UNet integration
-    """
-
-    def __init__(
-        self,
-        vae: AutoencoderKL,
-        text_encoder: CLIPTextModel,
-        tokenizer: CLIPTokenizer,
-        unet: UNet2DConditionModel,
-        scheduler: KarrasDiffusionSchedulers,
-        feature_extractor: CLIPImageProcessor,
-        safety_checker=None,
-        use_torch_compile=False,
-    ):
-        DiffusionPipeline.__init__(self)
-
-        safety_checker = None
-        self.register_modules(
-            vae=torch.compile(vae) if use_torch_compile else vae,
-            text_encoder=text_encoder,
-            tokenizer=tokenizer,
-            unet=unet,
-            scheduler=scheduler,
-            safety_checker=safety_checker,
-            feature_extractor=torch.compile(feature_extractor) if use_torch_compile else feature_extractor,
-        )
-
-        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
-        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
-
-        if isinstance(self.unet, UNet2DConditionModel):
-            self.unet = UNet2p5DConditionModel(self.unet, None, self.scheduler)
-
-    def eval(self):
-        self.unet.eval()
-        self.vae.eval()
-
-    def set_pbr_settings(self, pbr_settings: List[str]):
-        self.pbr_settings = pbr_settings
-
-    def set_learned_parameters(self):
-
-        """Configures parameter freezing strategy.
-        
-        Freezes:
-        - Standard attention layers
-        - Dual-stream reference UNet
-        
-        Unfreezes:
-        - Material-specific parameters
-        - DINO integration components
-        """
-
-        freezed_names = ["attn1", "unet_dual"]
-        added_learned_names = ["albedo", "mr", "dino"]
-
-        for name, params in self.unet.named_parameters():
-            if any(freeze_name in name for freeze_name in freezed_names) and all(
-                learned_name not in name for learned_name in added_learned_names
-            ):
-                params.requires_grad = False
-            else:
-                params.requires_grad = True
-
-    def prepare(self):
-        if isinstance(self.unet, UNet2DConditionModel):
-            self.unet = UNet2p5DConditionModel(self.unet, None, self.scheduler).eval()
-
-    @torch.no_grad()
-    def encode_images(self, images):
-
-        """Encodes multiview image batches into latent space.
-        
-        Args:
-            images: Input images [B, N_views, C, H, W]
-            
-        Returns:
-            torch.Tensor: Latent representations [B, N_views, C, H_latent, W_latent]
-        """
-
-        B = images.shape[0]
-        images = rearrange(images, "b n c h w -> (b n) c h w")
-
-        dtype = next(self.vae.parameters()).dtype
-        images = (images - 0.5) * 2.0
-        posterior = self.vae.encode(images.to(dtype)).latent_dist
-        latents = posterior.sample() * self.vae.config.scaling_factor
-
-        latents = rearrange(latents, "(b n) c h w -> b n c h w", b=B)
-        return latents
-
-    @torch.no_grad()
-    def __call__(
-        self,
-        images=None,
-        prompt=None,
-        negative_prompt="watermark, ugly, deformed, noisy, blurry, low contrast",
-        *args,
-        num_images_per_prompt: Optional[int] = 1,
-        guidance_scale=3.0,
-        output_type: Optional[str] = "pil",
-        width=512,
-        height=512,
-        num_inference_steps=15,
-        return_dict=True,
-        sync_condition=None,
-        **cached_condition,
-    ):
-
-        """Main generation method for multiview PBR textures.
-        
-        Steps:
-        1. Input validation and preparation
-        2. Reference image encoding
-        3. Condition processing (normal/position maps)
-        4. Prompt embedding setup
-        5. Classifier-free guidance preparation
-        6. Diffusion sampling loop
-        
-        Args:
-            images: List of reference PIL images
-            prompt: Text prompt (overridden by learned embeddings)
-            cached_condition: Dictionary containing:
-                - images_normal: Normal maps (PIL or tensor)
-                - images_position: Position maps (PIL or tensor)
-            
-        Returns:
-            List[PIL.Image]: Generated multiview PBR textures
-        """
-
-        self.prepare()
-        if images is None:
-            raise ValueError("Inputting embeddings not supported for this pipeline. Please pass an image.")
-        assert not isinstance(images, torch.Tensor)
-
-        if not isinstance(images, List):
-            images = [images]
-
-        images = [to_rgb_image(image) for image in images]
-        images_vae = [torch.tensor(np.array(image) / 255.0) for image in images]
-        images_vae = [image_vae.unsqueeze(0).permute(0, 3, 1, 2).unsqueeze(0) for image_vae in images_vae]
-        images_vae = torch.cat(images_vae, dim=1)
-        images_vae = images_vae.to(device=self.vae.device, dtype=self.unet.dtype)
-
-        batch_size = images_vae.shape[0]
-        N_ref = images_vae.shape[1]
-
-        assert batch_size == 1
-        assert num_images_per_prompt == 1
-
-        if self.unet.use_ra:
-            ref_latents = self.encode_images(images_vae)
-            cached_condition["ref_latents"] = ref_latents
-
-        def convert_pil_list_to_tensor(images):
-            bg_c = [1.0, 1.0, 1.0]
-            images_tensor = []
-            for batch_imgs in images:
-                view_imgs = []
-                for pil_img in batch_imgs:
-                    img = numpy.asarray(pil_img, dtype=numpy.float32) / 255.0
-                    if img.shape[2] > 3:
-                        alpha = img[:, :, 3:]
-                        img = img[:, :, :3] * alpha + bg_c * (1 - alpha)
-                    img = torch.from_numpy(img).permute(2, 0, 1).unsqueeze(0).contiguous().half().to("cuda")
-                    view_imgs.append(img)
-                view_imgs = torch.cat(view_imgs, dim=0)
-                images_tensor.append(view_imgs.unsqueeze(0))
-
-            images_tensor = torch.cat(images_tensor, dim=0)
-            return images_tensor
-
-        if "images_normal" in cached_condition:
-            if isinstance(cached_condition["images_normal"], List):
-                cached_condition["images_normal"] = convert_pil_list_to_tensor(cached_condition["images_normal"])
-
-            cached_condition["embeds_normal"] = self.encode_images(cached_condition["images_normal"])
-
-        if "images_position" in cached_condition:
-
-            if isinstance(cached_condition["images_position"], List):
-                cached_condition["images_position"] = convert_pil_list_to_tensor(cached_condition["images_position"])
-
-            cached_condition["position_maps"] = cached_condition["images_position"]
-            cached_condition["embeds_position"] = self.encode_images(cached_condition["images_position"])
-
-        if self.unet.use_learned_text_clip:
-
-            all_shading_tokens = []
-            for token in self.unet.pbr_setting:
-                all_shading_tokens.append(
-                    getattr(self.unet, f"learned_text_clip_{token}").unsqueeze(dim=0).repeat(batch_size, 1, 1)
-                )
-            prompt_embeds = torch.stack(all_shading_tokens, dim=1)
-            negative_prompt_embeds = torch.stack(all_shading_tokens, dim=1)
-            # negative_prompt_embeds = torch.zeros_like(prompt_embeds)
-
-        else:
-            if prompt is None:
-                prompt = "high quality"
-            if isinstance(prompt, str):
-                prompt = [prompt for _ in range(batch_size)]
-            device = self._execution_device
-            prompt_embeds, _ = self.encode_prompt(
-                prompt, device=device, num_images_per_prompt=num_images_per_prompt, do_classifier_free_guidance=False
-            )
-
-            if isinstance(negative_prompt, str):
-                negative_prompt = [negative_prompt for _ in range(batch_size)]
-            if negative_prompt is not None:
-                negative_prompt_embeds, _ = self.encode_prompt(
-                    negative_prompt,
-                    device=device,
-                    num_images_per_prompt=num_images_per_prompt,
-                    do_classifier_free_guidance=False,
-                )
-            else:
-                negative_prompt_embeds = torch.zeros_like(prompt_embeds)
-
-        if guidance_scale > 1:
-            if self.unet.use_ra:
-                cached_condition["ref_latents"] = cached_condition["ref_latents"].repeat(
-                    3, *([1] * (cached_condition["ref_latents"].dim() - 1))
-                )
-                cached_condition["ref_scale"] = torch.as_tensor([0.0, 1.0, 1.0]).to(cached_condition["ref_latents"])
-
-            if self.unet.use_dino:
-                zero_states = torch.zeros_like(cached_condition["dino_hidden_states"])
-                cached_condition["dino_hidden_states"] = torch.cat(
-                    [zero_states, zero_states, cached_condition["dino_hidden_states"]]
-                )
-
-                del zero_states
-            if "embeds_normal" in cached_condition:
-                cached_condition["embeds_normal"] = cached_condition["embeds_normal"].repeat(
-                    3, *([1] * (cached_condition["embeds_normal"].dim() - 1))
-                )
-
-            if "embeds_position" in cached_condition:
-                cached_condition["embeds_position"] = cached_condition["embeds_position"].repeat(
-                    3, *([1] * (cached_condition["embeds_position"].dim() - 1))
-                )
-
-            if "position_maps" in cached_condition:
-                cached_condition["position_maps"] = cached_condition["position_maps"].repeat(
-                    3, *([1] * (cached_condition["position_maps"].dim() - 1))
-                )
-
-        images = self.denoise(
-            None,
-            *args,
-            cross_attention_kwargs=None,
-            guidance_scale=guidance_scale,
-            num_images_per_prompt=num_images_per_prompt,
-            prompt_embeds=prompt_embeds,
-            negative_prompt_embeds=negative_prompt_embeds,
-            num_inference_steps=num_inference_steps,
-            output_type=output_type,
-            width=width,
-            height=height,
-            return_dict=return_dict,
-            **cached_condition,
-        )
-
-        return images
-
-    def denoise(
-        self,
-        prompt: Union[str, List[str]] = None,
-        height: Optional[int] = None,
-        width: Optional[int] = None,
-        num_inference_steps: int = 50,
-        timesteps: List[int] = None,
-        sigmas: List[float] = None,
-        guidance_scale: float = 7.5,
-        negative_prompt: Optional[Union[str, List[str]]] = None,
-        num_images_per_prompt: Optional[int] = 1,
-        eta: float = 0.0,
-        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
-        latents: Optional[torch.Tensor] = None,
-        prompt_embeds: Optional[torch.Tensor] = None,
-        negative_prompt_embeds: Optional[torch.Tensor] = None,
-        ip_adapter_image: Optional[PipelineImageInput] = None,
-        ip_adapter_image_embeds: Optional[List[torch.Tensor]] = None,
-        output_type: Optional[str] = "pil",
-        return_dict: bool = True,
-        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
-        guidance_rescale: float = 0.0,
-        clip_skip: Optional[int] = None,
-        callback_on_step_end: Optional[
-            Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks]
-        ] = None,
-        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
-        **kwargs,
-    ):
-        r"""
-        The call function to the pipeline for generation.
-
-        Args:
-            prompt (`str` or `List[str]`, *optional*):
-                The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
-            height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
-                The height in pixels of the generated image.
-            width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
-                The width in pixels of the generated image.
-            num_inference_steps (`int`, *optional*, defaults to 50):
-                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
-                expense of slower inference.
-            timesteps (`List[int]`, *optional*):
-                Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
-                in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
-                passed will be used. Must be in descending order.
-            sigmas (`List[float]`, *optional*):
-                Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in
-                their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed
-                will be used.
-            guidance_scale (`float`, *optional*, defaults to 7.5):
-                A higher guidance scale value encourages the model to generate images closely linked to the text
-                `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
-            negative_prompt (`str` or `List[str]`, *optional*):
-                The prompt or prompts to guide what to not include in image generation. If not defined, you need to
-                pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
-            num_images_per_prompt (`int`, *optional*, defaults to 1):
-                The number of images to generate per prompt.
-            eta (`float`, *optional*, defaults to 0.0):
-                Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
-                to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
-            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
-                A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
-                generation deterministic.
-            latents (`torch.Tensor`, *optional*):
-                Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image
-                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
-                tensor is generated by sampling using the supplied random `generator`.
-            prompt_embeds (`torch.Tensor`, *optional*):
-                Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
-                provided, text embeddings are generated from the `prompt` input argument.
-            negative_prompt_embeds (`torch.Tensor`, *optional*):
-                Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
-                not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
-            ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
-            ip_adapter_image_embeds (`List[torch.Tensor]`, *optional*):
-                Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of
-                IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should
-                contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not
-                provided, embeddings are computed from the `ip_adapter_image` input argument.
-            output_type (`str`, *optional*, defaults to `"pil"`):
-                The output format of the generated image. Choose between `PIL.Image` or `np.array`.
-            return_dict (`bool`, *optional*, defaults to `True`):
-                Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
-                plain tuple.
-            cross_attention_kwargs (`dict`, *optional*):
-                A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
-                [`self.processor`]
-                (https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
-            guidance_rescale (`float`, *optional*, defaults to 0.0):
-                Guidance rescale factor from [Common Diffusion Noise Schedules and Sample Steps are
-                Flawed](https://arxiv.org/pdf/2305.08891.pdf). Guidance rescale factor should fix overexposure when
-                using zero terminal SNR.
-            clip_skip (`int`, *optional*):
-                Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
-                the output of the pre-final layer will be used for computing the prompt embeddings.
-            callback_on_step_end (`Callable`, `PipelineCallback`, `MultiPipelineCallbacks`, *optional*):
-                A function or a subclass of `PipelineCallback` or `MultiPipelineCallbacks` that is called at the end of
-                each denoising step during the inference. with the following arguments: `callback_on_step_end(self:
-                DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict)`. `callback_kwargs` will include a
-                list of all tensors as specified by `callback_on_step_end_tensor_inputs`.
-            callback_on_step_end_tensor_inputs (`List`, *optional*):
-                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
-                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
-                `._callback_tensor_inputs` attribute of your pipeline class.
-
-        Examples:
-
-        Returns:
-            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
-                If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
-                otherwise a `tuple` is returned where the first element is a list with the generated images and the
-                second element is a list of `bool`s indicating whether the corresponding generated image contains
-                "not-safe-for-work" (nsfw) content.
-                
-        Core denoising procedure for multiview PBR texture generation.
-        
-        Handles the complete diffusion process including:
-        - Input validation and preparation
-        - Timestep scheduling
-        - Latent noise initialization
-        - Iterative denoising with specialized guidance
-        - Output decoding and post-processing
-        
-        Key innovations:
-        1. Triple-batch classifier-free guidance:
-           - Negative (unconditional)
-           - Reference-conditioned
-           - Full-conditioned
-        2. View-dependent guidance scaling:
-           - Adjusts influence based on camera azimuth
-        3. PBR-aware latent organization:
-           - Maintains material/view separation throughout
-        4. Optimized VRAM management:
-           - Selective tensor reshaping
-        
-        Processing Stages:
-        1. Setup & Validation: Configures pipeline components and validates inputs
-        2. Prompt Encoding: Processes text/material conditioning
-        3. Latent Initialization: Prepares noise for denoising process
-        4. Iterative Denoising: 
-            a) Scales and organizes latent variables
-            b) Predicts noise at current timestep
-            c) Applies view-dependent guidance
-            d) Computes previous latent state
-        5. Output Decoding: Converts latents to final images
-        6. Cleanup: Releases resources and formats output
-        
-        """
-
-        callback = kwargs.pop("callback", None)
-        callback_steps = kwargs.pop("callback_steps", None)
-
-        # open cache
-        kwargs["cache"] = {}
-
-        if callback is not None:
-            deprecate(
-                "callback",
-                "1.0.0",
-                "Passing `callback` as an input argument to `__call__` is deprecated,"
-                "consider using `callback_on_step_end`",
-            )
-        if callback_steps is not None:
-            deprecate(
-                "callback_steps",
-                "1.0.0",
-                "Passing `callback` as an input argument to `__call__` is deprecated,"
-                "consider using `callback_on_step_end`",
-            )
-
-        if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
-            callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
-
-        # 0. Default height and width to unet
-        height = height or self.unet.config.sample_size * self.vae_scale_factor
-        width = width or self.unet.config.sample_size * self.vae_scale_factor
-        # to deal with lora scaling and other possible forward hooks
-
-        # 1. Check inputs. Raise error if not correct
-        self.check_inputs(
-            prompt,
-            height,
-            width,
-            callback_steps,
-            negative_prompt,
-            prompt_embeds,
-            negative_prompt_embeds,
-            ip_adapter_image,
-            ip_adapter_image_embeds,
-            callback_on_step_end_tensor_inputs,
-        )
-
-        self._guidance_scale = guidance_scale
-        self._guidance_rescale = guidance_rescale
-        self._clip_skip = clip_skip
-        self._cross_attention_kwargs = cross_attention_kwargs
-        self._interrupt = False
-
-        # 2. Define call parameters
-        if prompt is not None and isinstance(prompt, str):
-            batch_size = 1
-        elif prompt is not None and isinstance(prompt, list):
-            batch_size = len(prompt)
-        else:
-            batch_size = prompt_embeds.shape[0]
-
-        device = self._execution_device
-
-        # 3. Encode input prompt
-        lora_scale = self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None
-
-        """
-        prompt_embeds, negative_prompt_embeds = self.encode_prompt(
-            prompt,
-            device,
-            num_images_per_prompt,
-            self.do_classifier_free_guidance,
-            negative_prompt,
-            prompt_embeds=prompt_embeds,
-            negative_prompt_embeds=negative_prompt_embeds,
-            lora_scale=lora_scale,
-            clip_skip=self.clip_skip,
-        )'
-        """
-
-        # For classifier free guidance, we need to do two forward passes.
-        # Here we concatenate the unconditional and text embeddings into a single batch
-        # to avoid doing two forward passes
-        if self.do_classifier_free_guidance:
-            # prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
-            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds, prompt_embeds])
-
-        if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
-            image_embeds = self.prepare_ip_adapter_image_embeds(
-                ip_adapter_image,
-                ip_adapter_image_embeds,
-                device,
-                batch_size * num_images_per_prompt,
-                self.do_classifier_free_guidance,
-            )
-
-        # 4. Prepare timesteps
-        timesteps, num_inference_steps = retrieve_timesteps(
-            self.scheduler, num_inference_steps, device, timesteps, sigmas
-        )
-        assert num_images_per_prompt == 1
-        # 5. Prepare latent variables
-        n_pbr = len(self.unet.pbr_setting)
-        num_channels_latents = self.unet.config.in_channels
-        latents = self.prepare_latents(
-            batch_size * kwargs["num_in_batch"] * n_pbr,  # num_images_per_prompt,
-            num_channels_latents,
-            height,
-            width,
-            prompt_embeds.dtype,
-            device,
-            generator,
-            latents,
-        )
-
-        # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
-        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
-
-        # 6.1 Add image embeds for IP-Adapter
-        added_cond_kwargs = (
-            {"image_embeds": image_embeds}
-            if (ip_adapter_image is not None or ip_adapter_image_embeds is not None)
-            else None
-        )
-
-        # 6.2 Optionally get Guidance Scale Embedding
-        timestep_cond = None
-        if self.unet.config.time_cond_proj_dim is not None:
-            guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt)
-            timestep_cond = self.get_guidance_scale_embedding(
-                guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim
-            ).to(device=device, dtype=latents.dtype)
-
-        # 7. Denoising loop
-        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
-        self._num_timesteps = len(timesteps)
-        with self.progress_bar(total=num_inference_steps) as progress_bar:
-            for i, t in enumerate(timesteps):
-                if self.interrupt:
-                    continue
-
-                # expand the latents if we are doing classifier free guidance
-                latents = rearrange(
-                    latents, "(b n_pbr n) c h w -> b n_pbr n c h w", n=kwargs["num_in_batch"], n_pbr=n_pbr
-                )
-                # latent_model_input = torch.cat([latents] * 3) if self.do_classifier_free_guidance else latents
-                latent_model_input = latents.repeat(3, 1, 1, 1, 1, 1) if self.do_classifier_free_guidance else latents
-                latent_model_input = rearrange(latent_model_input, "b n_pbr n c h w -> (b n_pbr n) c h w")
-                latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
-                latent_model_input = rearrange(
-                    latent_model_input, "(b n_pbr n) c h w ->b n_pbr n c h w", n=kwargs["num_in_batch"], n_pbr=n_pbr
-                )
-
-                # predict the noise residual
-
-                noise_pred = self.unet(
-                    latent_model_input,
-                    t,
-                    encoder_hidden_states=prompt_embeds,
-                    timestep_cond=timestep_cond,
-                    cross_attention_kwargs=self.cross_attention_kwargs,
-                    added_cond_kwargs=added_cond_kwargs,
-                    return_dict=False,
-                    **kwargs,
-                )[0]
-                latents = rearrange(latents, "b n_pbr n c h w -> (b n_pbr n) c h w")
-                # perform guidance
-                if self.do_classifier_free_guidance:
-                    # noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
-                    # noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
-                    noise_pred_uncond, noise_pred_ref, noise_pred_full = noise_pred.chunk(3)
-
-                    if "camera_azims" in kwargs.keys():
-                        camera_azims = kwargs["camera_azims"]
-                    else:
-                        camera_azims = [0] * kwargs["num_in_batch"]
-
-                    def cam_mapping(azim):
-                        if azim < 90 and azim >= 0:
-                            return float(azim) / 90.0 + 1
-                        elif azim >= 90 and azim < 330:
-                            return 2.0
-                        else:
-                            return -float(azim) / 90.0 + 5.0
-
-                    view_scale_tensor = (
-                        torch.from_numpy(np.asarray([cam_mapping(azim) for azim in camera_azims]))
-                        .unsqueeze(0)
-                        .repeat(n_pbr, 1)
-                        .view(-1)
-                        .to(noise_pred_uncond)[:, None, None, None]
-                    )
-                    noise_pred = noise_pred_uncond + self.guidance_scale * view_scale_tensor * (
-                        noise_pred_ref - noise_pred_uncond
-                    )
-                    noise_pred += self.guidance_scale * view_scale_tensor * (noise_pred_full - noise_pred_ref)
-
-                if self.do_classifier_free_guidance and self.guidance_rescale > 0.0:
-                    # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
-                    noise_pred = rescale_noise_cfg(noise_pred, noise_pred_ref, guidance_rescale=self.guidance_rescale)
-
-                # compute the previous noisy sample x_t -> x_t-1
-                latents = self.scheduler.step(
-                    noise_pred, t, latents[:, :num_channels_latents, :, :], **extra_step_kwargs, return_dict=False
-                )[0]
-
-                if callback_on_step_end is not None:
-                    callback_kwargs = {}
-                    for k in callback_on_step_end_tensor_inputs:
-                        callback_kwargs[k] = locals()[k]
-                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
-
-                    latents = callback_outputs.pop("latents", latents)
-                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
-                    negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
-
-                # call the callback, if provided
-                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
-                    progress_bar.update()
-                    if callback is not None and i % callback_steps == 0:
-                        step_idx = i // getattr(self.scheduler, "order", 1)
-                        callback(step_idx, t, latents)
-
-        if not output_type == "latent":
-            image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False, generator=generator)[0]
-            image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
-        else:
-            image = latents
-            has_nsfw_concept = None
-
-        if has_nsfw_concept is None:
-            do_denormalize = [True] * image.shape[0]
-        else:
-            do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
-
-        image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
-
-        # Offload all models
-        self.maybe_free_model_hooks()
-
-        if not return_dict:
-            return (image, has_nsfw_concept)
-
-        return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)

hy3dpaint/hunyuanpaintpbr/unet/attn_processor.py

@@ -1,839 +0,0 @@
-# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT
-# except for the third-party components listed below.
-# Hunyuan 3D does not impose any additional limitations beyond what is outlined
-# in the repsective licenses of these third-party components.
-# Users must comply with all terms and conditions of original licenses of these third-party
-# components and must ensure that the usage of the third party components adheres to
-# all relevant laws and regulations.
-
-# For avoidance of doubts, Hunyuan 3D means the large language models and
-# their software and algorithms, including trained model weights, parameters (including
-# optimizer states), machine-learning model code, inference-enabling code, training-enabling code,
-# fine-tuning enabling code and other elements of the foregoing made publicly available
-# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT.
-
-import torch
-import torch.nn as nn
-import torch.nn.functional as F
-from typing import Optional, Dict, Tuple, Union, Literal, List, Callable
-from einops import rearrange
-from diffusers.utils import deprecate
-from diffusers.models.attention_processor import Attention, AttnProcessor
-
-
-class AttnUtils:
-    """
-    Shared utility functions for attention processing.
-
-    This class provides common operations used across different attention processors
-    to eliminate code duplication and improve maintainability.
-    """
-
-    @staticmethod
-    def check_pytorch_compatibility():
-        """
-        Check PyTorch compatibility for scaled_dot_product_attention.
-
-        Raises:
-            ImportError: If PyTorch version doesn't support scaled_dot_product_attention
-        """
-        if not hasattr(F, "scaled_dot_product_attention"):
-            raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
-
-    @staticmethod
-    def handle_deprecation_warning(args, kwargs):
-        """
-        Handle deprecation warning for the 'scale' argument.
-
-        Args:
-            args: Positional arguments passed to attention processor
-            kwargs: Keyword arguments passed to attention processor
-        """
-        if len(args) > 0 or kwargs.get("scale", None) is not None:
-            deprecation_message = (
-                "The `scale` argument is deprecated and will be ignored."
-                "Please remove it, as passing it will raise an error in the future."
-                "`scale` should directly be passed while calling the underlying pipeline component"
-                "i.e., via `cross_attention_kwargs`."
-            )
-            deprecate("scale", "1.0.0", deprecation_message)
-
-    @staticmethod
-    def prepare_hidden_states(
-        hidden_states, attn, temb, spatial_norm_attr="spatial_norm", group_norm_attr="group_norm"
-    ):
-        """
-        Common preprocessing of hidden states for attention computation.
-
-        Args:
-            hidden_states: Input hidden states tensor
-            attn: Attention module instance
-            temb: Optional temporal embedding tensor
-            spatial_norm_attr: Attribute name for spatial normalization
-            group_norm_attr: Attribute name for group normalization
-
-        Returns:
-            Tuple of (processed_hidden_states, residual, input_ndim, shape_info)
-        """
-        residual = hidden_states
-
-        spatial_norm = getattr(attn, spatial_norm_attr, None)
-        if spatial_norm is not None:
-            hidden_states = spatial_norm(hidden_states, temb)
-
-        input_ndim = hidden_states.ndim
-
-        if input_ndim == 4:
-            batch_size, channel, height, width = hidden_states.shape
-            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
-        else:
-            batch_size, channel, height, width = None, None, None, None
-
-        group_norm = getattr(attn, group_norm_attr, None)
-        if group_norm is not None:
-            hidden_states = group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
-
-        return hidden_states, residual, input_ndim, (batch_size, channel, height, width)
-
-    @staticmethod
-    def prepare_attention_mask(attention_mask, attn, sequence_length, batch_size):
-        """
-        Prepare attention mask for scaled_dot_product_attention.
-
-        Args:
-            attention_mask: Input attention mask tensor or None
-            attn: Attention module instance
-            sequence_length: Length of the sequence
-            batch_size: Batch size
-
-        Returns:
-            Prepared attention mask tensor reshaped for multi-head attention
-        """
-        if attention_mask is not None:
-            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
-            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
-        return attention_mask
-
-    @staticmethod
-    def reshape_qkv_for_attention(tensor, batch_size, attn_heads, head_dim):
-        """
-        Reshape Q/K/V tensors for multi-head attention computation.
-
-        Args:
-            tensor: Input tensor to reshape
-            batch_size: Batch size
-            attn_heads: Number of attention heads
-            head_dim: Dimension per attention head
-
-        Returns:
-            Reshaped tensor with shape [batch_size, attn_heads, seq_len, head_dim]
-        """
-        return tensor.view(batch_size, -1, attn_heads, head_dim).transpose(1, 2)
-
-    @staticmethod
-    def apply_norms(query, key, norm_q, norm_k):
-        """
-        Apply Q/K normalization layers if available.
-
-        Args:
-            query: Query tensor
-            key: Key tensor
-            norm_q: Query normalization layer (optional)
-            norm_k: Key normalization layer (optional)
-
-        Returns:
-            Tuple of (normalized_query, normalized_key)
-        """
-        if norm_q is not None:
-            query = norm_q(query)
-        if norm_k is not None:
-            key = norm_k(key)
-        return query, key
-
-    @staticmethod
-    def finalize_output(hidden_states, input_ndim, shape_info, attn, residual, to_out):
-        """
-        Common output processing including projection, dropout, reshaping, and residual connection.
-
-        Args:
-            hidden_states: Processed hidden states from attention
-            input_ndim: Original input tensor dimensions
-            shape_info: Tuple containing original shape information
-            attn: Attention module instance
-            residual: Residual connection tensor
-            to_out: Output projection layers [linear, dropout]
-
-        Returns:
-            Final output tensor after all processing steps
-        """
-        batch_size, channel, height, width = shape_info
-
-        # Apply output projection and dropout
-        hidden_states = to_out[0](hidden_states)
-        hidden_states = to_out[1](hidden_states)
-
-        # Reshape back if needed
-        if input_ndim == 4:
-            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
-
-        # Apply residual connection
-        if attn.residual_connection:
-            hidden_states = hidden_states + residual
-
-        # Apply rescaling
-        hidden_states = hidden_states / attn.rescale_output_factor
-        return hidden_states
-
-
-# Base class for attention processors (eliminating initialization duplication)
-class BaseAttnProcessor(nn.Module):
-    """
-    Base class for attention processors with common initialization.
-
-    This base class provides shared parameter initialization and module registration
-    functionality to reduce code duplication across different attention processor types.
-    """
-
-    def __init__(
-        self,
-        query_dim: int,
-        pbr_setting: List[str] = ["albedo", "mr"],
-        cross_attention_dim: Optional[int] = None,
-        heads: int = 8,
-        kv_heads: Optional[int] = None,
-        dim_head: int = 64,
-        dropout: float = 0.0,
-        bias: bool = False,
-        upcast_attention: bool = False,
-        upcast_softmax: bool = False,
-        cross_attention_norm: Optional[str] = None,
-        cross_attention_norm_num_groups: int = 32,
-        qk_norm: Optional[str] = None,
-        added_kv_proj_dim: Optional[int] = None,
-        added_proj_bias: Optional[bool] = True,
-        norm_num_groups: Optional[int] = None,
-        spatial_norm_dim: Optional[int] = None,
-        out_bias: bool = True,
-        scale_qk: bool = True,
-        only_cross_attention: bool = False,
-        eps: float = 1e-5,
-        rescale_output_factor: float = 1.0,
-        residual_connection: bool = False,
-        _from_deprecated_attn_block: bool = False,
-        processor: Optional["AttnProcessor"] = None,
-        out_dim: int = None,
-        out_context_dim: int = None,
-        context_pre_only=None,
-        pre_only=False,
-        elementwise_affine: bool = True,
-        is_causal: bool = False,
-        **kwargs,
-    ):
-        """
-        Initialize base attention processor with common parameters.
-
-        Args:
-            query_dim: Dimension of query features
-            pbr_setting: List of PBR material types to process (e.g., ["albedo", "mr"])
-            cross_attention_dim: Dimension of cross-attention features (optional)
-            heads: Number of attention heads
-            kv_heads: Number of key-value heads for grouped query attention (optional)
-            dim_head: Dimension per attention head
-            dropout: Dropout rate
-            bias: Whether to use bias in linear projections
-            upcast_attention: Whether to upcast attention computation to float32
-            upcast_softmax: Whether to upcast softmax computation to float32
-            cross_attention_norm: Type of cross-attention normalization (optional)
-            cross_attention_norm_num_groups: Number of groups for cross-attention norm
-            qk_norm: Type of query-key normalization (optional)
-            added_kv_proj_dim: Dimension for additional key-value projections (optional)
-            added_proj_bias: Whether to use bias in additional projections
-            norm_num_groups: Number of groups for normalization (optional)
-            spatial_norm_dim: Dimension for spatial normalization (optional)
-            out_bias: Whether to use bias in output projection
-            scale_qk: Whether to scale query-key products
-            only_cross_attention: Whether to only perform cross-attention
-            eps: Small epsilon value for numerical stability
-            rescale_output_factor: Factor to rescale output values
-            residual_connection: Whether to use residual connections
-            _from_deprecated_attn_block: Flag for deprecated attention blocks
-            processor: Optional attention processor instance
-            out_dim: Output dimension (optional)
-            out_context_dim: Output context dimension (optional)
-            context_pre_only: Whether to only process context in pre-processing
-            pre_only: Whether to only perform pre-processing
-            elementwise_affine: Whether to use element-wise affine transformations
-            is_causal: Whether to use causal attention masking
-            **kwargs: Additional keyword arguments
-        """
-        super().__init__()
-        AttnUtils.check_pytorch_compatibility()
-
-        # Store common attributes
-        self.pbr_setting = pbr_setting
-        self.n_pbr_tokens = len(self.pbr_setting)
-        self.inner_dim = out_dim if out_dim is not None else dim_head * heads
-        self.inner_kv_dim = self.inner_dim if kv_heads is None else dim_head * kv_heads
-        self.query_dim = query_dim
-        self.use_bias = bias
-        self.is_cross_attention = cross_attention_dim is not None
-        self.cross_attention_dim = cross_attention_dim if cross_attention_dim is not None else query_dim
-        self.upcast_attention = upcast_attention
-        self.upcast_softmax = upcast_softmax
-        self.rescale_output_factor = rescale_output_factor
-        self.residual_connection = residual_connection
-        self.dropout = dropout
-        self.fused_projections = False
-        self.out_dim = out_dim if out_dim is not None else query_dim
-        self.out_context_dim = out_context_dim if out_context_dim is not None else query_dim
-        self.context_pre_only = context_pre_only
-        self.pre_only = pre_only
-        self.is_causal = is_causal
-        self._from_deprecated_attn_block = _from_deprecated_attn_block
-        self.scale_qk = scale_qk
-        self.scale = dim_head**-0.5 if self.scale_qk else 1.0
-        self.heads = out_dim // dim_head if out_dim is not None else heads
-        self.sliceable_head_dim = heads
-        self.added_kv_proj_dim = added_kv_proj_dim
-        self.only_cross_attention = only_cross_attention
-        self.added_proj_bias = added_proj_bias
-
-        # Validation
-        if self.added_kv_proj_dim is None and self.only_cross_attention:
-            raise ValueError(
-                "`only_cross_attention` can only be set to True if `added_kv_proj_dim` is not None."
-                "Make sure to set either `only_cross_attention=False` or define `added_kv_proj_dim`."
-            )
-
-    def register_pbr_modules(self, module_types: List[str], **kwargs):
-        """
-        Generic PBR module registration to eliminate code repetition.
-
-        Dynamically registers PyTorch modules for different PBR material types
-        based on the specified module types and PBR settings.
-
-        Args:
-            module_types: List of module types to register ("qkv", "v_only", "out", "add_kv")
-            **kwargs: Additional arguments for module configuration
-        """
-        for pbr_token in self.pbr_setting:
-            if pbr_token == "albedo":
-                continue
-
-            for module_type in module_types:
-                if module_type == "qkv":
-                    self.register_module(
-                        f"to_q_{pbr_token}", nn.Linear(self.query_dim, self.inner_dim, bias=self.use_bias)
-                    )
-                    self.register_module(
-                        f"to_k_{pbr_token}", nn.Linear(self.cross_attention_dim, self.inner_dim, bias=self.use_bias)
-                    )
-                    self.register_module(
-                        f"to_v_{pbr_token}", nn.Linear(self.cross_attention_dim, self.inner_dim, bias=self.use_bias)
-                    )
-                elif module_type == "v_only":
-                    self.register_module(
-                        f"to_v_{pbr_token}", nn.Linear(self.cross_attention_dim, self.inner_dim, bias=self.use_bias)
-                    )
-                elif module_type == "out":
-                    if not self.pre_only:
-                        self.register_module(
-                            f"to_out_{pbr_token}",
-                            nn.ModuleList(
-                                [
-                                    nn.Linear(self.inner_dim, self.out_dim, bias=kwargs.get("out_bias", True)),
-                                    nn.Dropout(self.dropout),
-                                ]
-                            ),
-                        )
-                    else:
-                        self.register_module(f"to_out_{pbr_token}", None)
-                elif module_type == "add_kv":
-                    if self.added_kv_proj_dim is not None:
-                        self.register_module(
-                            f"add_k_proj_{pbr_token}",
-                            nn.Linear(self.added_kv_proj_dim, self.inner_kv_dim, bias=self.added_proj_bias),
-                        )
-                        self.register_module(
-                            f"add_v_proj_{pbr_token}",
-                            nn.Linear(self.added_kv_proj_dim, self.inner_kv_dim, bias=self.added_proj_bias),
-                        )
-                    else:
-                        self.register_module(f"add_k_proj_{pbr_token}", None)
-                        self.register_module(f"add_v_proj_{pbr_token}", None)
-
-
-# Rotary Position Embedding utilities (specialized for PoseRoPE)
-class RotaryEmbedding:
-    """
-    Rotary position embedding utilities for 3D spatial attention.
-
-    Provides functions to compute and apply rotary position embeddings (RoPE)
-    for 1D, 3D spatial coordinates used in 3D-aware attention mechanisms.
-    """
-
-    @staticmethod
-    def get_1d_rotary_pos_embed(dim: int, pos: torch.Tensor, theta: float = 10000.0, linear_factor=1.0, ntk_factor=1.0):
-        """
-        Compute 1D rotary position embeddings.
-
-        Args:
-            dim: Embedding dimension (must be even)
-            pos: Position tensor
-            theta: Base frequency for rotary embeddings
-            linear_factor: Linear scaling factor
-            ntk_factor: NTK (Neural Tangent Kernel) scaling factor
-
-        Returns:
-            Tuple of (cos_embeddings, sin_embeddings)
-        """
-        assert dim % 2 == 0
-        theta = theta * ntk_factor
-        freqs = (
-            1.0
-            / (theta ** (torch.arange(0, dim, 2, dtype=pos.dtype, device=pos.device)[: (dim // 2)] / dim))
-            / linear_factor
-        )
-        freqs = torch.outer(pos, freqs)
-        freqs_cos = freqs.cos().repeat_interleave(2, dim=1).float()
-        freqs_sin = freqs.sin().repeat_interleave(2, dim=1).float()
-        return freqs_cos, freqs_sin
-
-    @staticmethod
-    def get_3d_rotary_pos_embed(position, embed_dim, voxel_resolution, theta: int = 10000):
-        """
-        Compute 3D rotary position embeddings for spatial coordinates.
-
-        Args:
-            position: 3D position tensor with shape [..., 3]
-            embed_dim: Embedding dimension
-            voxel_resolution: Resolution of the voxel grid
-            theta: Base frequency for rotary embeddings
-
-        Returns:
-            Tuple of (cos_embeddings, sin_embeddings) for 3D positions
-        """
-        assert position.shape[-1] == 3
-        dim_xy = embed_dim // 8 * 3
-        dim_z = embed_dim // 8 * 2
-
-        grid = torch.arange(voxel_resolution, dtype=torch.float32, device=position.device)
-        freqs_xy = RotaryEmbedding.get_1d_rotary_pos_embed(dim_xy, grid, theta=theta)
-        freqs_z = RotaryEmbedding.get_1d_rotary_pos_embed(dim_z, grid, theta=theta)
-
-        xy_cos, xy_sin = freqs_xy
-        z_cos, z_sin = freqs_z
-
-        embed_flattn = position.view(-1, position.shape[-1])
-        x_cos = xy_cos[embed_flattn[:, 0], :]
-        x_sin = xy_sin[embed_flattn[:, 0], :]
-        y_cos = xy_cos[embed_flattn[:, 1], :]
-        y_sin = xy_sin[embed_flattn[:, 1], :]
-        z_cos = z_cos[embed_flattn[:, 2], :]
-        z_sin = z_sin[embed_flattn[:, 2], :]
-
-        cos = torch.cat((x_cos, y_cos, z_cos), dim=-1)
-        sin = torch.cat((x_sin, y_sin, z_sin), dim=-1)
-
-        cos = cos.view(*position.shape[:-1], embed_dim)
-        sin = sin.view(*position.shape[:-1], embed_dim)
-        return cos, sin
-
-    @staticmethod
-    def apply_rotary_emb(x: torch.Tensor, freqs_cis: Union[torch.Tensor, Tuple[torch.Tensor]]):
-        """
-        Apply rotary position embeddings to input tensor.
-
-        Args:
-            x: Input tensor to apply rotary embeddings to
-            freqs_cis: Tuple of (cos_embeddings, sin_embeddings) or single tensor
-
-        Returns:
-            Tensor with rotary position embeddings applied
-        """
-        cos, sin = freqs_cis
-        cos, sin = cos.to(x.device), sin.to(x.device)
-        cos = cos.unsqueeze(1)
-        sin = sin.unsqueeze(1)
-
-        x_real, x_imag = x.reshape(*x.shape[:-1], -1, 2).unbind(-1)
-        x_rotated = torch.stack([-x_imag, x_real], dim=-1).flatten(3)
-
-        out = (x.float() * cos + x_rotated.float() * sin).to(x.dtype)
-        return out
-
-
-# Core attention processing logic (eliminating major duplication)
-class AttnCore:
-    """
-    Core attention processing logic shared across processors.
-
-    This class provides the fundamental attention computation pipeline
-    that can be reused across different attention processor implementations.
-    """
-
-    @staticmethod
-    def process_attention_base(
-        attn: Attention,
-        hidden_states: torch.Tensor,
-        encoder_hidden_states: Optional[torch.Tensor] = None,
-        attention_mask: Optional[torch.Tensor] = None,
-        temb: Optional[torch.Tensor] = None,
-        get_qkv_fn: Callable = None,
-        apply_rope_fn: Optional[Callable] = None,
-        **kwargs,
-    ):
-        """
-        Generic attention processing core shared across different processors.
-
-        This function implements the common attention computation pipeline including:
-        1. Hidden state preprocessing
-        2. Attention mask preparation
-        3. Q/K/V computation via provided function
-        4. Tensor reshaping for multi-head attention
-        5. Optional normalization and RoPE application
-        6. Scaled dot-product attention computation
-
-        Args:
-            attn: Attention module instance
-            hidden_states: Input hidden states tensor
-            encoder_hidden_states: Optional encoder hidden states for cross-attention
-            attention_mask: Optional attention mask tensor
-            temb: Optional temporal embedding tensor
-            get_qkv_fn: Function to compute Q, K, V tensors
-            apply_rope_fn: Optional function to apply rotary position embeddings
-            **kwargs: Additional keyword arguments passed to subfunctions
-
-        Returns:
-            Tuple containing (attention_output, residual, input_ndim, shape_info,
-            batch_size, num_heads, head_dim)
-        """
-        # Prepare hidden states
-        hidden_states, residual, input_ndim, shape_info = AttnUtils.prepare_hidden_states(hidden_states, attn, temb)
-
-        batch_size, sequence_length, _ = (
-            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
-        )
-
-        # Prepare attention mask
-        attention_mask = AttnUtils.prepare_attention_mask(attention_mask, attn, sequence_length, batch_size)
-
-        # Get Q, K, V
-        if encoder_hidden_states is None:
-            encoder_hidden_states = hidden_states
-        elif attn.norm_cross:
-            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
-
-        query, key, value = get_qkv_fn(attn, hidden_states, encoder_hidden_states, **kwargs)
-
-        # Reshape for attention
-        inner_dim = key.shape[-1]
-        head_dim = inner_dim // attn.heads
-
-        query = AttnUtils.reshape_qkv_for_attention(query, batch_size, attn.heads, head_dim)
-        key = AttnUtils.reshape_qkv_for_attention(key, batch_size, attn.heads, head_dim)
-        value = AttnUtils.reshape_qkv_for_attention(value, batch_size, attn.heads, value.shape[-1] // attn.heads)
-
-        # Apply normalization
-        query, key = AttnUtils.apply_norms(query, key, getattr(attn, "norm_q", None), getattr(attn, "norm_k", None))
-
-        # Apply RoPE if provided
-        if apply_rope_fn is not None:
-            query, key = apply_rope_fn(query, key, head_dim, **kwargs)
-
-        # Compute attention
-        hidden_states = F.scaled_dot_product_attention(
-            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
-        )
-
-        return hidden_states, residual, input_ndim, shape_info, batch_size, attn.heads, head_dim
-
-
-# Specific processor implementations (minimal unique code)
-class PoseRoPEAttnProcessor2_0:
-    """
-    Attention processor with Rotary Position Encoding (RoPE) for 3D spatial awareness.
-
-    This processor extends standard attention with 3D rotary position embeddings
-    to provide spatial awareness for 3D scene understanding tasks.
-    """
-
-    def __init__(self):
-        """Initialize the RoPE attention processor."""
-        AttnUtils.check_pytorch_compatibility()
-
-    def __call__(
-        self,
-        attn: Attention,
-        hidden_states: torch.Tensor,
-        encoder_hidden_states: Optional[torch.Tensor] = None,
-        attention_mask: Optional[torch.Tensor] = None,
-        position_indices: Dict = None,
-        temb: Optional[torch.Tensor] = None,
-        n_pbrs=1,
-        *args,
-        **kwargs,
-    ) -> torch.Tensor:
-        """
-        Apply RoPE-enhanced attention computation.
-
-        Args:
-            attn: Attention module instance
-            hidden_states: Input hidden states tensor
-            encoder_hidden_states: Optional encoder hidden states for cross-attention
-            attention_mask: Optional attention mask tensor
-            position_indices: Dictionary containing 3D position information for RoPE
-            temb: Optional temporal embedding tensor
-            n_pbrs: Number of PBR material types
-            *args: Additional positional arguments
-            **kwargs: Additional keyword arguments
-
-        Returns:
-            Attention output tensor with applied rotary position encodings
-        """
-        AttnUtils.handle_deprecation_warning(args, kwargs)
-
-        def get_qkv(attn, hidden_states, encoder_hidden_states, **kwargs):
-            return attn.to_q(hidden_states), attn.to_k(encoder_hidden_states), attn.to_v(encoder_hidden_states)
-
-        def apply_rope(query, key, head_dim, **kwargs):
-            if position_indices is not None:
-                if head_dim in position_indices:
-                    image_rotary_emb = position_indices[head_dim]
-                else:
-                    image_rotary_emb = RotaryEmbedding.get_3d_rotary_pos_embed(
-                        rearrange(
-                            position_indices["voxel_indices"].unsqueeze(1).repeat(1, n_pbrs, 1, 1),
-                            "b n_pbrs l c -> (b n_pbrs) l c",
-                        ),
-                        head_dim,
-                        voxel_resolution=position_indices["voxel_resolution"],
-                    )
-                    position_indices[head_dim] = image_rotary_emb
-
-                query = RotaryEmbedding.apply_rotary_emb(query, image_rotary_emb)
-                key = RotaryEmbedding.apply_rotary_emb(key, image_rotary_emb)
-            return query, key
-
-        # Core attention processing
-        hidden_states, residual, input_ndim, shape_info, batch_size, heads, head_dim = AttnCore.process_attention_base(
-            attn,
-            hidden_states,
-            encoder_hidden_states,
-            attention_mask,
-            temb,
-            get_qkv_fn=get_qkv,
-            apply_rope_fn=apply_rope,
-            position_indices=position_indices,
-            n_pbrs=n_pbrs,
-        )
-
-        # Finalize output
-        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, heads * head_dim)
-        hidden_states = hidden_states.to(hidden_states.dtype)
-
-        return AttnUtils.finalize_output(hidden_states, input_ndim, shape_info, attn, residual, attn.to_out)
-
-
-class SelfAttnProcessor2_0(BaseAttnProcessor):
-    """
-    Self-attention processor with PBR (Physically Based Rendering) material support.
-
-    This processor handles multiple PBR material types (e.g., albedo, metallic-roughness)
-    with separate attention computation paths for each material type.
-    """
-
-    def __init__(self, **kwargs):
-        """
-        Initialize self-attention processor with PBR support.
-
-        Args:
-            **kwargs: Arguments passed to BaseAttnProcessor initialization
-        """
-        super().__init__(**kwargs)
-        self.register_pbr_modules(["qkv", "out", "add_kv"], **kwargs)
-
-    def process_single(
-        self,
-        attn: Attention,
-        hidden_states: torch.Tensor,
-        encoder_hidden_states: Optional[torch.Tensor] = None,
-        attention_mask: Optional[torch.Tensor] = None,
-        temb: Optional[torch.Tensor] = None,
-        token: Literal["albedo", "mr"] = "albedo",
-        multiple_devices=False,
-        *args,
-        **kwargs,
-    ):
-        """
-        Process attention for a single PBR material type.
-
-        Args:
-            attn: Attention module instance
-            hidden_states: Input hidden states tensor
-            encoder_hidden_states: Optional encoder hidden states for cross-attention
-            attention_mask: Optional attention mask tensor
-            temb: Optional temporal embedding tensor
-            token: PBR material type to process ("albedo", "mr", etc.)
-            multiple_devices: Whether to use multiple GPU devices
-            *args: Additional positional arguments
-            **kwargs: Additional keyword arguments
-
-        Returns:
-            Processed attention output for the specified PBR material type
-        """
-        target = attn if token == "albedo" else attn.processor
-        token_suffix = "" if token == "albedo" else "_" + token
-
-        # Device management (if needed)
-        if multiple_devices:
-            device = torch.device("cuda:0") if token == "albedo" else torch.device("cuda:1")
-            for attr in [f"to_q{token_suffix}", f"to_k{token_suffix}", f"to_v{token_suffix}", f"to_out{token_suffix}"]:
-                getattr(target, attr).to(device)
-
-        def get_qkv(attn, hidden_states, encoder_hidden_states, **kwargs):
-            return (
-                getattr(target, f"to_q{token_suffix}")(hidden_states),
-                getattr(target, f"to_k{token_suffix}")(encoder_hidden_states),
-                getattr(target, f"to_v{token_suffix}")(encoder_hidden_states),
-            )
-
-        # Core processing using shared logic
-        hidden_states, residual, input_ndim, shape_info, batch_size, heads, head_dim = AttnCore.process_attention_base(
-            attn, hidden_states, encoder_hidden_states, attention_mask, temb, get_qkv_fn=get_qkv
-        )
-
-        # Finalize
-        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, heads * head_dim)
-        hidden_states = hidden_states.to(hidden_states.dtype)
-
-        return AttnUtils.finalize_output(
-            hidden_states, input_ndim, shape_info, attn, residual, getattr(target, f"to_out{token_suffix}")
-        )
-
-    def __call__(
-        self,
-        attn: Attention,
-        hidden_states: torch.Tensor,
-        encoder_hidden_states: Optional[torch.Tensor] = None,
-        attention_mask: Optional[torch.Tensor] = None,
-        temb: Optional[torch.Tensor] = None,
-        *args,
-        **kwargs,
-    ) -> torch.Tensor:
-        """
-        Apply self-attention with PBR material processing.
-
-        Processes multiple PBR material types sequentially, applying attention
-        computation for each material type separately and combining results.
-
-        Args:
-            attn: Attention module instance
-            hidden_states: Input hidden states tensor with PBR dimension
-            encoder_hidden_states: Optional encoder hidden states for cross-attention
-            attention_mask: Optional attention mask tensor
-            temb: Optional temporal embedding tensor
-            *args: Additional positional arguments
-            **kwargs: Additional keyword arguments
-
-        Returns:
-            Combined attention output for all PBR material types
-        """
-        AttnUtils.handle_deprecation_warning(args, kwargs)
-
-        B = hidden_states.size(0)
-        pbr_hidden_states = torch.split(hidden_states, 1, dim=1)
-
-        # Process each PBR setting
-        results = []
-        for token, pbr_hs in zip(self.pbr_setting, pbr_hidden_states):
-            processed_hs = rearrange(pbr_hs, "b n_pbrs n l c -> (b n_pbrs n) l c").to("cuda:0")
-            result = self.process_single(attn, processed_hs, None, attention_mask, temb, token, False)
-            results.append(result)
-
-        outputs = [rearrange(result, "(b n_pbrs n) l c -> b n_pbrs n l c", b=B, n_pbrs=1) for result in results]
-        return torch.cat(outputs, dim=1)
-
-
-class RefAttnProcessor2_0(BaseAttnProcessor):
-    """
-    Reference attention processor with shared value computation across PBR materials.
-
-    This processor computes query and key once, but uses separate value projections
-    for different PBR material types, enabling efficient multi-material processing.
-    """
-
-    def __init__(self, **kwargs):
-        """
-        Initialize reference attention processor.
-
-        Args:
-            **kwargs: Arguments passed to BaseAttnProcessor initialization
-        """
-        super().__init__(**kwargs)
-        self.pbr_settings = self.pbr_setting  # Alias for compatibility
-        self.register_pbr_modules(["v_only", "out"], **kwargs)
-
-    def __call__(
-        self,
-        attn: Attention,
-        hidden_states: torch.Tensor,
-        encoder_hidden_states: Optional[torch.Tensor] = None,
-        attention_mask: Optional[torch.Tensor] = None,
-        temb: Optional[torch.Tensor] = None,
-        *args,
-        **kwargs,
-    ) -> torch.Tensor:
-        """
-        Apply reference attention with shared Q/K and separate V projections.
-
-        This method computes query and key tensors once and reuses them across
-        all PBR material types, while using separate value projections for each
-        material type to maintain material-specific information.
-
-        Args:
-            attn: Attention module instance
-            hidden_states: Input hidden states tensor
-            encoder_hidden_states: Optional encoder hidden states for cross-attention
-            attention_mask: Optional attention mask tensor
-            temb: Optional temporal embedding tensor
-            *args: Additional positional arguments
-            **kwargs: Additional keyword arguments
-
-        Returns:
-            Stacked attention output for all PBR material types
-        """
-        AttnUtils.handle_deprecation_warning(args, kwargs)
-
-        def get_qkv(attn, hidden_states, encoder_hidden_states, **kwargs):
-            query = attn.to_q(hidden_states)
-            key = attn.to_k(encoder_hidden_states)
-
-            # Concatenate values from all PBR settings
-            value_list = [attn.to_v(encoder_hidden_states)]
-            for token in ["_" + token for token in self.pbr_settings if token != "albedo"]:
-                value_list.append(getattr(attn.processor, f"to_v{token}")(encoder_hidden_states))
-            value = torch.cat(value_list, dim=-1)
-
-            return query, key, value
-
-        # Core processing
-        hidden_states, residual, input_ndim, shape_info, batch_size, heads, head_dim = AttnCore.process_attention_base(
-            attn, hidden_states, encoder_hidden_states, attention_mask, temb, get_qkv_fn=get_qkv
-        )
-
-        # Split and process each PBR setting output
-        hidden_states_list = torch.split(hidden_states, head_dim, dim=-1)
-        output_hidden_states_list = []
-
-        for i, hs in enumerate(hidden_states_list):
-            hs = hs.transpose(1, 2).reshape(batch_size, -1, heads * head_dim).to(hs.dtype)
-            token_suffix = "_" + self.pbr_settings[i] if self.pbr_settings[i] != "albedo" else ""
-            target = attn if self.pbr_settings[i] == "albedo" else attn.processor
-
-            hs = AttnUtils.finalize_output(
-                hs, input_ndim, shape_info, attn, residual, getattr(target, f"to_out{token_suffix}")
-            )
-            output_hidden_states_list.append(hs)
-
-        return torch.stack(output_hidden_states_list, dim=1)

hy3dpaint/hunyuanpaintpbr/unet/model.py

@@ -1,622 +0,0 @@
-# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT
-# except for the third-party components listed below.
-# Hunyuan 3D does not impose any additional limitations beyond what is outlined
-# in the repsective licenses of these third-party components.
-# Users must comply with all terms and conditions of original licenses of these third-party
-# components and must ensure that the usage of the third party components adheres to
-# all relevant laws and regulations.
-
-# For avoidance of doubts, Hunyuan 3D means the large language models and
-# their software and algorithms, including trained model weights, parameters (including
-# optimizer states), machine-learning model code, inference-enabling code, training-enabling code,
-# fine-tuning enabling code and other elements of the foregoing made publicly available
-# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT.
-
-import os
-
-# import ipdb
-import numpy as np
-import torch
-import torch.nn as nn
-import torch.nn.functional as F
-import pytorch_lightning as pl
-from tqdm import tqdm
-from torchvision.transforms import v2
-from torchvision.utils import make_grid, save_image
-from einops import rearrange
-
-from diffusers import (
-    DiffusionPipeline,
-    EulerAncestralDiscreteScheduler,
-    DDPMScheduler,
-    UNet2DConditionModel,
-    ControlNetModel,
-)
-
-from .modules import Dino_v2, UNet2p5DConditionModel
-import math
-
-
-def extract_into_tensor(a, t, x_shape):
-    b, *_ = t.shape
-    out = a.gather(-1, t)
-    return out.reshape(b, *((1,) * (len(x_shape) - 1)))
-
-
-class HunyuanPaint(pl.LightningModule):
-    def __init__(
-        self,
-        stable_diffusion_config,
-        control_net_config=None,
-        num_view=6,
-        view_size=320,
-        drop_cond_prob=0.1,
-        with_normal_map=None,
-        with_position_map=None,
-        pbr_settings=["albedo", "mr"],
-        **kwargs,
-    ):
-        """Initializes the HunyuanPaint Lightning Module.
-
-        Args:
-            stable_diffusion_config: Configuration for loading the Stable Diffusion pipeline
-            control_net_config: Configuration for ControlNet (optional)
-            num_view: Number of views to process
-            view_size: Size of input views (height/width)
-            drop_cond_prob: Probability of dropping conditioning input during training
-            with_normal_map: Flag indicating whether normal maps are used
-            with_position_map: Flag indicating whether position maps are used
-            pbr_settings: List of PBR materials to generate (e.g., albedo, metallic-roughness)
-            **kwargs: Additional keyword arguments
-        """
-        super(HunyuanPaint, self).__init__()
-
-        self.num_view = num_view
-        self.view_size = view_size
-        self.drop_cond_prob = drop_cond_prob
-        self.pbr_settings = pbr_settings
-
-        # init modules
-        pipeline = DiffusionPipeline.from_pretrained(**stable_diffusion_config)
-        pipeline.set_pbr_settings(self.pbr_settings)
-        pipeline.scheduler = EulerAncestralDiscreteScheduler.from_config(
-            pipeline.scheduler.config, timestep_spacing="trailing"
-        )
-
-        self.with_normal_map = with_normal_map
-        self.with_position_map = with_position_map
-
-        self.pipeline = pipeline
-
-        self.pipeline.vae.use_slicing = True
-
-        train_sched = DDPMScheduler.from_config(self.pipeline.scheduler.config)
-
-        if isinstance(self.pipeline.unet, UNet2DConditionModel):
-            self.pipeline.unet = UNet2p5DConditionModel(
-                self.pipeline.unet, train_sched, self.pipeline.scheduler, self.pbr_settings
-            )
-        self.train_scheduler = train_sched  # use ddpm scheduler during training
-
-        self.register_schedule()
-
-        pipeline.set_learned_parameters()
-
-        if control_net_config is not None:
-            pipeline.unet = pipeline.unet.bfloat16().requires_grad_(control_net_config.train_unet)
-            self.pipeline.add_controlnet(
-                ControlNetModel.from_pretrained(control_net_config.pretrained_model_name_or_path),
-                conditioning_scale=0.75,
-            )
-
-        self.unet = pipeline.unet
-
-        self.pipeline.set_progress_bar_config(disable=True)
-        self.pipeline.vae = self.pipeline.vae.bfloat16()
-        self.pipeline.text_encoder = self.pipeline.text_encoder.bfloat16()
-
-        if self.unet.use_dino:
-            self.dino_v2 = Dino_v2("facebook/dinov2-giant")
-            self.dino_v2 = self.dino_v2.bfloat16()
-
-        self.validation_step_outputs = []
-
-    def register_schedule(self):
-
-        self.num_timesteps = self.train_scheduler.config.num_train_timesteps
-
-        betas = self.train_scheduler.betas.detach().cpu()
-
-        alphas = 1.0 - betas
-        alphas_cumprod = torch.cumprod(alphas, dim=0)
-        alphas_cumprod_prev = torch.cat([torch.ones(1, dtype=torch.float64), alphas_cumprod[:-1]], 0)
-
-        self.register_buffer("betas", betas.float())
-        self.register_buffer("alphas_cumprod", alphas_cumprod.float())
-        self.register_buffer("alphas_cumprod_prev", alphas_cumprod_prev.float())
-
-        # calculations for diffusion q(x_t | x_{t-1}) and others
-        self.register_buffer("sqrt_alphas_cumprod", torch.sqrt(alphas_cumprod).float())
-        self.register_buffer("sqrt_one_minus_alphas_cumprod", torch.sqrt(1 - alphas_cumprod).float())
-
-        self.register_buffer("sqrt_recip_alphas_cumprod", torch.sqrt(1.0 / alphas_cumprod).float())
-        self.register_buffer("sqrt_recipm1_alphas_cumprod", torch.sqrt(1.0 / alphas_cumprod - 1).float())
-
-    def on_fit_start(self):
-        device = torch.device(f"cuda:{self.local_rank}")
-        self.pipeline.to(device)
-        if self.global_rank == 0:
-            os.makedirs(os.path.join(self.logdir, "images_val"), exist_ok=True)
-
-    def prepare_batch_data(self, batch):
-        """Preprocesses a batch of input data for training/inference.
-
-        Args:
-            batch: Raw input batch dictionary
-
-        Returns:
-            tuple: Contains:
-                - cond_imgs: Primary conditioning images (B, 1, C, H, W)
-                - cond_imgs_another: Secondary conditioning images (B, 1, C, H, W)
-                - target_imgs: Dictionary of target PBR images resized and clamped
-                - images_normal: Preprocessed normal maps (if available)
-                - images_position: Preprocessed position maps (if available)
-        """
-
-        images_cond = batch["images_cond"].to(self.device)  # (B, M, C, H, W), where M is the number of reference images
-        cond_imgs, cond_imgs_another = images_cond[:, 0:1, ...], images_cond[:, 1:2, ...]
-
-        cond_size = self.view_size
-        cond_imgs = v2.functional.resize(cond_imgs, cond_size, interpolation=3, antialias=True).clamp(0, 1)
-        cond_imgs_another = v2.functional.resize(cond_imgs_another, cond_size, interpolation=3, antialias=True).clamp(
-            0, 1
-        )
-
-        target_imgs = {}
-        for pbr_token in self.pbr_settings:
-            target_imgs[pbr_token] = batch[f"images_{pbr_token}"].to(self.device)
-            target_imgs[pbr_token] = v2.functional.resize(
-                target_imgs[pbr_token], self.view_size, interpolation=3, antialias=True
-            ).clamp(0, 1)
-
-        images_normal = None
-        if "images_normal" in batch:
-            images_normal = batch["images_normal"]  # (B, N, C, H, W)
-            images_normal = v2.functional.resize(images_normal, self.view_size, interpolation=3, antialias=True).clamp(
-                0, 1
-            )
-            images_normal = [images_normal]
-
-        images_position = None
-        if "images_position" in batch:
-            images_position = batch["images_position"]  # (B, N, C, H, W)
-            images_position = v2.functional.resize(
-                images_position, self.view_size, interpolation=3, antialias=True
-            ).clamp(0, 1)
-            images_position = [images_position]
-
-        return cond_imgs, cond_imgs_another, target_imgs, images_normal, images_position
-
-    @torch.no_grad()
-    def forward_text_encoder(self, prompts):
-        device = next(self.pipeline.vae.parameters()).device
-        text_embeds = self.pipeline.encode_prompt(prompts, device, 1, False)[0]
-        return text_embeds
-
-    @torch.no_grad()
-    def encode_images(self, images):
-        """Encodes input images into latent representations using the VAE.
-
-        Handles both standard input (B, N, C, H, W) and PBR input (B, N_pbrs, N, C, H, W)
-        Maintains original batch structure in output latents.
-
-        Args:
-            images: Input images tensor
-
-        Returns:
-            torch.Tensor: Latent representations with original batch dimensions preserved
-        """
-
-        B = images.shape[0]
-        image_ndims = images.ndim
-        if image_ndims != 5:
-            N_pbrs, N = images.shape[1:3]
-        images = (
-            rearrange(images, "b n c h w -> (b n) c h w")
-            if image_ndims == 5
-            else rearrange(images, "b n_pbrs n c h w -> (b n_pbrs n) c h w")
-        )
-        dtype = next(self.pipeline.vae.parameters()).dtype
-
-        images = (images - 0.5) * 2.0
-        posterior = self.pipeline.vae.encode(images.to(dtype)).latent_dist
-        latents = posterior.sample() * self.pipeline.vae.config.scaling_factor
-
-        latents = (
-            rearrange(latents, "(b n) c h w -> b n c h w", b=B)
-            if image_ndims == 5
-            else rearrange(latents, "(b n_pbrs n) c h w -> b n_pbrs n c h w", b=B, n_pbrs=N_pbrs)
-        )
-
-        return latents
-
-    def forward_unet(self, latents, t, **cached_condition):
-        """Runs the UNet model to predict noise/latent residuals.
-
-        Args:
-            latents: Noisy latent representations (B, C, H, W)
-            t: Timestep tensor (B,)
-            **cached_condition: Dictionary of conditioning inputs (text embeds, reference images, etc)
-
-        Returns:
-            torch.Tensor: UNet output (predicted noise or velocity)
-        """
-
-        dtype = next(self.unet.parameters()).dtype
-        latents = latents.to(dtype)
-        shading_embeds = cached_condition["shading_embeds"]
-        pred_noise = self.pipeline.unet(latents, t, encoder_hidden_states=shading_embeds, **cached_condition)
-        return pred_noise[0]
-
-    def predict_start_from_z_and_v(self, x_t, t, v):
-        """
-        Predicts clean image (x0) from noisy latents (x_t) and
-        velocity prediction (v) using the v-prediction formula.
-
-        Args:
-            x_t: Noisy latents at timestep t
-            t: Current timestep
-            v: Predicted velocity (v) from UNet
-
-        Returns:
-            torch.Tensor: Predicted clean image (x0)
-        """
-
-        return (
-            extract_into_tensor(self.sqrt_alphas_cumprod, t, x_t.shape) * x_t
-            - extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_t.shape) * v
-        )
-
-    def get_v(self, x, noise, t):
-        """Computes the target velocity (v) for v-prediction training.
-
-        Args:
-            x: Clean latents (x0)
-            noise: Added noise
-            t: Current timestep
-
-        Returns:
-            torch.Tensor: Target velocity
-        """
-
-        return (
-            extract_into_tensor(self.sqrt_alphas_cumprod, t, x.shape) * noise
-            - extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x.shape) * x
-        )
-
-    def training_step(self, batch, batch_idx):
-        """Performs a single training step with both conditioning paths.
-
-        Implements:
-        1. Dual-conditioning path training (main ref + secondary ref)
-        2. Velocity-prediction with consistency loss
-        3. Conditional dropout for robust learning
-        4. PBR-specific losses (albedo/metallic-roughness)
-
-        Args:
-            batch: Input batch from dataloader
-            batch_idx: Index of current batch
-
-        Returns:
-            torch.Tensor: Combined loss value
-        """
-
-        cond_imgs, cond_imgs_another, target_imgs, normal_imgs, position_imgs = self.prepare_batch_data(batch)
-
-        B, N_ref = cond_imgs.shape[:2]
-        _, N_gen, _, H, W = target_imgs["albedo"].shape
-        N_pbrs = len(self.pbr_settings)
-        t = torch.randint(0, self.num_timesteps, size=(B,)).long().to(self.device)
-        t = t.unsqueeze(-1).repeat(1, N_pbrs, N_gen)
-        t = rearrange(t, "b n_pbrs n -> (b n_pbrs n)")
-
-        all_target_pbrs = []
-        for pbr_token in self.pbr_settings:
-            all_target_pbrs.append(target_imgs[pbr_token])
-        all_target_pbrs = torch.stack(all_target_pbrs, dim=0).transpose(1, 0)
-        gen_latents = self.encode_images(all_target_pbrs)  #! B, N_pbrs N C H W
-        ref_latents = self.encode_images(cond_imgs)  #! B, M, C, H, W
-        ref_latents_another = self.encode_images(cond_imgs_another)  #! B, M, C, H, W
-
-        all_shading_tokens = []
-        for token in self.pbr_settings:
-            if token in ["albedo", "mr"]:
-                all_shading_tokens.append(
-                    getattr(self.unet, f"learned_text_clip_{token}").unsqueeze(dim=0).repeat(B, 1, 1)
-                )
-        shading_embeds = torch.stack(all_shading_tokens, dim=1)
-
-        if self.unet.use_dino:
-            dino_hidden_states = self.dino_v2(cond_imgs[:, :1, ...])
-            dino_hidden_states_another = self.dino_v2(cond_imgs_another[:, :1, ...])
-
-        gen_latents = rearrange(gen_latents, "b n_pbrs n c h w -> (b n_pbrs n) c h w")
-        noise = torch.randn_like(gen_latents).to(self.device)
-        latents_noisy = self.train_scheduler.add_noise(gen_latents, noise, t).to(self.device)
-        latents_noisy = rearrange(latents_noisy, "(b n_pbrs n) c h w -> b n_pbrs n c h w", b=B, n_pbrs=N_pbrs)
-
-        cached_condition = {}
-
-        if normal_imgs is not None:
-            normal_embeds = self.encode_images(normal_imgs[0])
-            cached_condition["embeds_normal"] = normal_embeds  #! B, N, C, H, W
-
-        if position_imgs is not None:
-            position_embeds = self.encode_images(position_imgs[0])
-            cached_condition["embeds_position"] = position_embeds  #! B, N, C, H, W
-            cached_condition["position_maps"] = position_imgs[0]  #! B, N, C, H, W
-
-        for b in range(B):
-            prob = np.random.rand()
-            if prob < self.drop_cond_prob:
-                if "normal_imgs" in cached_condition:
-                    cached_condition["embeds_normal"][b, ...] = torch.zeros_like(
-                        cached_condition["embeds_normal"][b, ...]
-                    )
-                if "position_imgs" in cached_condition:
-                    cached_condition["embeds_position"][b, ...] = torch.zeros_like(
-                        cached_condition["embeds_position"][b, ...]
-                    )
-
-            prob = np.random.rand()
-            if prob < self.drop_cond_prob:
-                if "position_maps" in cached_condition:
-                    cached_condition["position_maps"][b, ...] = torch.zeros_like(
-                        cached_condition["position_maps"][b, ...]
-                    )
-
-            prob = np.random.rand()
-            if prob < self.drop_cond_prob:
-                dino_hidden_states[b, ...] = torch.zeros_like(dino_hidden_states[b, ...])
-            prob = np.random.rand()
-            if prob < self.drop_cond_prob:
-                dino_hidden_states_another[b, ...] = torch.zeros_like(dino_hidden_states_another[b, ...])
-
-        # MVA & Ref Attention
-        prob = np.random.rand()
-        cached_condition["mva_scale"] = 1.0
-        cached_condition["ref_scale"] = 1.0
-        if prob < self.drop_cond_prob:
-            cached_condition["mva_scale"] = 0.0
-            cached_condition["ref_scale"] = 0.0
-        elif prob > 1.0 - self.drop_cond_prob:
-            prob = np.random.rand()
-            if prob < 0.5:
-                cached_condition["mva_scale"] = 0.0
-            else:
-                cached_condition["ref_scale"] = 0.0
-        else:
-            pass
-
-        if self.train_scheduler.config.prediction_type == "v_prediction":
-
-            cached_condition["shading_embeds"] = shading_embeds
-            cached_condition["ref_latents"] = ref_latents
-            cached_condition["dino_hidden_states"] = dino_hidden_states
-            v_pred = self.forward_unet(latents_noisy, t, **cached_condition)
-            v_pred_albedo, v_pred_mr = torch.split(
-                rearrange(
-                    v_pred, "(b n_pbr n) c h w -> b n_pbr n c h w", n_pbr=len(self.pbr_settings), n=self.num_view
-                ),
-                1,
-                dim=1,
-            )
-            v_target = self.get_v(gen_latents, noise, t)
-            v_target_albedo, v_target_mr = torch.split(
-                rearrange(
-                    v_target, "(b n_pbr n) c h w -> b n_pbr n c h w", n_pbr=len(self.pbr_settings), n=self.num_view
-                ),
-                1,
-                dim=1,
-            )
-
-            albedo_loss_1, _ = self.compute_loss(v_pred_albedo, v_target_albedo)
-            mr_loss_1, _ = self.compute_loss(v_pred_mr, v_target_mr)
-
-            cached_condition["ref_latents"] = ref_latents_another
-            cached_condition["dino_hidden_states"] = dino_hidden_states_another
-            v_pred_another = self.forward_unet(latents_noisy, t, **cached_condition)
-            v_pred_another_albedo, v_pred_another_mr = torch.split(
-                rearrange(
-                    v_pred_another,
-                    "(b n_pbr n) c h w -> b n_pbr n c h w",
-                    n_pbr=len(self.pbr_settings),
-                    n=self.num_view,
-                ),
-                1,
-                dim=1,
-            )
-
-            albedo_loss_2, _ = self.compute_loss(v_pred_another_albedo, v_target_albedo)
-            mr_loss_2, _ = self.compute_loss(v_pred_another_mr, v_target_mr)
-
-            consistency_loss, _ = self.compute_loss(v_pred_another, v_pred)
-
-            albedo_loss = (albedo_loss_1 + albedo_loss_2) * 0.5
-            mr_loss = (mr_loss_1 + mr_loss_2) * 0.5
-
-            log_loss_dict = {}
-            log_loss_dict.update({f"train/albedo_loss": albedo_loss})
-            log_loss_dict.update({f"train/mr_loss": mr_loss})
-            log_loss_dict.update({f"train/cons_loss": consistency_loss})
-
-            loss_dict = log_loss_dict
-
-        elif self.train_scheduler.config.prediction_type == "epsilon":
-            e_pred = self.forward_unet(latents_noisy, t, **cached_condition)
-            loss, loss_dict = self.compute_loss(e_pred, noise)
-        else:
-            raise f"No {self.train_scheduler.config.prediction_type}"
-
-        # logging
-        self.log_dict(loss_dict, prog_bar=True, logger=True, on_step=True, on_epoch=True)
-        self.log("global_step", self.global_step, prog_bar=True, logger=True, on_step=True, on_epoch=False)
-        lr = self.optimizers().param_groups[0]["lr"]
-        self.log("lr_abs", lr, prog_bar=True, logger=True, on_step=True, on_epoch=False)
-
-        return 0.85 * (albedo_loss + mr_loss) + 0.15 * consistency_loss
-
-    def compute_loss(self, noise_pred, noise_gt):
-        loss = F.mse_loss(noise_pred, noise_gt)
-        prefix = "train"
-        loss_dict = {}
-        loss_dict.update({f"{prefix}/loss": loss})
-        return loss, loss_dict
-
-    @torch.no_grad()
-    def validation_step(self, batch, batch_idx):
-        """Performs validation on a single batch.
-
-        Generates predicted images using:
-        1. Reference conditioning images
-        2. Optional normal/position maps
-        3. Frozen DINO features (if enabled)
-        4. Text prompt conditioning
-
-        Compares predictions against ground truth targets and prepares visualization.
-        Stores results for epoch-level aggregation.
-
-        Args:
-            batch: Input batch from validation dataloader
-            batch_idx: Index of current batch
-        """
-        # [Validation image generation and comparison logic...]
-        # Key steps:
-        # 1. Preprocess conditioning images to PIL format
-        # 2. Set up conditioning inputs (normal maps, position maps, DINO features)
-        # 3. Run pipeline inference with fixed prompt ("high quality")
-        # 4. Decode latent outputs to image space
-        # 5. Arrange predictions and ground truths for visualization
-
-        cond_imgs_tensor, _, target_imgs, normal_imgs, position_imgs = self.prepare_batch_data(batch)
-        resolution = self.view_size
-        image_pils = []
-        for i in range(cond_imgs_tensor.shape[0]):
-            image_pils.append([])
-            for j in range(cond_imgs_tensor.shape[1]):
-                image_pils[-1].append(v2.functional.to_pil_image(cond_imgs_tensor[i, j, ...]))
-
-        outputs, gts = [], []
-        for idx in range(len(image_pils)):
-            cond_imgs = image_pils[idx]
-
-            cached_condition = dict(num_in_batch=self.num_view, N_pbrs=len(self.pbr_settings))
-            if normal_imgs is not None:
-                cached_condition["images_normal"] = normal_imgs[0][idx, ...].unsqueeze(0)
-            if position_imgs is not None:
-                cached_condition["images_position"] = position_imgs[0][idx, ...].unsqueeze(0)
-            if self.pipeline.unet.use_dino:
-                dino_hidden_states = self.dino_v2([cond_imgs][0])
-                cached_condition["dino_hidden_states"] = dino_hidden_states
-
-            latent = self.pipeline(
-                cond_imgs,
-                prompt="high quality",
-                num_inference_steps=30,
-                output_type="latent",
-                height=resolution,
-                width=resolution,
-                **cached_condition,
-            ).images
-
-            image = self.pipeline.vae.decode(latent / self.pipeline.vae.config.scaling_factor, return_dict=False)[
-                0
-            ]  # [-1, 1]
-            image = (image * 0.5 + 0.5).clamp(0, 1)
-
-            image = rearrange(
-                image, "(b n_pbr n) c h w -> b n_pbr n c h w", n_pbr=len(self.pbr_settings), n=self.num_view
-            )
-            image = torch.cat((torch.ones_like(image[:, :, :1, ...]) * 0.5, image), dim=2)
-            image = rearrange(image, "b n_pbr n c h w -> (b n_pbr n) c h w")
-            image = rearrange(
-                image,
-                "(b n_pbr n) c h w -> b c (n_pbr h) (n w)",
-                b=1,
-                n_pbr=len(self.pbr_settings),
-                n=self.num_view + 1,
-            )
-            outputs.append(image)
-
-        all_target_pbrs = []
-        for pbr_token in self.pbr_settings:
-            all_target_pbrs.append(target_imgs[pbr_token])
-        all_target_pbrs = torch.stack(all_target_pbrs, dim=0).transpose(1, 0)
-        all_target_pbrs = torch.cat(
-            (cond_imgs_tensor.unsqueeze(1).repeat(1, len(self.pbr_settings), 1, 1, 1, 1), all_target_pbrs), dim=2
-        )
-        all_target_pbrs = rearrange(all_target_pbrs, "b n_pbrs n c h w -> b c (n_pbrs h) (n w)")
-        gts = all_target_pbrs
-        outputs = torch.cat(outputs, dim=0).to(self.device)
-        images = torch.cat([gts, outputs], dim=-2)
-        self.validation_step_outputs.append(images)
-
-    @torch.no_grad()
-    def on_validation_epoch_end(self):
-        """Aggregates validation results at epoch end.
-
-        Gathers outputs from all GPUs (if distributed training),
-        creates a unified visualization grid, and saves to disk.
-        Only rank 0 process performs saving.
-        """
-        # [Result aggregation and visualization...]
-        # Key steps:
-        # 1. Gather validation outputs from all processes
-        # 2. Create image grid combining ground truths and predictions
-        # 3. Save visualization with step-numbered filename
-        # 4. Clear memory for next validation cycle
-
-        images = torch.cat(self.validation_step_outputs, dim=0)
-        all_images = self.all_gather(images)
-        all_images = rearrange(all_images, "r b c h w -> (r b) c h w")
-
-        if self.global_rank == 0:
-            grid = make_grid(all_images, nrow=8, normalize=True, value_range=(0, 1))
-            save_image(grid, os.path.join(self.logdir, "images_val", f"val_{self.global_step:07d}.png"))
-
-        self.validation_step_outputs.clear()  # free memory
-
-    def configure_optimizers(self):
-        lr = self.learning_rate
-        optimizer = torch.optim.AdamW(self.unet.parameters(), lr=lr)
-
-        def lr_lambda(step):
-            warm_up_step = 1000
-            T_step = 9000
-            gamma = 0.9
-            min_lr = 0.1 if step >= warm_up_step else 0.0
-            max_lr = 1.0
-            normalized_step = step % (warm_up_step + T_step)
-            current_max_lr = max_lr * gamma ** (step // (warm_up_step + T_step))
-            if current_max_lr < min_lr:
-                current_max_lr = min_lr
-            if normalized_step < warm_up_step:
-                lr_step = min_lr + (normalized_step / warm_up_step) * (current_max_lr - min_lr)
-            else:
-                step_wc_wp = normalized_step - warm_up_step
-                ratio = step_wc_wp / T_step
-                lr_step = min_lr + 0.5 * (current_max_lr - min_lr) * (1 + math.cos(math.pi * ratio))
-            return lr_step
-
-        lr_scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer, lr_lambda)
-
-        lr_scheduler_config = {
-            "scheduler": lr_scheduler,
-            "interval": "step",
-            "frequency": 1,
-            "monitor": "val_loss",
-            "strict": False,
-            "name": None,
-        }
-
-        return {"optimizer": optimizer, "lr_scheduler": lr_scheduler_config}

hy3dpaint/hunyuanpaintpbr/unet/modules.py

@@ -1,1102 +0,0 @@
-# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT
-# except for the third-party components listed below.
-# Hunyuan 3D does not impose any additional limitations beyond what is outlined
-# in the repsective licenses of these third-party components.
-# Users must comply with all terms and conditions of original licenses of these third-party
-# components and must ensure that the usage of the third party components adheres to
-# all relevant laws and regulations.
-
-# For avoidance of doubts, Hunyuan 3D means the large language models and
-# their software and algorithms, including trained model weights, parameters (including
-# optimizer states), machine-learning model code, inference-enabling code, training-enabling code,
-# fine-tuning enabling code and other elements of the foregoing made publicly available
-# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT.
-
-import os
-import json
-import copy
-import numpy as np
-import torch
-import torch.nn as nn
-from einops import rearrange
-from typing import Any, Callable, Dict, List, Optional, Union, Tuple, Literal
-import diffusers
-from diffusers.utils import deprecate
-from diffusers import (
-    DDPMScheduler,
-    EulerAncestralDiscreteScheduler,
-    UNet2DConditionModel,
-)
-from diffusers.models import UNet2DConditionModel
-from diffusers.models.attention_processor import Attention, AttnProcessor
-from diffusers.models.transformers.transformer_2d import BasicTransformerBlock
-from .attn_processor import SelfAttnProcessor2_0, RefAttnProcessor2_0, PoseRoPEAttnProcessor2_0
-
-from transformers import AutoImageProcessor, AutoModel
-
-
-class Dino_v2(nn.Module):
-
-    """Wrapper for DINOv2 vision transformer (frozen weights).
-    
-    Provides feature extraction for reference images.
-    
-    Args:
-        dino_v2_path: Custom path to DINOv2 model weights (uses default if None)
-    """
-
-
-    def __init__(self, dino_v2_path):
-        super(Dino_v2, self).__init__()
-        self.dino_processor = AutoImageProcessor.from_pretrained(dino_v2_path)
-        self.dino_v2 = AutoModel.from_pretrained(dino_v2_path)
-
-        for param in self.parameters():
-            param.requires_grad = False
-
-        self.dino_v2.eval()
-
-    def forward(self, images):
-
-        """Processes input images through DINOv2 ViT.
-        
-        Handles both tensor input (B, N, C, H, W) and PIL image lists.
-        Extracts patch embeddings and flattens spatial dimensions.
-        
-        Returns:
-            torch.Tensor: Feature vectors [B, N*(num_patches), feature_dim]
-        """
-
-        if isinstance(images, torch.Tensor):
-            batch_size = images.shape[0]
-            dino_proceesed_images = self.dino_processor(
-                images=rearrange(images, "b n c h w -> (b n) c h w"), return_tensors="pt", do_rescale=False
-            ).pixel_values
-        else:
-            batch_size = 1
-            dino_proceesed_images = self.dino_processor(images=images, return_tensors="pt").pixel_values
-            dino_proceesed_images = torch.stack(
-                [torch.from_numpy(np.array(image)) for image in dino_proceesed_images], dim=0
-            )
-        dino_param = next(self.dino_v2.parameters())
-        dino_proceesed_images = dino_proceesed_images.to(dino_param)
-        dino_hidden_states = self.dino_v2(dino_proceesed_images)[0]
-        dino_hidden_states = rearrange(dino_hidden_states.to(dino_param), "(b n) l c -> b (n l) c", b=batch_size)
-
-        return dino_hidden_states
-
-
-def _chunked_feed_forward(ff: nn.Module, hidden_states: torch.Tensor, chunk_dim: int, chunk_size: int):
-    # "feed_forward_chunk_size" can be used to save memory
-
-    """Memory-efficient feedforward execution via chunking.
-    
-    Divides input along specified dimension for sequential processing.
-    
-    Args:
-        ff: Feedforward module to apply
-        hidden_states: Input tensor
-        chunk_dim: Dimension to split
-        chunk_size: Size of each chunk
-        
-    Returns:
-        torch.Tensor: Reassembled output tensor
-    """
-
-    if hidden_states.shape[chunk_dim] % chunk_size != 0:
-        raise ValueError(
-            f"`hidden_states` dimension to be chunked: {hidden_states.shape[chunk_dim]}"
-            f"has to be divisible by chunk size: {chunk_size}."
-            "Make sure to set an appropriate `chunk_size` when calling `unet.enable_forward_chunking`."
-        )
-
-    num_chunks = hidden_states.shape[chunk_dim] // chunk_size
-    ff_output = torch.cat(
-        [ff(hid_slice) for hid_slice in hidden_states.chunk(num_chunks, dim=chunk_dim)],
-        dim=chunk_dim,
-    )
-    return ff_output
-
-
-@torch.no_grad()
-def compute_voxel_grid_mask(position, grid_resolution=8):
-
-    """Generates view-to-view attention mask based on 3D position similarity.
-    
-    Uses voxel grid downsampling to determine spatially adjacent regions.
-    Mask indicates where features should interact across different views.
-    
-    Args:
-        position: Position maps [B, N, 3, H, W] (normalized 0-1)
-        grid_resolution: Spatial reduction factor
-        
-    Returns:
-        torch.Tensor: Attention mask [B, N*grid_res**2, N*grid_res**2]
-    """
-
-    position = position.half()
-    B, N, _, H, W = position.shape
-    assert H % grid_resolution == 0 and W % grid_resolution == 0
-
-    valid_mask = (position != 1).all(dim=2, keepdim=True)
-    valid_mask = valid_mask.expand_as(position)
-    position[valid_mask == False] = 0
-
-    position = rearrange(
-        position,
-        "b n c (num_h grid_h) (num_w grid_w) -> b n num_h num_w c grid_h grid_w",
-        num_h=grid_resolution,
-        num_w=grid_resolution,
-    )
-    valid_mask = rearrange(
-        valid_mask,
-        "b n c (num_h grid_h) (num_w grid_w) -> b n num_h num_w c grid_h grid_w",
-        num_h=grid_resolution,
-        num_w=grid_resolution,
-    )
-
-    grid_position = position.sum(dim=(-2, -1))
-    count_masked = valid_mask.sum(dim=(-2, -1))
-
-    grid_position = grid_position / count_masked.clamp(min=1)
-    grid_position[count_masked < 5] = 0
-
-    grid_position = grid_position.permute(0, 1, 4, 2, 3)
-    grid_position = rearrange(grid_position, "b n c h w -> b n (h w) c")
-
-    grid_position_expanded_1 = grid_position.unsqueeze(2).unsqueeze(4)  # 形状变为 B, N, 1, L, 1, 3
-    grid_position_expanded_2 = grid_position.unsqueeze(1).unsqueeze(3)  # 形状变为 B, 1, N, 1, L, 3
-
-    # 计算欧氏距离
-    distances = torch.norm(grid_position_expanded_1 - grid_position_expanded_2, dim=-1)  # 形状为 B, N, N, L, L
-
-    weights = distances
-    grid_distance = 1.73 / grid_resolution
-    weights = weights < grid_distance
-
-    return weights
-
-
-def compute_multi_resolution_mask(position_maps, grid_resolutions=[32, 16, 8]):
-
-    """Generates attention masks at multiple spatial resolutions.
-    
-    Creates pyramid of position-based masks for hierarchical attention.
-    
-    Args:
-        position_maps: Position maps [B, N, 3, H, W]
-        grid_resolutions: List of downsampling factors
-        
-    Returns:
-        dict: Resolution-specific masks keyed by flattened dimension size
-    """
-
-    position_attn_mask = {}
-    with torch.no_grad():
-        for grid_resolution in grid_resolutions:
-            position_mask = compute_voxel_grid_mask(position_maps, grid_resolution)
-            position_mask = rearrange(position_mask, "b ni nj li lj -> b (ni li) (nj lj)")
-            position_attn_mask[position_mask.shape[1]] = position_mask
-    return position_attn_mask
-
-
-@torch.no_grad()
-def compute_discrete_voxel_indice(position, grid_resolution=8, voxel_resolution=128):
-
-    """Quantizes position maps to discrete voxel indices.
-    
-    Creates sparse 3D coordinate representations for efficient hashing.
-    
-    Args:
-        position: Position maps [B, N, 3, H, W]
-        grid_resolution: Spatial downsampling factor
-        voxel_resolution: Quantization resolution
-        
-    Returns:
-        torch.Tensor: Voxel indices [B, N, grid_res, grid_res, 3]
-    """
-
-    position = position.half()
-    B, N, _, H, W = position.shape
-    assert H % grid_resolution == 0 and W % grid_resolution == 0
-
-    valid_mask = (position != 1).all(dim=2, keepdim=True)
-    valid_mask = valid_mask.expand_as(position)
-    position[valid_mask == False] = 0
-
-    position = rearrange(
-        position,
-        "b n c (num_h grid_h) (num_w grid_w) -> b n num_h num_w c grid_h grid_w",
-        num_h=grid_resolution,
-        num_w=grid_resolution,
-    )
-    valid_mask = rearrange(
-        valid_mask,
-        "b n c (num_h grid_h) (num_w grid_w) -> b n num_h num_w c grid_h grid_w",
-        num_h=grid_resolution,
-        num_w=grid_resolution,
-    )
-
-    grid_position = position.sum(dim=(-2, -1))
-    count_masked = valid_mask.sum(dim=(-2, -1))
-
-    grid_position = grid_position / count_masked.clamp(min=1)
-    voxel_mask_thres = (H // grid_resolution) * (W // grid_resolution) // (4 * 4)
-    grid_position[count_masked < voxel_mask_thres] = 0
-
-    grid_position = grid_position.permute(0, 1, 4, 2, 3).clamp(0, 1)  # B N C H W
-    voxel_indices = grid_position * (voxel_resolution - 1)
-    voxel_indices = torch.round(voxel_indices).long()
-    return voxel_indices
-
-
-def calc_multires_voxel_idxs(position_maps, grid_resolutions=[64, 32, 16, 8], voxel_resolutions=[512, 256, 128, 64]):
-
-    """Generates multi-resolution voxel indices for position encoding.
-    
-    Creates pyramid of quantized position representations.
-    
-    Args:
-        position_maps: Input position maps
-        grid_resolutions: Spatial resolution levels
-        voxel_resolutions: Quantization levels
-        
-    Returns:
-        dict: Voxel indices keyed by flattened dimension size, with resolution metadata
-    """
-
-    voxel_indices = {}
-    with torch.no_grad():
-        for grid_resolution, voxel_resolution in zip(grid_resolutions, voxel_resolutions):
-            voxel_indice = compute_discrete_voxel_indice(position_maps, grid_resolution, voxel_resolution)
-            voxel_indice = rearrange(voxel_indice, "b n c h w -> b (n h w) c")
-            voxel_indices[voxel_indice.shape[1]] = {"voxel_indices": voxel_indice, "voxel_resolution": voxel_resolution}
-    return voxel_indices
-
-
-class Basic2p5DTransformerBlock(torch.nn.Module):
-
-
-    """Enhanced transformer block for multiview 2.5D image generation.
-    
-    Extends standard transformer blocks with:
-    - Material-specific attention (MDA)
-    - Multiview attention (MA)
-    - Reference attention (RA)
-    - DINO feature integration
-    
-    Args:
-        transformer: Base transformer block
-        layer_name: Identifier for layer
-        use_ma: Enable multiview attention
-        use_ra: Enable reference attention
-        use_mda: Enable material-aware attention
-        use_dino: Enable DINO feature integration
-        pbr_setting: List of PBR materials
-    """
-
-    def __init__(
-        self,
-        transformer: BasicTransformerBlock,
-        layer_name,
-        use_ma=True,
-        use_ra=True,
-        use_mda=True,
-        use_dino=True,
-        pbr_setting=None,
-    ) -> None:
-        
-        """
-        Initialization:
-        1. Material-Dimension Attention (MDA):
-           - Processes each PBR material with separate projection weights
-           - Uses custom SelfAttnProcessor2_0 with material awareness
-           
-        2. Multiview Attention (MA):
-           - Adds cross-view attention with PoseRoPE
-           - Initialized as zero-initialized residual pathway
-           
-        3. Reference Attention (RA):
-           - Conditions on reference view features
-           - Uses RefAttnProcessor2_0 for material-specific conditioning
-           
-        4. DINO Attention:
-           - Incorporates DINO-ViT features
-           - Initialized as zero-initialized residual pathway
-        """
-
-        super().__init__()
-        self.transformer = transformer
-        self.layer_name = layer_name
-        self.use_ma = use_ma
-        self.use_ra = use_ra
-        self.use_mda = use_mda
-        self.use_dino = use_dino
-        self.pbr_setting = pbr_setting
-
-        if self.use_mda:
-            self.attn1.set_processor(
-                SelfAttnProcessor2_0(
-                    query_dim=self.dim,
-                    heads=self.num_attention_heads,
-                    dim_head=self.attention_head_dim,
-                    dropout=self.dropout,
-                    bias=self.attention_bias,
-                    cross_attention_dim=None,
-                    upcast_attention=self.attn1.upcast_attention,
-                    out_bias=True,
-                    pbr_setting=self.pbr_setting,
-                )
-            )
-
-        # multiview attn
-        if self.use_ma:
-            self.attn_multiview = Attention(
-                query_dim=self.dim,
-                heads=self.num_attention_heads,
-                dim_head=self.attention_head_dim,
-                dropout=self.dropout,
-                bias=self.attention_bias,
-                cross_attention_dim=None,
-                upcast_attention=self.attn1.upcast_attention,
-                out_bias=True,
-                processor=PoseRoPEAttnProcessor2_0(),
-            )
-
-        # ref attn
-        if self.use_ra:
-            self.attn_refview = Attention(
-                query_dim=self.dim,
-                heads=self.num_attention_heads,
-                dim_head=self.attention_head_dim,
-                dropout=self.dropout,
-                bias=self.attention_bias,
-                cross_attention_dim=None,
-                upcast_attention=self.attn1.upcast_attention,
-                out_bias=True,
-                processor=RefAttnProcessor2_0(
-                    query_dim=self.dim,
-                    heads=self.num_attention_heads,
-                    dim_head=self.attention_head_dim,
-                    dropout=self.dropout,
-                    bias=self.attention_bias,
-                    cross_attention_dim=None,
-                    upcast_attention=self.attn1.upcast_attention,
-                    out_bias=True,
-                    pbr_setting=self.pbr_setting,
-                ),
-            )
-
-        # dino attn
-        if self.use_dino:
-            self.attn_dino = Attention(
-                query_dim=self.dim,
-                heads=self.num_attention_heads,
-                dim_head=self.attention_head_dim,
-                dropout=self.dropout,
-                bias=self.attention_bias,
-                cross_attention_dim=self.cross_attention_dim,
-                upcast_attention=self.attn2.upcast_attention,
-                out_bias=True,
-            )
-
-        self._initialize_attn_weights()
-
-    def _initialize_attn_weights(self):
-
-        """Initializes specialized attention heads with base weights.
-        
-        Uses weight sharing strategy:
-        - Copies base transformer weights to specialized heads
-        - Initializes newly-added parameters to zero
-        """
-
-        if self.use_mda:
-            for token in self.pbr_setting:
-                if token == "albedo":
-                    continue
-                getattr(self.attn1.processor, f"to_q_{token}").load_state_dict(self.attn1.to_q.state_dict())
-                getattr(self.attn1.processor, f"to_k_{token}").load_state_dict(self.attn1.to_k.state_dict())
-                getattr(self.attn1.processor, f"to_v_{token}").load_state_dict(self.attn1.to_v.state_dict())
-                getattr(self.attn1.processor, f"to_out_{token}").load_state_dict(self.attn1.to_out.state_dict())
-
-        if self.use_ma:
-            self.attn_multiview.load_state_dict(self.attn1.state_dict(), strict=False)
-            with torch.no_grad():
-                for layer in self.attn_multiview.to_out:
-                    for param in layer.parameters():
-                        param.zero_()
-
-        if self.use_ra:
-            self.attn_refview.load_state_dict(self.attn1.state_dict(), strict=False)
-            for token in self.pbr_setting:
-                if token == "albedo":
-                    continue
-                getattr(self.attn_refview.processor, f"to_v_{token}").load_state_dict(
-                    self.attn_refview.to_q.state_dict()
-                )
-                getattr(self.attn_refview.processor, f"to_out_{token}").load_state_dict(
-                    self.attn_refview.to_out.state_dict()
-                )
-            with torch.no_grad():
-                for layer in self.attn_refview.to_out:
-                    for param in layer.parameters():
-                        param.zero_()
-                for token in self.pbr_setting:
-                    if token == "albedo":
-                        continue
-                    for layer in getattr(self.attn_refview.processor, f"to_out_{token}"):
-                        for param in layer.parameters():
-                            param.zero_()
-
-        if self.use_dino:
-            self.attn_dino.load_state_dict(self.attn2.state_dict(), strict=False)
-            with torch.no_grad():
-                for layer in self.attn_dino.to_out:
-                    for param in layer.parameters():
-                        param.zero_()
-
-        if self.use_dino:
-            self.attn_dino.load_state_dict(self.attn2.state_dict(), strict=False)
-            with torch.no_grad():
-                for layer in self.attn_dino.to_out:
-                    for param in layer.parameters():
-                        param.zero_()
-
-    def __getattr__(self, name: str):
-        try:
-            return super().__getattr__(name)
-        except AttributeError:
-            return getattr(self.transformer, name)
-
-    def forward(
-        self,
-        hidden_states: torch.Tensor,
-        attention_mask: Optional[torch.Tensor] = None,
-        encoder_hidden_states: Optional[torch.Tensor] = None,
-        encoder_attention_mask: Optional[torch.Tensor] = None,
-        timestep: Optional[torch.LongTensor] = None,
-        cross_attention_kwargs: Dict[str, Any] = None,
-        class_labels: Optional[torch.LongTensor] = None,
-        added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
-    ) -> torch.Tensor:
-
-        """Forward pass with multi-mechanism attention.
-        
-        Processing stages:
-        1. Material-aware self-attention (MDA)
-        2. Reference attention (RA)
-        3. Multiview attention (MA) with position-aware attention
-        4. Text conditioning (base attention)
-        5. DINO feature conditioning (optional)
-        6. Position-aware conditioning
-        7. Feed-forward network
-        
-        Args:
-            hidden_states: Input features [B * N_materials * N_views, Seq_len, Feat_dim]
-            See base transformer for other parameters
-            
-        Returns:
-            torch.Tensor: Output features
-        """
-        # [Full multi-mechanism processing pipeline...]
-        # Key processing stages:
-        # 1. Material-aware self-attention (handles albedo/mr separation)
-        # 2. Reference attention (conditioned on reference features)
-        # 3. View-to-view attention with geometric constraints
-        # 4. Text-to-image cross-attention
-        # 5. DINO feature fusion (when enabled)
-        # 6. Positional conditioning (RoPE-style)
-        # 7. Feed-forward network with conditional normalization
-
-        # Notice that normalization is always applied before the real computation in the following blocks.
-        # 0. Self-Attention
-        batch_size = hidden_states.shape[0]
-
-        cross_attention_kwargs = cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {}
-        num_in_batch = cross_attention_kwargs.pop("num_in_batch", 1)
-        mode = cross_attention_kwargs.pop("mode", None)
-        mva_scale = cross_attention_kwargs.pop("mva_scale", 1.0)
-        ref_scale = cross_attention_kwargs.pop("ref_scale", 1.0)
-        condition_embed_dict = cross_attention_kwargs.pop("condition_embed_dict", None)
-        dino_hidden_states = cross_attention_kwargs.pop("dino_hidden_states", None)
-        position_voxel_indices = cross_attention_kwargs.pop("position_voxel_indices", None)
-        N_pbr = len(self.pbr_setting) if self.pbr_setting is not None else 1
-
-        if self.norm_type == "ada_norm":
-            norm_hidden_states = self.norm1(hidden_states, timestep)
-        elif self.norm_type == "ada_norm_zero":
-            norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(
-                hidden_states, timestep, class_labels, hidden_dtype=hidden_states.dtype
-            )
-        elif self.norm_type in ["layer_norm", "layer_norm_i2vgen"]:
-            norm_hidden_states = self.norm1(hidden_states)
-        elif self.norm_type == "ada_norm_continuous":
-            norm_hidden_states = self.norm1(hidden_states, added_cond_kwargs["pooled_text_emb"])
-        elif self.norm_type == "ada_norm_single":
-            shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
-                self.scale_shift_table[None] + timestep.reshape(batch_size, 6, -1)
-            ).chunk(6, dim=1)
-            norm_hidden_states = self.norm1(hidden_states)
-            norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa
-        else:
-            raise ValueError("Incorrect norm used")
-
-        if self.pos_embed is not None:
-            norm_hidden_states = self.pos_embed(norm_hidden_states)
-
-        # 1. Prepare GLIGEN inputs
-        cross_attention_kwargs = cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {}
-        gligen_kwargs = cross_attention_kwargs.pop("gligen", None)
-
-        if self.use_mda:
-            mda_norm_hidden_states = rearrange(
-                norm_hidden_states, "(b n_pbr n) l c -> b n_pbr n l c", n=num_in_batch, n_pbr=N_pbr
-            )
-            attn_output = self.attn1(
-                mda_norm_hidden_states,
-                encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,
-                attention_mask=attention_mask,
-                **cross_attention_kwargs,
-            )
-            attn_output = rearrange(attn_output, "b n_pbr n l c -> (b n_pbr n) l c")
-        else:
-            attn_output = self.attn1(
-                norm_hidden_states,
-                encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,
-                attention_mask=attention_mask,
-                **cross_attention_kwargs,
-            )
-
-        if self.norm_type == "ada_norm_zero":
-            attn_output = gate_msa.unsqueeze(1) * attn_output
-        elif self.norm_type == "ada_norm_single":
-            attn_output = gate_msa * attn_output
-
-        hidden_states = attn_output + hidden_states
-        if hidden_states.ndim == 4:
-            hidden_states = hidden_states.squeeze(1)
-
-        # 1.2 Reference Attention
-        if "w" in mode:
-            condition_embed_dict[self.layer_name] = rearrange(
-                norm_hidden_states, "(b n) l c -> b (n l) c", n=num_in_batch
-            )  # B, (N L), C
-
-        if "r" in mode and self.use_ra:
-            condition_embed = condition_embed_dict[self.layer_name]
-
-            #! Only using albedo features for reference attention
-            ref_norm_hidden_states = rearrange(
-                norm_hidden_states, "(b n_pbr n) l c -> b n_pbr (n l) c", n=num_in_batch, n_pbr=N_pbr
-            )[:, 0, ...]
-
-            attn_output = self.attn_refview(
-                ref_norm_hidden_states,
-                encoder_hidden_states=condition_embed,
-                attention_mask=None,
-                **cross_attention_kwargs,
-            )  # b (n l) c
-            attn_output = rearrange(attn_output, "b n_pbr (n l) c -> (b n_pbr n) l c", n=num_in_batch, n_pbr=N_pbr)
-
-            ref_scale_timing = ref_scale
-            if isinstance(ref_scale, torch.Tensor):
-                ref_scale_timing = ref_scale.unsqueeze(1).repeat(1, num_in_batch * N_pbr).view(-1)
-                for _ in range(attn_output.ndim - 1):
-                    ref_scale_timing = ref_scale_timing.unsqueeze(-1)
-            hidden_states = ref_scale_timing * attn_output + hidden_states
-            if hidden_states.ndim == 4:
-                hidden_states = hidden_states.squeeze(1)
-
-        # 1.3 Multiview Attention
-        if num_in_batch > 1 and self.use_ma:
-            multivew_hidden_states = rearrange(
-                norm_hidden_states, "(b n_pbr n) l c -> (b n_pbr) (n l) c", n_pbr=N_pbr, n=num_in_batch
-            )
-            position_indices = None
-            if position_voxel_indices is not None:
-                if multivew_hidden_states.shape[1] in position_voxel_indices:
-                    position_indices = position_voxel_indices[multivew_hidden_states.shape[1]]
-
-            attn_output = self.attn_multiview(
-                multivew_hidden_states,
-                encoder_hidden_states=multivew_hidden_states,
-                position_indices=position_indices,
-                n_pbrs=N_pbr,
-                **cross_attention_kwargs,
-            )
-
-            attn_output = rearrange(attn_output, "(b n_pbr) (n l) c -> (b n_pbr n) l c", n_pbr=N_pbr, n=num_in_batch)
-
-            hidden_states = mva_scale * attn_output + hidden_states
-            if hidden_states.ndim == 4:
-                hidden_states = hidden_states.squeeze(1)
-
-        # 1.2 GLIGEN Control
-        if gligen_kwargs is not None:
-            hidden_states = self.fuser(hidden_states, gligen_kwargs["objs"])
-
-        # 3. Cross-Attention
-        if self.attn2 is not None:
-            if self.norm_type == "ada_norm":
-                norm_hidden_states = self.norm2(hidden_states, timestep)
-            elif self.norm_type in ["ada_norm_zero", "layer_norm", "layer_norm_i2vgen"]:
-                norm_hidden_states = self.norm2(hidden_states)
-            elif self.norm_type == "ada_norm_single":
-                # For PixArt norm2 isn't applied here:
-                # https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L70C1-L76C103
-                norm_hidden_states = hidden_states
-            elif self.norm_type == "ada_norm_continuous":
-                norm_hidden_states = self.norm2(hidden_states, added_cond_kwargs["pooled_text_emb"])
-            else:
-                raise ValueError("Incorrect norm")
-
-            if self.pos_embed is not None and self.norm_type != "ada_norm_single":
-                norm_hidden_states = self.pos_embed(norm_hidden_states)
-
-            attn_output = self.attn2(
-                norm_hidden_states,
-                encoder_hidden_states=encoder_hidden_states,
-                attention_mask=encoder_attention_mask,
-                **cross_attention_kwargs,
-            )
-            hidden_states = attn_output + hidden_states
-
-            # dino attn
-            if self.use_dino:
-                dino_hidden_states = dino_hidden_states.unsqueeze(1).repeat(1, N_pbr * num_in_batch, 1, 1)
-                dino_hidden_states = rearrange(dino_hidden_states, "b n l c -> (b n) l c")
-                attn_output = self.attn_dino(
-                    norm_hidden_states,
-                    encoder_hidden_states=dino_hidden_states,
-                    attention_mask=None,
-                    **cross_attention_kwargs,
-                )
-
-                hidden_states = attn_output + hidden_states
-
-        # 4. Feed-forward
-        # i2vgen doesn't have this norm 🤷‍♂️
-        if self.norm_type == "ada_norm_continuous":
-            norm_hidden_states = self.norm3(hidden_states, added_cond_kwargs["pooled_text_emb"])
-        elif not self.norm_type == "ada_norm_single":
-            norm_hidden_states = self.norm3(hidden_states)
-
-        if self.norm_type == "ada_norm_zero":
-            norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
-
-        if self.norm_type == "ada_norm_single":
-            norm_hidden_states = self.norm2(hidden_states)
-            norm_hidden_states = norm_hidden_states * (1 + scale_mlp) + shift_mlp
-
-        if self._chunk_size is not None:
-            # "feed_forward_chunk_size" can be used to save memory
-            ff_output = _chunked_feed_forward(self.ff, norm_hidden_states, self._chunk_dim, self._chunk_size)
-        else:
-            ff_output = self.ff(norm_hidden_states)
-
-        if self.norm_type == "ada_norm_zero":
-            ff_output = gate_mlp.unsqueeze(1) * ff_output
-        elif self.norm_type == "ada_norm_single":
-            ff_output = gate_mlp * ff_output
-
-        hidden_states = ff_output + hidden_states
-        if hidden_states.ndim == 4:
-            hidden_states = hidden_states.squeeze(1)
-
-        return hidden_states
-
-
-class ImageProjModel(torch.nn.Module):
-
-    """Projects image embeddings into cross-attention space.
-    
-    Transforms CLIP embeddings into additional context tokens for conditioning.
-    
-    Args:
-        cross_attention_dim: Dimension of attention space
-        clip_embeddings_dim: Dimension of input CLIP embeddings
-        clip_extra_context_tokens: Number of context tokens to generate
-    """
-
-    def __init__(self, cross_attention_dim=1024, clip_embeddings_dim=1024, clip_extra_context_tokens=4):
-        super().__init__()
-
-        self.generator = None
-        self.cross_attention_dim = cross_attention_dim
-        self.clip_extra_context_tokens = clip_extra_context_tokens
-        self.proj = torch.nn.Linear(clip_embeddings_dim, self.clip_extra_context_tokens * cross_attention_dim)
-        self.norm = torch.nn.LayerNorm(cross_attention_dim)
-
-    def forward(self, image_embeds):
-
-        """Projects image embeddings to cross-attention context tokens.
-        
-        Args:
-            image_embeds: Input embeddings [B, N, C] or [B, C]
-            
-        Returns:
-            torch.Tensor: Context tokens [B, N*clip_extra_context_tokens, cross_attention_dim]
-        """
-
-        embeds = image_embeds
-        num_token = 1
-        if embeds.dim() == 3:
-            num_token = embeds.shape[1]
-            embeds = rearrange(embeds, "b n c -> (b n) c")
-
-        clip_extra_context_tokens = self.proj(embeds).reshape(
-            -1, self.clip_extra_context_tokens, self.cross_attention_dim
-        )
-        clip_extra_context_tokens = self.norm(clip_extra_context_tokens)
-
-        clip_extra_context_tokens = rearrange(clip_extra_context_tokens, "(b nt) n c -> b (nt n) c", nt=num_token)
-
-        return clip_extra_context_tokens
-
-
-class UNet2p5DConditionModel(torch.nn.Module):
-
-    """2.5D UNet extension for multiview PBR generation.
-    
-    Enhances standard 2D UNet with:
-    - Multiview attention mechanisms
-    - Material-aware processing
-    - Position-aware conditioning
-    - Dual-stream reference processing
-    
-    Args:
-        unet: Base 2D UNet model
-        train_sched: Training scheduler (DDPM)
-        val_sched: Validation scheduler (EulerAncestral)
-    """
-
-    def __init__(
-        self,
-        unet: UNet2DConditionModel,
-        train_sched: DDPMScheduler = None,
-        val_sched: EulerAncestralDiscreteScheduler = None,
-    ) -> None:
-        super().__init__()
-        self.unet = unet
-        self.train_sched = train_sched
-        self.val_sched = val_sched
-
-        self.use_ma = True
-        self.use_ra = True
-        self.use_mda = True
-        self.use_dino = True
-        self.use_position_rope = True
-        self.use_learned_text_clip = True
-        self.use_dual_stream = True
-        self.pbr_setting = ["albedo", "mr"]
-        self.pbr_token_channels = 77
-
-        if self.use_dual_stream and self.use_ra:
-            self.unet_dual = copy.deepcopy(unet)
-            self.init_attention(self.unet_dual)
-
-        self.init_attention(
-            self.unet,
-            use_ma=self.use_ma,
-            use_ra=self.use_ra,
-            use_dino=self.use_dino,
-            use_mda=self.use_mda,
-            pbr_setting=self.pbr_setting,
-        )
-        self.init_condition(use_dino=self.use_dino)
-
-    @staticmethod
-    def from_pretrained(pretrained_model_name_or_path, **kwargs):
-        torch_dtype = kwargs.pop("torch_dtype", torch.float32)
-        config_path = os.path.join(pretrained_model_name_or_path, "config.json")
-        unet_ckpt_path = os.path.join(pretrained_model_name_or_path, "diffusion_pytorch_model.bin")
-        with open(config_path, "r", encoding="utf-8") as file:
-            config = json.load(file)
-        unet = UNet2DConditionModel(**config)
-        unet_2p5d = UNet2p5DConditionModel(unet)
-        unet_2p5d.unet.conv_in = torch.nn.Conv2d(
-            12,
-            unet.conv_in.out_channels,
-            kernel_size=unet.conv_in.kernel_size,
-            stride=unet.conv_in.stride,
-            padding=unet.conv_in.padding,
-            dilation=unet.conv_in.dilation,
-            groups=unet.conv_in.groups,
-            bias=unet.conv_in.bias is not None,
-        )
-        unet_ckpt = torch.load(unet_ckpt_path, map_location="cpu", weights_only=True)
-        unet_2p5d.load_state_dict(unet_ckpt, strict=True)
-        unet_2p5d = unet_2p5d.to(torch_dtype)
-        return unet_2p5d
-
-    def init_condition(self, use_dino):
-
-        """Initializes conditioning mechanisms for multiview PBR generation.
-        
-        Sets up:
-        1. Learned text embeddings: Material-specific tokens (albedo, mr) initialized to zeros
-        2. DINO projector: Model to process DINO-ViT features for cross-attention
-        
-        Args:
-            use_dino: Flag to enable DINO feature integration
-        """
-
-        if self.use_learned_text_clip:
-            for token in self.pbr_setting:
-                self.unet.register_parameter(
-                    f"learned_text_clip_{token}", nn.Parameter(torch.zeros(self.pbr_token_channels, 1024))
-                )
-            self.unet.learned_text_clip_ref = nn.Parameter(torch.zeros(self.pbr_token_channels, 1024))
-
-        if use_dino:
-            self.unet.image_proj_model_dino = ImageProjModel(
-                cross_attention_dim=self.unet.config.cross_attention_dim,
-                clip_embeddings_dim=1536,
-                clip_extra_context_tokens=4,
-            )
-
-    def init_attention(self, unet, use_ma=False, use_ra=False, use_mda=False, use_dino=False, pbr_setting=None):
-
-        """Recursively replaces standard transformers with enhanced 2.5D blocks.
-        
-        Processes UNet architecture:
-        1. Downsampling blocks: Replaces transformers in attention layers
-        2. Middle block: Upgrades central transformers
-        3. Upsampling blocks: Modifies decoder transformers
-        
-        Args:
-            unet: UNet model to enhance
-            use_ma: Enable multiview attention
-            use_ra: Enable reference attention
-            use_mda: Enable material-specific attention
-            use_dino: Enable DINO feature integration
-            pbr_setting: List of PBR materials
-        """
-
-        for down_block_i, down_block in enumerate(unet.down_blocks):
-            if hasattr(down_block, "has_cross_attention") and down_block.has_cross_attention:
-                for attn_i, attn in enumerate(down_block.attentions):
-                    for transformer_i, transformer in enumerate(attn.transformer_blocks):
-                        if isinstance(transformer, BasicTransformerBlock):
-                            attn.transformer_blocks[transformer_i] = Basic2p5DTransformerBlock(
-                                transformer,
-                                f"down_{down_block_i}_{attn_i}_{transformer_i}",
-                                use_ma,
-                                use_ra,
-                                use_mda,
-                                use_dino,
-                                pbr_setting,
-                            )
-
-        if hasattr(unet.mid_block, "has_cross_attention") and unet.mid_block.has_cross_attention:
-            for attn_i, attn in enumerate(unet.mid_block.attentions):
-                for transformer_i, transformer in enumerate(attn.transformer_blocks):
-                    if isinstance(transformer, BasicTransformerBlock):
-                        attn.transformer_blocks[transformer_i] = Basic2p5DTransformerBlock(
-                            transformer, f"mid_{attn_i}_{transformer_i}", use_ma, use_ra, use_mda, use_dino, pbr_setting
-                        )
-
-        for up_block_i, up_block in enumerate(unet.up_blocks):
-            if hasattr(up_block, "has_cross_attention") and up_block.has_cross_attention:
-                for attn_i, attn in enumerate(up_block.attentions):
-                    for transformer_i, transformer in enumerate(attn.transformer_blocks):
-                        if isinstance(transformer, BasicTransformerBlock):
-                            attn.transformer_blocks[transformer_i] = Basic2p5DTransformerBlock(
-                                transformer,
-                                f"up_{up_block_i}_{attn_i}_{transformer_i}",
-                                use_ma,
-                                use_ra,
-                                use_mda,
-                                use_dino,
-                                pbr_setting,
-                            )
-
-    def __getattr__(self, name: str):
-        try:
-            return super().__getattr__(name)
-        except AttributeError:
-            return getattr(self.unet, name)
-
-    def forward(
-        self,
-        sample,
-        timestep,
-        encoder_hidden_states,
-        *args,
-        added_cond_kwargs=None,
-        cross_attention_kwargs=None,
-        down_intrablock_additional_residuals=None,
-        down_block_res_samples=None,
-        mid_block_res_sample=None,
-        **cached_condition,
-    ):
-
-        """Forward pass with multiview/material conditioning.
-        
-        Key stages:
-        1. Input preparation (concat normal/position maps)
-        2. Reference feature extraction (dual-stream)
-        3. Position encoding (voxel indices)
-        4. DINO feature projection
-        5. Main UNet processing with attention conditioning
-        
-        Args:
-            sample: Input latents [B, N_pbr, N_gen, C, H, W]
-            cached_condition: Dictionary containing:
-                - embeds_normal: Normal map embeddings
-                - embeds_position: Position map embeddings
-                - ref_latents: Reference image latents
-                - dino_hidden_states: DINO features
-                - position_maps: 3D position maps
-                - mva_scale: Multiview attention scale
-                - ref_scale: Reference attention scale
-                
-        Returns:
-            torch.Tensor: Output features
-        """
-
-        B, N_pbr, N_gen, _, H, W = sample.shape
-        assert H == W
-
-        if "cache" not in cached_condition:
-            cached_condition["cache"] = {}
-
-        sample = [sample]
-        if "embeds_normal" in cached_condition:
-            sample.append(cached_condition["embeds_normal"].unsqueeze(1).repeat(1, N_pbr, 1, 1, 1, 1))
-        if "embeds_position" in cached_condition:
-            sample.append(cached_condition["embeds_position"].unsqueeze(1).repeat(1, N_pbr, 1, 1, 1, 1))
-        sample = torch.cat(sample, dim=-3)
-
-        sample = rearrange(sample, "b n_pbr n c h w -> (b n_pbr n) c h w")
-
-        encoder_hidden_states_gen = encoder_hidden_states.unsqueeze(-3).repeat(1, 1, N_gen, 1, 1)
-        encoder_hidden_states_gen = rearrange(encoder_hidden_states_gen, "b n_pbr n l c -> (b n_pbr n) l c")
-
-        if added_cond_kwargs is not None:
-            text_embeds_gen = added_cond_kwargs["text_embeds"].unsqueeze(1).repeat(1, N_gen, 1)
-            text_embeds_gen = rearrange(text_embeds_gen, "b n c -> (b n) c")
-            time_ids_gen = added_cond_kwargs["time_ids"].unsqueeze(1).repeat(1, N_gen, 1)
-            time_ids_gen = rearrange(time_ids_gen, "b n c -> (b n) c")
-            added_cond_kwargs_gen = {"text_embeds": text_embeds_gen, "time_ids": time_ids_gen}
-        else:
-            added_cond_kwargs_gen = None
-
-        if self.use_position_rope:
-            if "position_voxel_indices" in cached_condition["cache"]:
-                position_voxel_indices = cached_condition["cache"]["position_voxel_indices"]
-            else:
-                if "position_maps" in cached_condition:
-                    position_voxel_indices = calc_multires_voxel_idxs(
-                        cached_condition["position_maps"],
-                        grid_resolutions=[H, H // 2, H // 4, H // 8],
-                        voxel_resolutions=[H * 8, H * 4, H * 2, H],
-                    )
-                    cached_condition["cache"]["position_voxel_indices"] = position_voxel_indices
-        else:
-            position_voxel_indices = None
-
-        if self.use_dino:
-            if "dino_hidden_states_proj" in cached_condition["cache"]:
-                dino_hidden_states = cached_condition["cache"]["dino_hidden_states_proj"]
-            else:
-                assert "dino_hidden_states" in cached_condition
-                dino_hidden_states = cached_condition["dino_hidden_states"]
-                dino_hidden_states = self.image_proj_model_dino(dino_hidden_states)
-                cached_condition["cache"]["dino_hidden_states_proj"] = dino_hidden_states
-        else:
-            dino_hidden_states = None
-
-        if self.use_ra:
-            if "condition_embed_dict" in cached_condition["cache"]:
-                condition_embed_dict = cached_condition["cache"]["condition_embed_dict"]
-            else:
-                condition_embed_dict = {}
-                ref_latents = cached_condition["ref_latents"]
-                N_ref = ref_latents.shape[1]
-
-                if not self.use_dual_stream:
-                    ref_latents = [ref_latents]
-                    if "embeds_normal" in cached_condition:
-                        ref_latents.append(torch.zeros_like(ref_latents[0]))
-                    if "embeds_position" in cached_condition:
-                        ref_latents.append(torch.zeros_like(ref_latents[0]))
-                    ref_latents = torch.cat(ref_latents, dim=2)
-
-                ref_latents = rearrange(ref_latents, "b n c h w -> (b n) c h w")
-
-                encoder_hidden_states_ref = self.unet.learned_text_clip_ref.repeat(B, N_ref, 1, 1)
-
-                encoder_hidden_states_ref = rearrange(encoder_hidden_states_ref, "b n l c -> (b n) l c")
-
-                if added_cond_kwargs is not None:
-                    text_embeds_ref = added_cond_kwargs["text_embeds"].unsqueeze(1).repeat(1, N_ref, 1)
-                    text_embeds_ref = rearrange(text_embeds_ref, "b n c -> (b n) c")
-                    time_ids_ref = added_cond_kwargs["time_ids"].unsqueeze(1).repeat(1, N_ref, 1)
-                    time_ids_ref = rearrange(time_ids_ref, "b n c -> (b n) c")
-                    added_cond_kwargs_ref = {
-                        "text_embeds": text_embeds_ref,
-                        "time_ids": time_ids_ref,
-                    }
-                else:
-                    added_cond_kwargs_ref = None
-
-                noisy_ref_latents = ref_latents
-                timestep_ref = 0
-                if self.use_dual_stream:
-                    unet_ref = self.unet_dual
-                else:
-                    unet_ref = self.unet
-                unet_ref(
-                    noisy_ref_latents,
-                    timestep_ref,
-                    encoder_hidden_states=encoder_hidden_states_ref,
-                    class_labels=None,
-                    added_cond_kwargs=added_cond_kwargs_ref,
-                    # **kwargs
-                    return_dict=False,
-                    cross_attention_kwargs={
-                        "mode": "w",
-                        "num_in_batch": N_ref,
-                        "condition_embed_dict": condition_embed_dict,
-                    },
-                )
-                cached_condition["cache"]["condition_embed_dict"] = condition_embed_dict
-        else:
-            condition_embed_dict = None
-
-        mva_scale = cached_condition.get("mva_scale", 1.0)
-        ref_scale = cached_condition.get("ref_scale", 1.0)
-
-        return self.unet(
-            sample,
-            timestep,
-            encoder_hidden_states_gen,
-            *args,
-            class_labels=None,
-            added_cond_kwargs=added_cond_kwargs_gen,
-            down_intrablock_additional_residuals=(
-                [sample.to(dtype=self.unet.dtype) for sample in down_intrablock_additional_residuals]
-                if down_intrablock_additional_residuals is not None
-                else None
-            ),
-            down_block_additional_residuals=(
-                [sample.to(dtype=self.unet.dtype) for sample in down_block_res_samples]
-                if down_block_res_samples is not None
-                else None
-            ),
-            mid_block_additional_residual=(
-                mid_block_res_sample.to(dtype=self.unet.dtype) if mid_block_res_sample is not None else None
-            ),
-            return_dict=False,
-            cross_attention_kwargs={
-                "mode": "r",
-                "num_in_batch": N_gen,
-                "dino_hidden_states": dino_hidden_states,
-                "condition_embed_dict": condition_embed_dict,
-                "mva_scale": mva_scale,
-                "ref_scale": ref_scale,
-                "position_voxel_indices": position_voxel_indices,
-            },
-        )

hy3dpaint/packages/custom_rasterizer/custom_rasterizer/__init__.py

@@ -1,4 +0,0 @@
-"""
-from .render import rasterize, interpolate
-"""
-from .render import *

hy3dpaint/packages/custom_rasterizer/custom_rasterizer/render.py

@@ -1,32 +0,0 @@
-# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT
-# except for the third-party components listed below.
-# Hunyuan 3D does not impose any additional limitations beyond what is outlined
-# in the repsective licenses of these third-party components.
-# Users must comply with all terms and conditions of original licenses of these third-party
-# components and must ensure that the usage of the third party components adheres to
-# all relevant laws and regulations.
-
-# For avoidance of doubts, Hunyuan 3D means the large language models and
-# their software and algorithms, including trained model weights, parameters (including
-# optimizer states), machine-learning model code, inference-enabling code, training-enabling code,
-# fine-tuning enabling code and other elements of the foregoing made publicly available
-# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT.
-
-import custom_rasterizer_kernel
-import torch
-
-
-def rasterize(pos, tri, resolution, clamp_depth=torch.zeros(0), use_depth_prior=0):
-    assert pos.device == tri.device
-    findices, barycentric = custom_rasterizer_kernel.rasterize_image(
-        pos[0], tri, clamp_depth, resolution[1], resolution[0], 1e-6, use_depth_prior
-    )
-    return findices, barycentric
-
-
-def interpolate(col, findices, barycentric, tri):
-    f = findices - 1 + (findices == 0)
-    vcol = col[0, tri.long()[f.long()]]
-    result = barycentric.view(*barycentric.shape, 1) * vcol
-    result = torch.sum(result, axis=-2)
-    return result.view(1, *result.shape)

hy3dpaint/packages/custom_rasterizer/setup.py

@@ -1,40 +0,0 @@
-# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT
-# except for the third-party components listed below.
-# Hunyuan 3D does not impose any additional limitations beyond what is outlined
-# in the repsective licenses of these third-party components.
-# Users must comply with all terms and conditions of original licenses of these third-party
-# components and must ensure that the usage of the third party components adheres to
-# all relevant laws and regulations.
-
-# For avoidance of doubts, Hunyuan 3D means the large language models and
-# their software and algorithms, including trained model weights, parameters (including
-# optimizer states), machine-learning model code, inference-enabling code, training-enabling code,
-# fine-tuning enabling code and other elements of the foregoing made publicly available
-# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT.
-
-from setuptools import setup, find_packages
-import torch
-from torch.utils.cpp_extension import BuildExtension, CUDAExtension, CppExtension
-
-# build custom rasterizer
-
-custom_rasterizer_module = CUDAExtension(
-    "custom_rasterizer_kernel",
-    [
-        "lib/custom_rasterizer_kernel/rasterizer.cpp",
-        "lib/custom_rasterizer_kernel/grid_neighbor.cpp",
-        "lib/custom_rasterizer_kernel/rasterizer_gpu.cu",
-    ],
-)
-
-setup(
-    packages=find_packages(),
-    version="0.1",
-    name="custom_rasterizer",
-    include_package_data=True,
-    package_dir={"": "."},
-    ext_modules=[
-        custom_rasterizer_module,
-    ],
-    cmdclass={"build_ext": BuildExtension},
-)

hy3dpaint/src/__init__.py

@@ -1,13 +0,0 @@
-# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT
-# except for the third-party components listed below.
-# Hunyuan 3D does not impose any additional limitations beyond what is outlined
-# in the repsective licenses of these third-party components.
-# Users must comply with all terms and conditions of original licenses of these third-party
-# components and must ensure that the usage of the third party components adheres to
-# all relevant laws and regulations.
-
-# For avoidance of doubts, Hunyuan 3D means the large language models and
-# their software and algorithms, including trained model weights, parameters (including
-# optimizer states), machine-learning model code, inference-enabling code, training-enabling code,
-# fine-tuning enabling code and other elements of the foregoing made publicly available
-# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT.

hy3dpaint/src/utils/__init__.py

@@ -1,13 +0,0 @@
-# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT
-# except for the third-party components listed below.
-# Hunyuan 3D does not impose any additional limitations beyond what is outlined
-# in the repsective licenses of these third-party components.
-# Users must comply with all terms and conditions of original licenses of these third-party
-# components and must ensure that the usage of the third party components adheres to
-# all relevant laws and regulations.
-
-# For avoidance of doubts, Hunyuan 3D means the large language models and
-# their software and algorithms, including trained model weights, parameters (including
-# optimizer states), machine-learning model code, inference-enabling code, training-enabling code,
-# fine-tuning enabling code and other elements of the foregoing made publicly available
-# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT.

hy3dpaint/src/utils/train_util.py

@@ -1,40 +0,0 @@
-# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT
-# except for the third-party components listed below.
-# Hunyuan 3D does not impose any additional limitations beyond what is outlined
-# in the repsective licenses of these third-party components.
-# Users must comply with all terms and conditions of original licenses of these third-party
-# components and must ensure that the usage of the third party components adheres to
-# all relevant laws and regulations.
-
-# For avoidance of doubts, Hunyuan 3D means the large language models and
-# their software and algorithms, including trained model weights, parameters (including
-# optimizer states), machine-learning model code, inference-enabling code, training-enabling code,
-# fine-tuning enabling code and other elements of the foregoing made publicly available
-# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT.
-
-import importlib
-
-
-def count_params(model, verbose=False):
-    total_params = sum(p.numel() for p in model.parameters())
-    if verbose:
-        print(f"{model.__class__.__name__} has {total_params*1.e-6:.2f} M params.")
-    return total_params
-
-
-def instantiate_from_config(config):
-    if not "target" in config:
-        if config == "__is_first_stage__":
-            return None
-        elif config == "__is_unconditional__":
-            return None
-        raise KeyError("Expected key `target` to instantiate.")
-    return get_obj_from_str(config["target"])(**config.get("params", dict()))
-
-
-def get_obj_from_str(string, reload=False):
-    module, cls = string.rsplit(".", 1)
-    if reload:
-        module_imp = importlib.import_module(module)
-        importlib.reload(module_imp)
-    return getattr(importlib.import_module(module, package=None), cls)

hy3dpaint/textureGenPipeline.py

@@ -1,192 +0,0 @@
-# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT
-# except for the third-party components listed below.
-# Hunyuan 3D does not impose any additional limitations beyond what is outlined
-# in the repsective licenses of these third-party components.
-# Users must comply with all terms and conditions of original licenses of these third-party
-# components and must ensure that the usage of the third party components adheres to
-# all relevant laws and regulations.
-
-# For avoidance of doubts, Hunyuan 3D means the large language models and
-# their software and algorithms, including trained model weights, parameters (including
-# optimizer states), machine-learning model code, inference-enabling code, training-enabling code,
-# fine-tuning enabling code and other elements of the foregoing made publicly available
-# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT.
-
-import os
-import torch
-import copy
-import trimesh
-import numpy as np
-from PIL import Image
-from typing import List
-from DifferentiableRenderer.MeshRender import MeshRender
-from utils.simplify_mesh_utils import remesh_mesh
-from utils.multiview_utils import multiviewDiffusionNet
-from utils.pipeline_utils import ViewProcessor
-from utils.image_super_utils import imageSuperNet
-from utils.uvwrap_utils import mesh_uv_wrap
-from DifferentiableRenderer.mesh_utils import convert_obj_to_glb
-import warnings
-
-warnings.filterwarnings("ignore")
-from diffusers.utils import logging as diffusers_logging
-
-diffusers_logging.set_verbosity(50)
-
-
-class Hunyuan3DPaintConfig:
-    def __init__(self, max_num_view, resolution):
-        self.device = "cuda"
-
-        self.multiview_cfg_path = "cfgs/hunyuan-paint-pbr.yaml"
-        self.custom_pipeline = "hunyuanpaintpbr"
-        self.multiview_pretrained_path = "tencent/Hunyuan3D-2.1"
-        self.dino_ckpt_path = "facebook/dinov2-giant"
-        self.realesrgan_ckpt_path = "ckpt/RealESRGAN_x4plus.pth"
-
-        self.raster_mode = "cr"
-        self.bake_mode = "back_sample"
-        self.render_size = 1024 * 2
-        self.texture_size = 1024 * 4
-        self.max_selected_view_num = max_num_view
-        self.resolution = resolution
-        self.bake_exp = 4
-        self.merge_method = "fast"
-
-        # view selection
-        self.candidate_camera_azims = [0, 90, 180, 270, 0, 180]
-        self.candidate_camera_elevs = [0, 0, 0, 0, 90, -90]
-        self.candidate_view_weights = [1, 0.1, 0.5, 0.1, 0.05, 0.05]
-
-        for azim in range(0, 360, 30):
-            self.candidate_camera_azims.append(azim)
-            self.candidate_camera_elevs.append(20)
-            self.candidate_view_weights.append(0.01)
-
-            self.candidate_camera_azims.append(azim)
-            self.candidate_camera_elevs.append(-20)
-            self.candidate_view_weights.append(0.01)
-
-
-class Hunyuan3DPaintPipeline:
-
-    def __init__(self, config=None) -> None:
-        self.config = config if config is not None else Hunyuan3DPaintConfig()
-        self.models = {}
-        self.stats_logs = {}
-        self.render = MeshRender(
-            default_resolution=self.config.render_size,
-            texture_size=self.config.texture_size,
-            bake_mode=self.config.bake_mode,
-            raster_mode=self.config.raster_mode,
-        )
-        self.view_processor = ViewProcessor(self.config, self.render)
-        self.load_models()
-
-    def load_models(self):
-        torch.cuda.empty_cache()
-        self.models["super_model"] = imageSuperNet(self.config)
-        self.models["multiview_model"] = multiviewDiffusionNet(self.config)
-        print("Models Loaded.")
-
-    @torch.no_grad()
-    def __call__(self, mesh_path=None, image_path=None, output_mesh_path=None, use_remesh=True, save_glb=True):
-        """Generate texture for 3D mesh using multiview diffusion"""
-        # Ensure image_prompt is a list
-        if isinstance(image_path, str):
-            image_prompt = Image.open(image_path)
-        elif isinstance(image_path, Image.Image):
-            image_prompt = image_path
-        if not isinstance(image_prompt, List):
-            image_prompt = [image_prompt]
-        else:
-            image_prompt = image_path
-
-        # Process mesh
-        path = os.path.dirname(mesh_path)
-        if use_remesh:
-            processed_mesh_path = os.path.join(path, "white_mesh_remesh.obj")
-            remesh_mesh(mesh_path, processed_mesh_path)
-        else:
-            processed_mesh_path = mesh_path
-
-        # Output path
-        if output_mesh_path is None:
-            output_mesh_path = os.path.join(path, f"textured_mesh.obj")
-
-        # Load mesh
-        mesh = trimesh.load(processed_mesh_path)
-        mesh = mesh_uv_wrap(mesh)
-        self.render.load_mesh(mesh=mesh)
-
-        ########### View Selection #########
-        selected_camera_elevs, selected_camera_azims, selected_view_weights = self.view_processor.bake_view_selection(
-            self.config.candidate_camera_elevs,
-            self.config.candidate_camera_azims,
-            self.config.candidate_view_weights,
-            self.config.max_selected_view_num,
-        )
-
-        normal_maps = self.view_processor.render_normal_multiview(
-            selected_camera_elevs, selected_camera_azims, use_abs_coor=True
-        )
-        position_maps = self.view_processor.render_position_multiview(selected_camera_elevs, selected_camera_azims)
-
-        ##########  Style  ###########
-        image_caption = "high quality"
-        image_style = []
-        for image in image_prompt:
-            image = image.resize((512, 512))
-            if image.mode == "RGBA":
-                white_bg = Image.new("RGB", image.size, (255, 255, 255))
-                white_bg.paste(image, mask=image.getchannel("A"))
-                image = white_bg
-            image_style.append(image)
-        image_style = [image.convert("RGB") for image in image_style]
-
-        ###########  Multiview  ##########
-        multiviews_pbr = self.models["multiview_model"](
-            image_style,
-            normal_maps + position_maps,
-            prompt=image_caption,
-            custom_view_size=self.config.resolution,
-            resize_input=True,
-        )
-        ###########  Enhance  ##########
-        enhance_images = {}
-        enhance_images["albedo"] = copy.deepcopy(multiviews_pbr["albedo"])
-        enhance_images["mr"] = copy.deepcopy(multiviews_pbr["mr"])
-
-        for i in range(len(enhance_images["albedo"])):
-            enhance_images["albedo"][i] = self.models["super_model"](enhance_images["albedo"][i])
-            enhance_images["mr"][i] = self.models["super_model"](enhance_images["mr"][i])
-
-        ###########  Bake  ##########
-        for i in range(len(enhance_images)):
-            enhance_images["albedo"][i] = enhance_images["albedo"][i].resize(
-                (self.config.render_size, self.config.render_size)
-            )
-            enhance_images["mr"][i] = enhance_images["mr"][i].resize((self.config.render_size, self.config.render_size))
-        texture, mask = self.view_processor.bake_from_multiview(
-            enhance_images["albedo"], selected_camera_elevs, selected_camera_azims, selected_view_weights
-        )
-        mask_np = (mask.squeeze(-1).cpu().numpy() * 255).astype(np.uint8)
-        texture_mr, mask_mr = self.view_processor.bake_from_multiview(
-            enhance_images["mr"], selected_camera_elevs, selected_camera_azims, selected_view_weights
-        )
-        mask_mr_np = (mask_mr.squeeze(-1).cpu().numpy() * 255).astype(np.uint8)
-
-        ##########  inpaint  ###########
-        texture = self.view_processor.texture_inpaint(texture, mask_np)
-        self.render.set_texture(texture, force_set=True)
-        if "mr" in enhance_images:
-            texture_mr = self.view_processor.texture_inpaint(texture_mr, mask_mr_np)
-            self.render.set_texture_mr(texture_mr)
-
-        self.render.save_mesh(output_mesh_path, downsample=True)
-
-        if save_glb:
-            convert_obj_to_glb(output_mesh_path, output_mesh_path.replace(".obj", ".glb"))
-            output_glb_path = output_mesh_path.replace(".obj", ".glb")
-
-        return output_mesh_path

hy3dpaint/train.py

@@ -1,401 +0,0 @@
-# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT
-# except for the third-party components listed below.
-# Hunyuan 3D does not impose any additional limitations beyond what is outlined
-# in the repsective licenses of these third-party components.
-# Users must comply with all terms and conditions of original licenses of these third-party
-# components and must ensure that the usage of the third party components adheres to
-# all relevant laws and regulations.
-
-# For avoidance of doubts, Hunyuan 3D means the large language models and
-# their software and algorithms, including trained model weights, parameters (including
-# optimizer states), machine-learning model code, inference-enabling code, training-enabling code,
-# fine-tuning enabling code and other elements of the foregoing made publicly available
-# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT.
-
-import torch
-import os, sys
-import argparse
-import shutil
-import subprocess
-from omegaconf import OmegaConf
-
-from pytorch_lightning import seed_everything
-from pytorch_lightning.trainer import Trainer
-from pytorch_lightning.strategies import DDPStrategy
-from pytorch_lightning.callbacks import Callback
-from pytorch_lightning.utilities import rank_zero_only, rank_zero_warn
-
-from src.utils.train_util import instantiate_from_config
-import warnings
-
-warnings.filterwarnings("ignore")
-from diffusers.utils import logging as diffusers_logging
-
-diffusers_logging.set_verbosity(50)
-
-
-@rank_zero_only
-def rank_zero_print(*args):
-    print(*args)
-
-
-def get_parser(**parser_kwargs):
-    def str2bool(v):
-        if isinstance(v, bool):
-            return v
-        if v.lower() in ("yes", "true", "t", "y", "1"):
-            return True
-        elif v.lower() in ("no", "false", "f", "n", "0"):
-            return False
-        else:
-            raise argparse.ArgumentTypeError("Boolean value expected.")
-
-    parser = argparse.ArgumentParser(**parser_kwargs)
-    parser.add_argument(
-        "-r",
-        "--resume",
-        type=str,
-        default=None,
-        help="resume from checkpoint",
-    )
-    parser.add_argument(
-        "--resume_weights_only",
-        action="store_true",
-        help="only resume model weights",
-    )
-    parser.add_argument(
-        "-b",
-        "--base",
-        type=str,
-        default="base_config.yaml",
-        help="path to base configs",
-    )
-    parser.add_argument(
-        "-n",
-        "--name",
-        type=str,
-        default="",
-        help="experiment name",
-    )
-    parser.add_argument(
-        "--num_nodes",
-        type=int,
-        default=1,
-        help="number of nodes to use",
-    )
-    parser.add_argument(
-        "--gpus",
-        type=str,
-        default="0,",
-        help="gpu ids to use",
-    )
-    parser.add_argument(
-        "-s",
-        "--seed",
-        type=int,
-        default=42,
-        help="seed for seed_everything",
-    )
-    parser.add_argument(
-        "-l",
-        "--logdir",
-        type=str,
-        default="logs",
-        help="directory for logging data",
-    )
-    return parser
-
-
-class SetupCallback(Callback):
-    def __init__(self, resume, logdir, ckptdir, cfgdir, config):
-        super().__init__()
-        self.resume = resume
-        self.logdir = logdir
-        self.ckptdir = ckptdir
-        self.cfgdir = cfgdir
-        self.config = config
-
-    def on_fit_start(self, trainer, pl_module):
-        if trainer.global_rank == 0:
-            # Create logdirs and save configs
-            os.makedirs(self.logdir, exist_ok=True)
-            os.makedirs(self.ckptdir, exist_ok=True)
-            os.makedirs(self.cfgdir, exist_ok=True)
-
-            rank_zero_print("Project config")
-            rank_zero_print(OmegaConf.to_yaml(self.config))
-            OmegaConf.save(self.config, os.path.join(self.cfgdir, "project.yaml"))
-
-
-class CodeSnapshot(Callback):
-    """
-    Modified from https://github.com/threestudio-project/threestudio/blob/main/threestudio/utils/callbacks.py#L60
-    """
-
-    def __init__(self, savedir):
-        self.savedir = savedir
-
-    def get_file_list(self):
-        return [
-            b.decode()
-            for b in set(subprocess.check_output('git ls-files -- ":!:configs/*"', shell=True).splitlines())
-            | set(  # hard code, TODO: use config to exclude folders or files
-                subprocess.check_output("git ls-files --others --exclude-standard", shell=True).splitlines()
-            )
-        ]
-
-    @rank_zero_only
-    def save_code_snapshot(self):
-        os.makedirs(self.savedir, exist_ok=True)
-
-    #       for f in self.get_file_list():
-    #           if not os.path.exists(f) or os.path.isdir(f):
-    #               continue
-    #           os.makedirs(os.path.join(self.savedir, os.path.dirname(f)), exist_ok=True)
-    #           shutil.copyfile(f, os.path.join(self.savedir, f))
-
-    def on_fit_start(self, trainer, pl_module):
-        try:
-            self.save_code_snapshot()
-        except:
-            rank_zero_warn(
-                "Code snapshot is not saved. Please make sure you have git installed and are in a git repository."
-            )
-
-
-if __name__ == "__main__":
-    # add cwd for convenience and to make classes in this file available when
-    # running as `python main.py`
-    sys.path.append(os.getcwd())
-    torch.set_float32_matmul_precision("medium")
-
-    parser = get_parser()
-    opt, unknown = parser.parse_known_args()
-
-    cfg_fname = os.path.split(opt.base)[-1]
-    cfg_name = os.path.splitext(cfg_fname)[0]
-    exp_name = "-" + opt.name if opt.name != "" else ""
-    logdir = os.path.join(opt.logdir, cfg_name + exp_name)
-
-    # assert not os.path.exists(logdir) or 'test' in logdir, logdir
-    if os.path.exists(logdir) and opt.resume is None:
-        auto_resume_path = os.path.join(logdir, "checkpoints", "last.ckpt")
-        if os.path.exists(auto_resume_path):
-            opt.resume = auto_resume_path
-            print(f"Auto set resume ckpt {opt.resume}")
-
-    ckptdir = os.path.join(logdir, "checkpoints")
-    cfgdir = os.path.join(logdir, "configs")
-    codedir = os.path.join(logdir, "code")
-
-    node_rank = int(os.environ.get("NODE_RANK", 0))  # 当前节点的编号
-    local_rank = int(os.environ.get("LOCAL_RANK", 0))  # 当前节点上的 GPU 编号
-    num_gpus_per_node = torch.cuda.device_count()  # 每个节点上的 GPU 数量
-
-    global_rank = node_rank * num_gpus_per_node + local_rank
-    seed_everything(opt.seed + global_rank)
-
-    # init configs
-    config = OmegaConf.load(opt.base)
-    lightning_config = config.lightning
-    trainer_config = lightning_config.trainer
-
-    trainer_config["accelerator"] = "gpu"
-    rank_zero_print(f"Running on GPUs {opt.gpus}")
-    try:
-        ngpu = int(opt.gpus)
-    except:
-        ngpu = len(opt.gpus.strip(",").split(","))
-    trainer_config["devices"] = ngpu
-
-    trainer_opt = argparse.Namespace(**trainer_config)
-    lightning_config.trainer = trainer_config
-
-    # model
-    model = instantiate_from_config(config.model)
-
-    model_unet = model.unet.unet
-    model_unet_prefix = "unet.unet."
-    if hasattr(model_unet, "unet"):
-        model_unet = model_unet.unet
-        model_unet_prefix += "unet."
-
-    if getattr(config, "init_unet_from", None):
-        unet_ckpt_path = config.init_unet_from
-        sd = torch.load(unet_ckpt_path, map_location="cpu")
-        model_unet.load_state_dict(sd, strict=True)
-
-    if getattr(config, "init_vae_from", None):
-        vae_ckpt_path = config.init_vae_from
-        sd_vae = torch.load(vae_ckpt_path, map_location="cpu")
-
-        def replace_key(key_str):
-            replace_pairs = [("key", "to_k"), ("query", "to_q"), ("value", "to_v"), ("proj_attn", "to_out.0")]
-            for replace_pair in replace_pairs:
-                key_str = key_str.replace(replace_pair[0], replace_pair[1])
-            return key_str
-
-        sd_vae = {replace_key(k): v for k, v in sd_vae.items()}
-        model.pipeline.vae.load_state_dict(sd_vae, strict=True)
-
-    if hasattr(model.unet, "controlnet"):
-        if getattr(config, "init_control_from", None):
-            unet_ckpt_path = config.init_control_from
-            sd_control = torch.load(unet_ckpt_path, map_location="cpu")
-            model.unet.controlnet.load(sd_control, strict=True)
-
-    noise_in_channels = config.model.params.get("noise_in_channels", None)
-    if noise_in_channels is not None:
-        with torch.no_grad():
-            new_conv_in = torch.nn.Conv2d(
-                noise_in_channels,
-                model_unet.conv_in.out_channels,
-                model_unet.conv_in.kernel_size,
-                model_unet.conv_in.stride,
-                model_unet.conv_in.padding,
-            )
-            new_conv_in.weight.zero_()
-            new_conv_in.weight[:, : model_unet.conv_in.in_channels, :, :].copy_(model_unet.conv_in.weight)
-
-            new_conv_in.bias.zero_()
-            new_conv_in.bias[: model_unet.conv_in.bias.size(0)].copy_(model_unet.conv_in.bias)
-
-            model_unet.conv_in = new_conv_in
-
-    if hasattr(model.unet, "controlnet"):
-        if config.model.params.get("control_in_channels", None):
-            control_in_channels = config.model.params.control_in_channels
-            model.unet.controlnet.config["conditioning_channels"] = control_in_channels
-            condition_conv_in = model.unet.controlnet.controlnet_cond_embedding.conv_in
-
-            new_condition_conv_in = torch.nn.Conv2d(
-                control_in_channels,
-                condition_conv_in.out_channels,
-                kernel_size=condition_conv_in.kernel_size,
-                stride=condition_conv_in.stride,
-                padding=condition_conv_in.padding,
-            )
-
-            with torch.no_grad():
-                new_condition_conv_in.weight[:, : condition_conv_in.in_channels, :, :] = condition_conv_in.weight
-                if condition_conv_in.bias is not None:
-                    new_condition_conv_in.bias = condition_conv_in.bias
-
-            model.unet.controlnet.controlnet_cond_embedding.conv_in = new_condition_conv_in
-
-    rank_zero_print(f"Loaded Init ...")
-
-    if getattr(config, "resume_from", None):
-        cnet_ckpt_path = config.resume_from
-        sds = torch.load(cnet_ckpt_path, map_location="cpu")["state_dict"]
-        sd0 = {k[len(model_unet_prefix) :]: v for k, v in sds.items() if model_unet_prefix in k}
-        # model.unet.unet.unet.load_state_dict(sd0, strict=True)
-        model_unet.load_state_dict(sd0, strict=True)
-        if hasattr(model.unet, "controlnet"):
-            sd1 = {k[16:]: v for k, v in sds.items() if "unet.controlnet." in k}
-            model.unet.controlnet.load_state_dict(sd1, strict=True)
-        rank_zero_print(f"Loaded {cnet_ckpt_path} ...")
-
-    if opt.resume and opt.resume_weights_only:
-        model = model.__class__.load_from_checkpoint(opt.resume, **config.model.params)
-
-    model.logdir = logdir
-
-    # trainer and callbacks
-    trainer_kwargs = dict()
-
-    # logger
-    default_logger_cfg = {
-        "target": "pytorch_lightning.loggers.TensorBoardLogger",
-        "params": {
-            "name": "tensorboard",
-            "save_dir": logdir,
-            "version": "0",
-        },
-    }
-    logger_cfg = OmegaConf.merge(default_logger_cfg)
-    trainer_kwargs["logger"] = instantiate_from_config(logger_cfg)
-
-    # model checkpoint
-    default_modelckpt_cfg = {
-        "target": "pytorch_lightning.callbacks.ModelCheckpoint",
-        "params": {
-            "dirpath": ckptdir,
-            "filename": "{step:08}",
-            "verbose": True,
-            "save_last": True,
-            "every_n_train_steps": 5000,
-            "save_top_k": -1,  # save all checkpoints
-        },
-    }
-
-    if "modelcheckpoint" in lightning_config:
-        modelckpt_cfg = lightning_config.modelcheckpoint
-    else:
-        modelckpt_cfg = OmegaConf.create()
-    modelckpt_cfg = OmegaConf.merge(default_modelckpt_cfg, modelckpt_cfg)
-
-    # callbacks
-    default_callbacks_cfg = {
-        "setup_callback": {
-            "target": "train.SetupCallback",
-            "params": {
-                "resume": opt.resume,
-                "logdir": logdir,
-                "ckptdir": ckptdir,
-                "cfgdir": cfgdir,
-                "config": config,
-            },
-        },
-        "learning_rate_logger": {
-            "target": "pytorch_lightning.callbacks.LearningRateMonitor",
-            "params": {
-                "logging_interval": "step",
-            },
-        },
-        "code_snapshot": {
-            "target": "train.CodeSnapshot",
-            "params": {
-                "savedir": codedir,
-            },
-        },
-    }
-    default_callbacks_cfg["checkpoint_callback"] = modelckpt_cfg
-
-    if "callbacks" in lightning_config:
-        callbacks_cfg = lightning_config.callbacks
-    else:
-        callbacks_cfg = OmegaConf.create()
-    callbacks_cfg = OmegaConf.merge(default_callbacks_cfg, callbacks_cfg)
-
-    trainer_kwargs["callbacks"] = [instantiate_from_config(callbacks_cfg[k]) for k in callbacks_cfg]
-
-    trainer_kwargs["precision"] = "bf16"
-    trainer_kwargs["strategy"] = DDPStrategy(find_unused_parameters=False)
-
-    # trainer
-    trainer = Trainer(**trainer_config, **trainer_kwargs, num_nodes=opt.num_nodes, inference_mode=False)
-    trainer.logdir = logdir
-
-    # data
-    data = instantiate_from_config(config.data)
-    data.prepare_data()
-    data.setup("fit")
-
-    # configure learning rate
-    base_lr = config.model.base_learning_rate
-    if "accumulate_grad_batches" in lightning_config.trainer:
-        accumulate_grad_batches = lightning_config.trainer.accumulate_grad_batches
-    else:
-        accumulate_grad_batches = 1
-    rank_zero_print(f"accumulate_grad_batches = {accumulate_grad_batches}")
-    lightning_config.trainer.accumulate_grad_batches = accumulate_grad_batches
-    model.learning_rate = base_lr
-    rank_zero_print("++++ NOT USING LR SCALING ++++")
-    rank_zero_print(f"Setting learning rate to {model.learning_rate:.2e}")
-
-    # run training loop
-    if opt.resume and not opt.resume_weights_only:
-        trainer.fit(model, data, ckpt_path=opt.resume)
-    else:
-        trainer.fit(model, data)

hy3dpaint/train_examples/001/render_cond/transforms.json

@@ -1,838 +0,0 @@
-{
-    "aabb": [
-        [
-            -0.5,
-            -0.5,
-            -0.5
-        ],
-        [
-            0.5,
-            0.5,
-            0.5
-        ]
-    ],
-    "scale": 0.5121457915021708,
-    "offset": [
-        0.00032189488410949707,
-        -0.03408946841955185,
-        -0.2757369577884674
-    ],
-    "frames": [
-        {
-            "file_path": "000.png",
-            "camera_angle_x": 0.47011032380678586,
-            "proj_type": 0,
-            "azimuth": 2.2539364673716054,
-            "elevation": -1.3001257577149983,
-            "cam_dis": 3.71849691615005,
-            "transform_matrix": [
-                [
-                    -0.7755944728851318,
-                    -0.6082496643066406,
-                    -0.1687772274017334,
-                    -0.6275975704193115
-                ],
-                [
-                    -0.6312316656112671,
-                    0.7473564147949219,
-                    0.2073766440153122,
-                    0.7711292505264282
-                ],
-                [
-                    -2.3015424233108206e-08,
-                    0.26737767457962036,
-                    -0.9635918736457825,
-                    -3.583113193511963
-                ],
-                [
-                    0,
-                    0,
-                    0,
-                    1
-                ]
-            ]
-        },
-        {
-            "file_path": "001.png",
-            "camera_angle_x": 0.5079176868006297,
-            "proj_type": 0,
-            "azimuth": 5.395529120961399,
-            "elevation": -0.9221878379342444,
-            "cam_dis": 3.4470348893999567,
-            "transform_matrix": [
-                [
-                    0.7755942940711975,
-                    0.5030443668365479,
-                    0.3813132345676422,
-                    1.3143998384475708
-                ],
-                [
-                    0.6312316656112671,
-                    -0.6180905103683472,
-                    -0.4685195982456207,
-                    -1.615003228187561
-                ],
-                [
-                    -4.003486608894491e-08,
-                    0.6040781140327454,
-                    -0.7969250082969666,
-                    -2.7470288276672363
-                ],
-                [
-                    0,
-                    0,
-                    0,
-                    1
-                ]
-            ]
-        },
-        {
-            "file_path": "002.png",
-            "camera_angle_x": 0.7090660480350758,
-            "proj_type": 0,
-            "azimuth": 3.824732794166502,
-            "elevation": -0.6818860998422269,
-            "cam_dis": 2.494654312477568,
-            "transform_matrix": [
-                [
-                    0.6312316060066223,
-                    -0.4888249933719635,
-                    -0.6021600961685181,
-                    -1.5021814107894897
-                ],
-                [
-                    -0.7755944132804871,
-                    -0.39783918857574463,
-                    -0.49007895588874817,
-                    -1.222577691078186
-                ],
-                [
-                    -2.0854262317016037e-08,
-                    0.7763853669166565,
-                    -0.6302586197853088,
-                    -1.5722770690917969
-                ],
-                [
-                    0,
-                    0,
-                    0,
-                    1
-                ]
-            ]
-        },
-        {
-            "file_path": "003.png",
-            "camera_angle_x": 0.7207915782104436,
-            "proj_type": 0,
-            "azimuth": 6.966325447756295,
-            "elevation": -0.48204466653151123,
-            "cam_dis": 2.4558020914544034,
-            "transform_matrix": [
-                [
-                    -0.6312316060066223,
-                    0.35955917835235596,
-                    0.6872146725654602,
-                    1.687662959098816
-                ],
-                [
-                    0.7755944728851318,
-                    0.292633593082428,
-                    0.5593021512031555,
-                    1.37353515625
-                ],
-                [
-                    7.981980587601356e-08,
-                    0.8860490322113037,
-                    -0.46359172463417053,
-                    -1.1384897232055664
-                ],
-                [
-                    0,
-                    0,
-                    0,
-                    1
-                ]
-            ]
-        },
-        {
-            "file_path": "004.png",
-            "camera_angle_x": 1.1560563050396928,
-            "proj_type": 0,
-            "azimuth": 3.0393346307690536,
-            "elevation": -0.30147096310802857,
-            "cam_dis": 1.5850428518685533,
-            "transform_matrix": [
-                [
-                    -0.10207992047071457,
-                    -0.2953740656375885,
-                    -0.9499125480651855,
-                    -1.5056520700454712
-                ],
-                [
-                    -0.99477618932724,
-                    0.03031015396118164,
-                    0.09747619181871414,
-                    0.15450391173362732
-                ],
-                [
-                    4.873770365065866e-08,
-                    0.9549007415771484,
-                    -0.2969251275062561,
-                    -0.47063907980918884
-                ],
-                [
-                    0,
-                    0,
-                    0,
-                    1
-                ]
-            ]
-        },
-        {
-            "file_path": "005.png",
-            "camera_angle_x": 1.1080063733527263,
-            "proj_type": 0,
-            "azimuth": 6.180927284358847,
-            "elevation": -0.13062968094200955,
-            "cam_dis": 1.6461361738590004,
-            "transform_matrix": [
-                [
-                    0.10207971930503845,
-                    0.1295779049396515,
-                    0.9863009452819824,
-                    1.6235853433609009
-                ],
-                [
-                    0.9947763085365295,
-                    -0.013296757824718952,
-                    -0.10121013969182968,
-                    -0.16660575568675995
-                ],
-                [
-                    -1.0836848218787054e-07,
-                    0.9914802312850952,
-                    -0.1302584558725357,
-                    -0.21442320942878723
-                ],
-                [
-                    0,
-                    0,
-                    0,
-                    1
-                ]
-            ]
-        },
-        {
-            "file_path": "006.png",
-            "camera_angle_x": 1.0631377508801954,
-            "proj_type": 0,
-            "azimuth": 4.610130957563951,
-            "elevation": 0.012136358647873546,
-            "cam_dis": 1.7085198579433762,
-            "transform_matrix": [
-                [
-                    0.99477618932724,
-                    0.0012388412142172456,
-                    -0.10207240283489227,
-                    -0.1743927001953125
-                ],
-                [
-                    -0.10207992047071457,
-                    0.012072510085999966,
-                    -0.9947029948234558,
-                    -1.6994696855545044
-                ],
-                [
-                    -2.03229033601815e-09,
-                    0.9999264478683472,
-                    0.012135906144976616,
-                    0.020734701305627823
-                ],
-                [
-                    0,
-                    0,
-                    0,
-                    1
-                ]
-            ]
-        },
-        {
-            "file_path": "007.png",
-            "camera_angle_x": 0.5894741433793819,
-            "proj_type": 0,
-            "azimuth": 7.751723611153743,
-            "elevation": 0.06774341874087009,
-            "cam_dis": 2.9812749570228485,
-            "transform_matrix": [
-                [
-                    -0.99477618932724,
-                    -0.006909956689924002,
-                    0.10184576362371445,
-                    0.30363020300865173
-                ],
-                [
-                    0.10207990556955338,
-                    -0.06733796000480652,
-                    0.9924944639205933,
-                    2.9588990211486816
-                ],
-                [
-                    -8.816746444040291e-09,
-                    0.9977062940597534,
-                    0.06769154965877533,
-                    0.20180732011795044
-                ],
-                [
-                    0,
-                    0,
-                    0,
-                    1
-                ]
-            ]
-        },
-        {
-            "file_path": "008.png",
-            "camera_angle_x": 1.1731446025768484,
-            "proj_type": 0,
-            "azimuth": 2.6466355490703295,
-            "elevation": 0.12356134208787495,
-            "cam_dis": 1.5646078921687379,
-            "transform_matrix": [
-                [
-                    -0.47499409317970276,
-                    0.10845614224672318,
-                    -0.8732801675796509,
-                    -1.3663408756256104
-                ],
-                [
-                    -0.8799892067909241,
-                    -0.058541711419820786,
-                    0.4713726341724396,
-                    0.7375132441520691
-                ],
-                [
-                    5.2883926571212214e-08,
-                    0.9923761487007141,
-                    0.12324699014425278,
-                    0.19283349812030792
-                ],
-                [
-                    0,
-                    0,
-                    0,
-                    1
-                ]
-            ]
-        },
-        {
-            "file_path": "009.png",
-            "camera_angle_x": 0.9240898657044195,
-            "proj_type": 0,
-            "azimuth": 5.788228202660123,
-            "elevation": 0.17976943361645015,
-            "cam_dis": 1.9427212715365068,
-            "transform_matrix": [
-                [
-                    0.4749939441680908,
-                    -0.15734440088272095,
-                    0.8658080101013184,
-                    1.6820236444473267
-                ],
-                [
-                    0.8799890875816345,
-                    0.08493020385503769,
-                    -0.4673393964767456,
-                    -0.9079101085662842
-                ],
-                [
-                    -2.0080790363863343e-08,
-                    0.9838849902153015,
-                    0.17880268394947052,
-                    0.3473638594150543
-                ],
-                [
-                    0,
-                    0,
-                    0,
-                    1
-                ]
-            ]
-        },
-        {
-            "file_path": "010.png",
-            "camera_angle_x": 1.2064601116642644,
-            "proj_type": 0,
-            "azimuth": 4.217431875865226,
-            "elevation": 0.23655841484996287,
-            "cam_dis": 1.526559413456265,
-            "transform_matrix": [
-                [
-                    0.8799890875816345,
-                    0.11131872981786728,
-                    -0.46176549792289734,
-                    -0.7049124240875244
-                ],
-                [
-                    -0.47499391436576843,
-                    0.20623266696929932,
-                    -0.855481743812561,
-                    -1.3059436082839966
-                ],
-                [
-                    2.947104427164504e-08,
-                    0.9721503853797913,
-                    0.23435820639133453,
-                    0.35776183009147644
-                ],
-                [
-                    0,
-                    0,
-                    0,
-                    1
-                ]
-            ]
-        },
-        {
-            "file_path": "011.png",
-            "camera_angle_x": 0.8106133703864192,
-            "proj_type": 0,
-            "azimuth": 7.359024529455018,
-            "elevation": 0.2941368085661682,
-            "cam_dis": 2.196358226678013,
-            "transform_matrix": [
-                [
-                    -0.8799890875816345,
-                    -0.13770732283592224,
-                    0.454594224691391,
-                    0.9984517693519592
-                ],
-                [
-                    0.47499391436576843,
-                    -0.25512099266052246,
-                    0.8421959280967712,
-                    1.849764108657837
-                ],
-                [
-                    -1.5272490827555885e-08,
-                    0.9570527672767639,
-                    0.2899138629436493,
-                    0.6367546319961548
-                ],
-                [
-                    0,
-                    0,
-                    0,
-                    1
-                ]
-            ]
-        },
-        {
-            "file_path": "012.png",
-            "camera_angle_x": 0.6777962932590744,
-            "proj_type": 0,
-            "azimuth": 3.4320337124677778,
-            "elevation": 0.35273892986621935,
-            "cam_dis": 2.604994196713999,
-            "transform_matrix": [
-                [
-                    0.2863747477531433,
-                    0.33100035786628723,
-                    -0.8991264700889587,
-                    -2.3422188758850098
-                ],
-                [
-                    -0.9581177234649658,
-                    0.09893371164798737,
-                    -0.2687426805496216,
-                    -0.7000733017921448
-                ],
-                [
-                    5.095670019272802e-08,
-                    0.9384300708770752,
-                    0.3454693555831909,
-                    0.8999457955360413
-                ],
-                [
-                    0,
-                    0,
-                    0,
-                    1
-                ]
-            ]
-        },
-        {
-            "file_path": "013.png",
-            "camera_angle_x": 1.2201901414150222,
-            "proj_type": 0,
-            "azimuth": 6.57362636605757,
-            "elevation": 0.4126354310485474,
-            "cam_dis": 1.511532217466212,
-            "transform_matrix": [
-                [
-                    -0.2863748371601105,
-                    -0.38422903418540955,
-                    0.8777002096176147,
-                    1.3266719579696655
-                ],
-                [
-                    0.9581177234649658,
-                    -0.11484345048666,
-                    0.2623385488986969,
-                    0.39653322100639343
-                ],
-                [
-                    -1.850063036101801e-08,
-                    0.9160672426223755,
-                    0.40102484822273254,
-                    0.6061621308326721
-                ],
-                [
-                    0,
-                    0,
-                    0,
-                    1
-                ]
-            ]
-        },
-        {
-            "file_path": "014.png",
-            "camera_angle_x": 0.5705820549491503,
-            "proj_type": 0,
-            "azimuth": 5.002830039262674,
-            "elevation": 0.4741478890444091,
-            "cam_dis": 3.077158450915264,
-            "transform_matrix": [
-                [
-                    0.958117663860321,
-                    -0.13075314462184906,
-                    0.2547825574874878,
-                    0.7840063571929932
-                ],
-                [
-                    0.2863748073577881,
-                    0.4374578595161438,
-                    -0.8524201512336731,
-                    -2.6230318546295166
-                ],
-                [
-                    -1.4379679669218604e-08,
-                    0.8896821141242981,
-                    0.4565805196762085,
-                    1.4049705266952515
-                ],
-                [
-                    0,
-                    0,
-                    0,
-                    1
-                ]
-            ]
-        },
-        {
-            "file_path": "015.png",
-            "camera_angle_x": 1.1922232332059943,
-            "proj_type": 0,
-            "azimuth": 8.144422692852467,
-            "elevation": 0.5376699116109869,
-            "cam_dis": 1.5425377326560552,
-            "transform_matrix": [
-                [
-                    -0.9581177234649658,
-                    0.1466628760099411,
-                    -0.24596858024597168,
-                    -0.3794158399105072
-                ],
-                [
-                    -0.2863748371601105,
-                    -0.4906865656375885,
-                    0.8229314684867859,
-                    1.2694027423858643
-                ],
-                [
-                    1.9456553701502344e-08,
-                    0.8589044213294983,
-                    0.5121359825134277,
-                    0.789989173412323
-                ],
-                [
-                    0,
-                    0,
-                    0,
-                    1
-                ]
-            ]
-        },
-        {
-            "file_path": "016.png",
-            "camera_angle_x": 0.4843348255851337,
-            "proj_type": 0,
-            "azimuth": 2.4502860082209676,
-            "elevation": 0.6036991169616068,
-            "cam_dis": 3.6113379310693747,
-            "transform_matrix": [
-                [
-                    -0.6375443935394287,
-                    0.4373573660850525,
-                    -0.6342363953590393,
-                    -2.2904417514801025
-                ],
-                [
-                    -0.7704136967658997,
-                    -0.3619285523891449,
-                    0.5248528718948364,
-                    1.895421028137207
-                ],
-                [
-                    4.0876805940115446e-08,
-                    0.8232413530349731,
-                    0.5676915049552917,
-                    2.050126314163208
-                ],
-                [
-                    0,
-                    0,
-                    0,
-                    1
-                ]
-            ]
-        },
-        {
-            "file_path": "017.png",
-            "camera_angle_x": 1.0484863420149597,
-            "proj_type": 0,
-            "azimuth": 5.59187866181076,
-            "elevation": 0.6728881287032298,
-            "cam_dis": 1.7301201355596483,
-            "transform_matrix": [
-                [
-                    0.6375443935394287,
-                    -0.4801580607891083,
-                    0.6024826765060425,
-                    1.0423673391342163
-                ],
-                [
-                    0.7704136371612549,
-                    0.39734768867492676,
-                    -0.4985755980014801,
-                    -0.8625956773757935
-                ],
-                [
-                    3.110944390982695e-09,
-                    0.7820249199867249,
-                    0.6232471466064453,
-                    1.0782924890518188
-                ],
-                [
-                    0,
-                    0,
-                    0,
-                    1
-                ]
-            ]
-        },
-        {
-            "file_path": "018.png",
-            "camera_angle_x": 0.8327046868500112,
-            "proj_type": 0,
-            "azimuth": 4.021082335015864,
-            "elevation": 0.7461309541186378,
-            "cam_dis": 2.1413633661978952,
-            "transform_matrix": [
-                [
-                    0.7704136371612549,
-                    0.432766854763031,
-                    -0.46816205978393555,
-                    -1.0025050640106201
-                ],
-                [
-                    -0.6375443935394287,
-                    0.5229588747024536,
-                    -0.5657307505607605,
-                    -1.2114349603652954
-                ],
-                [
-                    2.405637644642411e-08,
-                    0.7343207597732544,
-                    0.6788026690483093,
-                    1.4535633325576782
-                ],
-                [
-                    0,
-                    0,
-                    0,
-                    1
-                ]
-            ]
-        },
-        {
-            "file_path": "019.png",
-            "camera_angle_x": 1.0378875875393918,
-            "proj_type": 0,
-            "azimuth": 7.162674988605656,
-            "elevation": 0.8247207741120075,
-            "cam_dis": 1.7461482839234206,
-            "transform_matrix": [
-                [
-                    -0.7704136371612549,
-                    -0.46818602085113525,
-                    0.4327409267425537,
-                    0.75562983751297
-                ],
-                [
-                    0.6375443935394287,
-                    -0.5657596588134766,
-                    0.5229275226593018,
-                    0.9131090044975281
-                ],
-                [
-                    5.7037539136217674e-09,
-                    0.6787620186805725,
-                    0.7343583106994629,
-                    1.2822984457015991
-                ],
-                [
-                    0,
-                    0,
-                    0,
-                    1
-                ]
-            ]
-        },
-        {
-            "file_path": "020.png",
-            "camera_angle_x": 1.181698605034973,
-            "proj_type": 0,
-            "azimuth": 3.235684171618416,
-            "elevation": 0.9106684775229277,
-            "cam_dis": 1.5546175155141897,
-            "transform_matrix": [
-                [
-                    0.09395270049571991,
-                    0.7864198088645935,
-                    -0.6105054020881653,
-                    -0.9491023421287537
-                ],
-                [
-                    -0.9955766797065735,
-                    0.07421454787254333,
-                    -0.057613469660282135,
-                    -0.08956695348024368
-                ],
-                [
-                    1.1830872281848315e-08,
-                    0.613217830657959,
-                    0.7899138331413269,
-                    1.2280138731002808
-                ],
-                [
-                    0,
-                    0,
-                    0,
-                    1
-                ]
-            ]
-        },
-        {
-            "file_path": "021.png",
-            "camera_angle_x": 1.197224930669693,
-            "proj_type": 0,
-            "azimuth": 6.377276825208208,
-            "elevation": 1.0074435322145288,
-            "cam_dis": 1.536877375373375,
-            "transform_matrix": [
-                [
-                    -0.09395267069339752,
-                    -0.8417297005653381,
-                    0.5316617488861084,
-                    0.81709885597229
-                ],
-                [
-                    0.9955767393112183,
-                    -0.0794341042637825,
-                    0.05017295107245445,
-                    0.07710976153612137
-                ],
-                [
-                    -3.2885168366192374e-08,
-                    0.5340239405632019,
-                    0.8454694151878357,
-                    1.2993828058242798
-                ],
-                [
-                    0,
-                    0,
-                    0,
-                    1
-                ]
-            ]
-        },
-        {
-            "file_path": "022.png",
-            "camera_angle_x": 0.7684080505969957,
-            "proj_type": 0,
-            "azimuth": 4.806480498413312,
-            "elevation": 1.122126648780772,
-            "cam_dis": 2.310502047946327,
-            "transform_matrix": [
-                [
-                    0.9955766797065735,
-                    -0.08465376496315002,
-                    0.040753625333309174,
-                    0.09416133165359497
-                ],
-                [
-                    0.0939527377486229,
-                    0.8970394134521484,
-                    -0.43184858560562134,
-                    -0.9977869987487793
-                ],
-                [
-                    -3.0048674659610697e-09,
-                    0.43376728892326355,
-                    0.9010249376296997,
-                    2.081820011138916
-                ],
-                [
-                    0,
-                    0,
-                    0,
-                    1
-                ]
-            ]
-        },
-        {
-            "file_path": "023.png",
-            "camera_angle_x": 1.2027037528892426,
-            "proj_type": 0,
-            "azimuth": 7.948073152003104,
-            "elevation": 1.2750349593459749,
-            "cam_dis": 1.5307354901339616,
-            "transform_matrix": [
-                [
-                    -0.9955767393112183,
-                    0.08987335115671158,
-                    -0.027384238317608833,
-                    -0.04191803187131882
-                ],
-                [
-                    -0.0939527302980423,
-                    -0.9523493051528931,
-                    0.29017895460128784,
-                    0.4441872537136078
-                ],
-                [
-                    -4.757865657012417e-09,
-                    0.29146820306777954,
-                    0.9565805196762085,
-                    1.4642717838287354
-                ],
-                [
-                    0,
-                    0,
-                    0,
-                    1
-                ]
-            ]
-        }
-    ]
-}

hy3dpaint/train_examples/001/render_tex/transforms.json

@@ -1,226 +0,0 @@
-{
-    "aabb": [
-        [
-            -0.5,
-            -0.5,
-            -0.5
-        ],
-        [
-            0.5,
-            0.5,
-            0.5
-        ]
-    ],
-    "scale": 0.5121457915021708,
-    "offset": [
-        0.00032189488410949707,
-        -0.03408946841955185,
-        -0.2757369577884674
-    ],
-    "frames": [
-        {
-            "file_path": "000.png",
-            "camera_angle_x": 0.9232195630055808,
-            "proj_type": 1,
-            "azimuth": -1.5707963267948966,
-            "elevation": 0,
-            "cam_dis": 1.5,
-            "transform_matrix": [
-                [
-                    1.0,
-                    0.0,
-                    6.123233601181349e-17,
-                    9.184850732644269e-17
-                ],
-                [
-                    6.123233601181349e-17,
-                    3.422854177870249e-08,
-                    -0.9999999403953552,
-                    -1.5
-                ],
-                [
-                    0.0,
-                    0.9999999403953552,
-                    3.422854177870249e-08,
-                    0.0
-                ],
-                [
-                    0,
-                    0,
-                    0,
-                    1
-                ]
-            ]
-        },
-        {
-            "file_path": "001.png",
-            "camera_angle_x": 0.9232195630055808,
-            "proj_type": 1,
-            "azimuth": 0,
-            "elevation": 0,
-            "cam_dis": 1.5,
-            "transform_matrix": [
-                [
-                    -2.220446049250313e-16,
-                    0.0,
-                    1.0,
-                    1.5
-                ],
-                [
-                    1.0,
-                    -2.220446049250313e-16,
-                    0.0,
-                    0.0
-                ],
-                [
-                    0.0,
-                    1.0,
-                    -2.220446049250313e-16,
-                    0.0
-                ],
-                [
-                    0,
-                    0,
-                    0,
-                    1
-                ]
-            ]
-        },
-        {
-            "file_path": "002.png",
-            "camera_angle_x": 0.9232195630055808,
-            "proj_type": 1,
-            "azimuth": 1.5707963267948966,
-            "elevation": 0,
-            "cam_dis": 1.5,
-            "transform_matrix": [
-                [
-                    -0.9999999403953552,
-                    0.0,
-                    6.123233601181349e-17,
-                    9.184850732644269e-17
-                ],
-                [
-                    6.123233601181349e-17,
-                    3.422854177870249e-08,
-                    0.9999999403953552,
-                    1.5
-                ],
-                [
-                    0.0,
-                    0.9999999403953552,
-                    3.422854177870249e-08,
-                    0.0
-                ],
-                [
-                    0,
-                    0,
-                    0,
-                    1
-                ]
-            ]
-        },
-        {
-            "file_path": "003.png",
-            "camera_angle_x": 0.9232195630055808,
-            "proj_type": 1,
-            "azimuth": 3.141592653589793,
-            "elevation": 0,
-            "cam_dis": 1.5,
-            "transform_matrix": [
-                [
-                    -2.220446049250313e-16,
-                    0.0,
-                    -1.0,
-                    -1.5
-                ],
-                [
-                    -1.0,
-                    -2.220446049250313e-16,
-                    0.0,
-                    1.8369701465288538e-16
-                ],
-                [
-                    0.0,
-                    1.0,
-                    -2.220446049250313e-16,
-                    0.0
-                ],
-                [
-                    0,
-                    0,
-                    0,
-                    1
-                ]
-            ]
-        },
-        {
-            "file_path": "004.png",
-            "camera_angle_x": 0.9232195630055808,
-            "proj_type": 1,
-            "azimuth": -1.5707963267948966,
-            "elevation": 1.5707963267948966,
-            "cam_dis": 1.5,
-            "transform_matrix": [
-                [
-                    1.0,
-                    -6.123233601181349e-17,
-                    2.812049544789739e-33,
-                    5.624099089579478e-33
-                ],
-                [
-                    6.123233601181349e-17,
-                    1.0,
-                    -3.0616171314629196e-17,
-                    -9.184850732644269e-17
-                ],
-                [
-                    -9.373498482632464e-34,
-                    3.0616171314629196e-17,
-                    1.0,
-                    1.5
-                ],
-                [
-                    0,
-                    0,
-                    0,
-                    1
-                ]
-            ]
-        },
-        {
-            "file_path": "005.png",
-            "camera_angle_x": 0.9232195630055808,
-            "proj_type": 1,
-            "azimuth": -1.5707963267948966,
-            "elevation": -1.5707963267948966,
-            "cam_dis": 1.5,
-            "transform_matrix": [
-                [
-                    1.0,
-                    6.123233601181349e-17,
-                    2.812049544789739e-33,
-                    5.624099089579478e-33
-                ],
-                [
-                    6.123233601181349e-17,
-                    -1.0,
-                    -3.0616171314629196e-17,
-                    -9.184850732644269e-17
-                ],
-                [
-                    9.373498482632464e-34,
-                    3.0616171314629196e-17,
-                    -1.0,
-                    -1.5
-                ],
-                [
-                    0,
-                    0,
-                    0,
-                    1
-                ]
-            ]
-        }
-    ]
-}

hy3dpaint/train_examples/examples.json

@@ -1,3 +0,0 @@
-[
-    "hy3dpaint/train_examples/001"
-]

hy3dpaint/utils/__init__.py

@@ -1,13 +0,0 @@
-# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT
-# except for the third-party components listed below.
-# Hunyuan 3D does not impose any additional limitations beyond what is outlined
-# in the repsective licenses of these third-party components.
-# Users must comply with all terms and conditions of original licenses of these third-party
-# components and must ensure that the usage of the third party components adheres to
-# all relevant laws and regulations.
-
-# For avoidance of doubts, Hunyuan 3D means the large language models and
-# their software and algorithms, including trained model weights, parameters (including
-# optimizer states), machine-learning model code, inference-enabling code, training-enabling code,
-# fine-tuning enabling code and other elements of the foregoing made publicly available
-# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT.

hy3dpaint/utils/image_super_utils.py

@@ -1,41 +0,0 @@
-# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT
-# except for the third-party components listed below.
-# Hunyuan 3D does not impose any additional limitations beyond what is outlined
-# in the repsective licenses of these third-party components.
-# Users must comply with all terms and conditions of original licenses of these third-party
-# components and must ensure that the usage of the third party components adheres to
-# all relevant laws and regulations.
-
-# For avoidance of doubts, Hunyuan 3D means the large language models and
-# their software and algorithms, including trained model weights, parameters (including
-# optimizer states), machine-learning model code, inference-enabling code, training-enabling code,
-# fine-tuning enabling code and other elements of the foregoing made publicly available
-# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT.
-
-import numpy as np
-from PIL import Image
-
-
-class imageSuperNet:
-    def __init__(self, config) -> None:
-        from realesrgan import RealESRGANer
-        from basicsr.archs.rrdbnet_arch import RRDBNet
-
-        model = RRDBNet(num_in_ch=3, num_out_ch=3, num_feat=64, num_block=23, num_grow_ch=32, scale=4)
-        upsampler = RealESRGANer(
-            scale=4,
-            model_path=config.realesrgan_ckpt_path,
-            dni_weight=None,
-            model=model,
-            tile=0,
-            tile_pad=10,
-            pre_pad=0,
-            half=True,
-            gpu_id=None,
-        )
-        self.upsampler = upsampler
-
-    def __call__(self, image):
-        output, _ = self.upsampler.enhance(np.array(image))
-        output = Image.fromarray(output)
-        return output

hy3dpaint/utils/multiview_utils.py

@@ -1,128 +0,0 @@
-# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT
-# except for the third-party components listed below.
-# Hunyuan 3D does not impose any additional limitations beyond what is outlined
-# in the repsective licenses of these third-party components.
-# Users must comply with all terms and conditions of original licenses of these third-party
-# components and must ensure that the usage of the third party components adheres to
-# all relevant laws and regulations.
-
-# For avoidance of doubts, Hunyuan 3D means the large language models and
-# their software and algorithms, including trained model weights, parameters (including
-# optimizer states), machine-learning model code, inference-enabling code, training-enabling code,
-# fine-tuning enabling code and other elements of the foregoing made publicly available
-# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT.
-
-import os
-import torch
-import random
-import numpy as np
-from PIL import Image
-from typing import List
-import huggingface_hub
-from omegaconf import OmegaConf
-from diffusers import DiffusionPipeline
-from diffusers import EulerAncestralDiscreteScheduler, DDIMScheduler, UniPCMultistepScheduler
-
-
-class multiviewDiffusionNet:
-    def __init__(self, config) -> None:
-        self.device = config.device
-
-        cfg_path = config.multiview_cfg_path
-        custom_pipeline = config.custom_pipeline
-        cfg = OmegaConf.load(cfg_path)
-        self.cfg = cfg
-        self.mode = self.cfg.model.params.stable_diffusion_config.custom_pipeline[2:]
-
-        model_path = huggingface_hub.snapshot_download(
-            repo_id=config.multiview_pretrained_path,
-            allow_patterns=["hunyuan3d-paintpbr-v2-1/*"],
-        )
-
-        model_path = os.path.join(model_path, "hunyuan3d-paintpbr-v2-1")
-        pipeline = DiffusionPipeline.from_pretrained(
-            model_path,
-            custom_pipeline=custom_pipeline, 
-            torch_dtype=torch.float16
-        )
-
-        pipeline.scheduler = UniPCMultistepScheduler.from_config(pipeline.scheduler.config, timestep_spacing="trailing")
-        pipeline.set_progress_bar_config(disable=True)
-        pipeline.eval()
-        setattr(pipeline, "view_size", cfg.model.params.get("view_size", 320))
-        self.pipeline = pipeline.to(self.device)
-
-        if hasattr(self.pipeline.unet, "use_dino") and self.pipeline.unet.use_dino:
-            from hunyuanpaintpbr.unet.modules import Dino_v2
-            self.dino_v2 = Dino_v2(config.dino_ckpt_path).to(torch.float16)
-            self.dino_v2 = self.dino_v2.to(self.device)
-
-    def seed_everything(self, seed):
-        random.seed(seed)
-        np.random.seed(seed)
-        torch.manual_seed(seed)
-        os.environ["PL_GLOBAL_SEED"] = str(seed)
-
-    @torch.no_grad()
-    def __call__(self, images, conditions, prompt=None, custom_view_size=None, resize_input=False):
-        pils = self.forward_one(
-            images, conditions, prompt=prompt, custom_view_size=custom_view_size, resize_input=resize_input
-        )
-        return pils
-
-    def forward_one(self, input_images, control_images, prompt=None, custom_view_size=None, resize_input=False):
-        self.seed_everything(0)
-        custom_view_size = custom_view_size if custom_view_size is not None else self.pipeline.view_size
-        if not isinstance(input_images, List):
-            input_images = [input_images]
-        if not resize_input:
-            input_images = [
-                input_image.resize((self.pipeline.view_size, self.pipeline.view_size)) for input_image in input_images
-            ]
-        else:
-            input_images = [input_image.resize((custom_view_size, custom_view_size)) for input_image in input_images]
-        for i in range(len(control_images)):
-            control_images[i] = control_images[i].resize((custom_view_size, custom_view_size))
-            if control_images[i].mode == "L":
-                control_images[i] = control_images[i].point(lambda x: 255 if x > 1 else 0, mode="1")
-        kwargs = dict(generator=torch.Generator(device=self.pipeline.device).manual_seed(0))
-
-        num_view = len(control_images) // 2
-        normal_image = [[control_images[i] for i in range(num_view)]]
-        position_image = [[control_images[i + num_view] for i in range(num_view)]]
-
-        kwargs["width"] = custom_view_size
-        kwargs["height"] = custom_view_size
-        kwargs["num_in_batch"] = num_view
-        kwargs["images_normal"] = normal_image
-        kwargs["images_position"] = position_image
-
-        if hasattr(self.pipeline.unet, "use_dino") and self.pipeline.unet.use_dino:
-            dino_hidden_states = self.dino_v2(input_images[0])
-            kwargs["dino_hidden_states"] = dino_hidden_states
-
-        sync_condition = None
-
-        infer_steps_dict = {
-            "EulerAncestralDiscreteScheduler": 30,
-            "UniPCMultistepScheduler": 15,
-            "DDIMScheduler": 50,
-            "ShiftSNRScheduler": 15,
-        }
-
-        mvd_image = self.pipeline(
-            input_images[0:1],
-            num_inference_steps=infer_steps_dict[self.pipeline.scheduler.__class__.__name__],
-            prompt=prompt,
-            sync_condition=sync_condition,
-            guidance_scale=3.0,
-            **kwargs,
-        ).images
-
-        if "pbr" in self.mode:
-            mvd_image = {"albedo": mvd_image[:num_view], "mr": mvd_image[num_view:]}
-            # mvd_image = {'albedo':mvd_image[:num_view]}
-        else:
-            mvd_image = {"hdr": mvd_image}
-
-        return mvd_image

hy3dpaint/utils/pipeline_utils.py

@@ -1,135 +0,0 @@
-# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT
-# except for the third-party components listed below.
-# Hunyuan 3D does not impose any additional limitations beyond what is outlined
-# in the repsective licenses of these third-party components.
-# Users must comply with all terms and conditions of original licenses of these third-party
-# components and must ensure that the usage of the third party components adheres to
-# all relevant laws and regulations.
-
-# For avoidance of doubts, Hunyuan 3D means the large language models and
-# their software and algorithms, including trained model weights, parameters (including
-# optimizer states), machine-learning model code, inference-enabling code, training-enabling code,
-# fine-tuning enabling code and other elements of the foregoing made publicly available
-# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT.
-
-import torch
-import numpy as np
-
-
-class ViewProcessor:
-    def __init__(self, config, render):
-        self.config = config
-        self.render = render
-
-    def render_normal_multiview(self, camera_elevs, camera_azims, use_abs_coor=True):
-        normal_maps = []
-        for elev, azim in zip(camera_elevs, camera_azims):
-            normal_map = self.render.render_normal(elev, azim, use_abs_coor=use_abs_coor, return_type="pl")
-            normal_maps.append(normal_map)
-
-        return normal_maps
-
-    def render_position_multiview(self, camera_elevs, camera_azims):
-        position_maps = []
-        for elev, azim in zip(camera_elevs, camera_azims):
-            position_map = self.render.render_position(elev, azim, return_type="pl")
-            position_maps.append(position_map)
-
-        return position_maps
-
-    def bake_view_selection(
-        self, candidate_camera_elevs, candidate_camera_azims, candidate_view_weights, max_selected_view_num
-    ):
-
-        original_resolution = self.render.default_resolution
-        self.render.set_default_render_resolution(1024)
-
-        selected_camera_elevs = []
-        selected_camera_azims = []
-        selected_view_weights = []
-        selected_alpha_maps = []
-        viewed_tri_idxs = []
-        viewed_masks = []
-
-        # 计算每个三角片的面积
-        face_areas = self.render.get_face_areas(from_one_index=True)
-        total_area = face_areas.sum()
-        face_area_ratios = face_areas / total_area
-
-        candidate_view_num = len(candidate_camera_elevs)
-        self.render.set_boundary_unreliable_scale(2)
-
-        for elev, azim in zip(candidate_camera_elevs, candidate_camera_azims):
-            viewed_tri_idx = self.render.render_alpha(elev, azim, return_type="np")
-            viewed_tri_idxs.append(set(np.unique(viewed_tri_idx.flatten())))
-            viewed_masks.append(viewed_tri_idx[0, :, :, 0] > 0)
-
-        is_selected = [False for _ in range(candidate_view_num)]
-        total_viewed_tri_idxs = set()
-        total_viewed_area = 0.0
-
-        for idx in range(6):
-            selected_camera_elevs.append(candidate_camera_elevs[idx])
-            selected_camera_azims.append(candidate_camera_azims[idx])
-            selected_view_weights.append(candidate_view_weights[idx])
-            selected_alpha_maps.append(viewed_masks[idx])
-            is_selected[idx] = True
-            total_viewed_tri_idxs.update(viewed_tri_idxs[idx])
-
-        total_viewed_area = face_area_ratios[list(total_viewed_tri_idxs)].sum()
-        for iter in range(max_selected_view_num - len(selected_view_weights)):
-            max_inc = 0
-            max_idx = -1
-
-            for idx, (elev, azim, weight) in enumerate(
-                zip(candidate_camera_elevs, candidate_camera_azims, candidate_view_weights)
-            ):
-                if is_selected[idx]:
-                    continue
-                new_tri_idxs = viewed_tri_idxs[idx] - total_viewed_tri_idxs
-                new_inc_area = face_area_ratios[list(new_tri_idxs)].sum()
-
-                if new_inc_area > max_inc:
-                    max_inc = new_inc_area
-                    max_idx = idx
-
-            if max_inc > 0.01:
-                is_selected[max_idx] = True
-                selected_camera_elevs.append(candidate_camera_elevs[max_idx])
-                selected_camera_azims.append(candidate_camera_azims[max_idx])
-                selected_view_weights.append(candidate_view_weights[max_idx])
-                selected_alpha_maps.append(viewed_masks[max_idx])
-                total_viewed_tri_idxs = total_viewed_tri_idxs.union(viewed_tri_idxs[max_idx])
-                total_viewed_area += max_inc
-            else:
-                break
-
-        self.render.set_default_render_resolution(original_resolution)
-
-        return selected_camera_elevs, selected_camera_azims, selected_view_weights
-
-    def bake_from_multiview(self, views, camera_elevs, camera_azims, view_weights):
-        project_textures, project_weighted_cos_maps = [], []
-        project_boundary_maps = []
-
-        for view, camera_elev, camera_azim, weight in zip(views, camera_elevs, camera_azims, view_weights):
-            project_texture, project_cos_map, project_boundary_map = self.render.back_project(
-                view, camera_elev, camera_azim
-            )
-            project_cos_map = weight * (project_cos_map**self.config.bake_exp)
-            project_textures.append(project_texture)
-            project_weighted_cos_maps.append(project_cos_map)
-            project_boundary_maps.append(project_boundary_map)
-            texture, ori_trust_map = self.render.fast_bake_texture(project_textures, project_weighted_cos_maps)
-        return texture, ori_trust_map > 1e-8
-
-    def texture_inpaint(self, texture, mask, defualt=None):
-        if defualt is not None:
-            mask = mask.astype(bool)
-            inpaint_value = torch.tensor(defualt, dtype=texture.dtype, device=texture.device)
-            texture[~mask] = inpaint_value
-        else:
-            texture_np = self.render.uv_inpaint(texture, mask)
-            texture = torch.tensor(texture_np / 255).float().to(texture.device)
-
-        return texture

hy3dpaint/utils/simplify_mesh_utils.py

@@ -1,37 +0,0 @@
-# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT
-# except for the third-party components listed below.
-# Hunyuan 3D does not impose any additional limitations beyond what is outlined
-# in the repsective licenses of these third-party components.
-# Users must comply with all terms and conditions of original licenses of these third-party
-# components and must ensure that the usage of the third party components adheres to
-# all relevant laws and regulations.
-
-# For avoidance of doubts, Hunyuan 3D means the large language models and
-# their software and algorithms, including trained model weights, parameters (including
-# optimizer states), machine-learning model code, inference-enabling code, training-enabling code,
-# fine-tuning enabling code and other elements of the foregoing made publicly available
-# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT.
-
-import trimesh
-import pymeshlab
-
-
-def remesh_mesh(mesh_path, remesh_path):
-    mesh = mesh_simplify_trimesh(mesh_path, remesh_path)
-
-
-def mesh_simplify_trimesh(inputpath, outputpath, target_count=40000):
-    # 先去除离散面
-    ms = pymeshlab.MeshSet()
-    if inputpath.endswith(".glb"):
-        ms.load_new_mesh(inputpath, load_in_a_single_layer=True)
-    else:
-        ms.load_new_mesh(inputpath)
-    ms.save_current_mesh(outputpath.replace(".glb", ".obj"), save_textures=False)
-    # 调用减面函数
-    courent = trimesh.load(outputpath.replace(".glb", ".obj"), force="mesh")
-    face_num = courent.faces.shape[0]
-
-    if face_num > target_count:
-        courent = courent.simplify_quadric_decimation(target_count)
-    courent.export(outputpath)

hy3dpaint/utils/torchvision_fix.py

@@ -1,111 +0,0 @@
-# Torchvision compatibility fix for functional_tensor module
-# This file helps resolve compatibility issues between different torchvision versions
-
-import sys
-import torch
-import torchvision
-
-def fix_torchvision_functional_tensor():
-    """
-    Fix torchvision.transforms.functional_tensor import issue
-    """
-    try:
-        # Check if the module exists in the expected location
-        import torchvision.transforms.functional_tensor
-        print("torchvision.transforms.functional_tensor is available")
-        return True
-    except ImportError:
-        print("torchvision.transforms.functional_tensor not found, applying compatibility fix...")
-        
-        try:
-            # Create a mock functional_tensor module with the required functions
-            import torchvision.transforms.functional as F
-            
-            class FunctionalTensorMock:
-                """Mock module to replace functional_tensor"""
-                
-                @staticmethod
-                def _get_grayscale_weights(img):
-                    """Helper to create grayscale weights based on image dimensions"""
-                    weights = torch.tensor([0.299, 0.587, 0.114], device=img.device, dtype=img.dtype)
-                    return weights.view(1, 3, 1, 1) if len(img.shape) == 4 else weights.view(3, 1, 1)
-                
-                @staticmethod
-                def _try_import_fallback(module_names, attr_name):
-                    """Helper to try importing from multiple modules"""
-                    for module_name in module_names:
-                        try:
-                            module = __import__(module_name, fromlist=[attr_name])
-                            if hasattr(module, attr_name):
-                                return getattr(module, attr_name)
-                        except ImportError:
-                            continue
-                    return None
-                
-                @staticmethod
-                def rgb_to_grayscale(img, num_output_channels=1):
-                    """Convert RGB image to grayscale"""
-                    if hasattr(F, 'rgb_to_grayscale'):
-                        return F.rgb_to_grayscale(img, num_output_channels)
-                    
-                    # Fallback implementation
-                    weights = FunctionalTensorMock._get_grayscale_weights(img)
-                    grayscale = torch.sum(img * weights, dim=-3, keepdim=True)
-                    
-                    if num_output_channels == 3:
-                        repeat_dims = (1, 3, 1, 1) if len(img.shape) == 4 else (3, 1, 1)
-                        grayscale = grayscale.repeat(*repeat_dims)
-                    
-                    return grayscale
-                
-                @staticmethod
-                def resize(img, size, interpolation=2, antialias=None):
-                    """Resize function wrapper"""
-                    # Try v2.functional first, then regular functional, then torch.nn.functional
-                    resize_func = FunctionalTensorMock._try_import_fallback([
-                        'torchvision.transforms.v2.functional',
-                        'torchvision.transforms.functional'
-                    ], 'resize')
-                    
-                    if resize_func:
-                        try:
-                            return resize_func(img, size, interpolation=interpolation, antialias=antialias)
-                        except TypeError:
-                            # Fallback for older versions without antialias parameter
-                            return resize_func(img, size, interpolation=interpolation)
-                    
-                    # Final fallback using torch.nn.functional
-                    import torch.nn.functional as torch_F
-                    size = (size, size) if isinstance(size, int) else size
-                    img_input = img.unsqueeze(0) if len(img.shape) == 3 else img
-                    return torch_F.interpolate(img_input, size=size, mode='bilinear', align_corners=False)
-                
-                def __getattr__(self, name):
-                    """Fallback to regular functional module"""
-                    func = self._try_import_fallback([
-                        'torchvision.transforms.functional',
-                        'torchvision.transforms.v2.functional'
-                    ], name)
-                    
-                    if func:
-                        return func
-                    
-                    raise AttributeError(f"'{name}' not found in functional_tensor mock")
-            
-            # Create the mock module instance and monkey patch
-            sys.modules['torchvision.transforms.functional_tensor'] = FunctionalTensorMock()
-            print("Applied compatibility fix: created functional_tensor mock module")
-            return True
-            
-        except Exception as e:
-            print(f"Failed to create functional_tensor mock: {e}")
-            return False
-
-def apply_fix():
-    """Apply the torchvision compatibility fix"""
-    print(f"Torchvision version: {torchvision.__version__}")
-    return fix_torchvision_functional_tensor()
-
-if __name__ == "__main__":
-    apply_fix() 
-    

hy3dpaint/utils/uvwrap_utils.py

@@ -1,32 +0,0 @@
-# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT
-# except for the third-party components listed below.
-# Hunyuan 3D does not impose any additional limitations beyond what is outlined
-# in the repsective licenses of these third-party components.
-# Users must comply with all terms and conditions of original licenses of these third-party
-# components and must ensure that the usage of the third party components adheres to
-# all relevant laws and regulations.
-
-# For avoidance of doubts, Hunyuan 3D means the large language models and
-# their software and algorithms, including trained model weights, parameters (including
-# optimizer states), machine-learning model code, inference-enabling code, training-enabling code,
-# fine-tuning enabling code and other elements of the foregoing made publicly available
-# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT.
-
-import trimesh
-import xatlas
-
-
-def mesh_uv_wrap(mesh):
-    if isinstance(mesh, trimesh.Scene):
-        mesh = mesh.dump(concatenate=True)
-
-    if len(mesh.faces) > 500000000:
-        raise ValueError("The mesh has more than 500,000,000 faces, which is not supported.")
-
-    vmapping, indices, uvs = xatlas.parametrize(mesh.vertices, mesh.faces)
-
-    mesh.vertices = mesh.vertices[vmapping]
-    mesh.faces = indices
-    mesh.visual.uv = uvs
-
-    return mesh

hy3dshape/.gitignore

@@ -1,169 +0,0 @@
-# Byte-compiled / optimized / DLL files
-__pycache__/
-*.py[cod]
-*$py.class
-
-# C extensions
-*.so
-
-# Distribution / packaging
-.Python
-build/
-develop-eggs/
-dist/
-downloads/
-eggs/
-.eggs/
-lib/
-!hy3dgen/texgen/custom_rasterizer/lib/
-lib64/
-parts/
-sdist/
-var/
-wheels/
-share/python-wheels/
-*.egg-info/
-.installed.cfg
-*.egg
-MANIFEST
-
-# PyInstaller
-#  Usually these files are written by a python script from a template
-#  before PyInstaller builds the exe, so as to inject date/other infos into it.
-*.manifest
-*.spec
-
-# Installer logs
-pip-log.txt
-pip-delete-this-directory.txt
-
-# Unit test / coverage reports
-htmlcov/
-.tox/
-.nox/
-.coverage
-.coverage.*
-.cache
-nosetests.xml
-coverage.xml
-*.cover
-*.py,cover
-.hypothesis/
-.pytest_cache/
-cover/
-
-# Translations
-*.mo
-*.pot
-
-# Django stuff:
-*.log
-local_settings.py
-db.sqlite3
-db.sqlite3-journal
-
-# Flask stuff:
-instance/
-.webassets-cache
-
-# Scrapy stuff:
-.scrapy
-
-# Sphinx documentation
-docs/_build/
-
-# PyBuilder
-.pybuilder/
-target/
-
-# Jupyter Notebook
-.ipynb_checkpoints
-
-# IPython
-profile_default/
-ipython_config.py
-
-# pyenv
-#   For a library or package, you might want to ignore these files since the code is
-#   intended to run in multiple environments; otherwise, check them in:
-# .python-version
-
-# pipenv
-#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
-#   However, in case of collaboration, if having platform-specific dependencies or dependencies
-#   having no cross-platform support, pipenv may install dependencies that don't work, or not
-#   install all needed dependencies.
-#Pipfile.lock
-
-# UV
-#   Similar to Pipfile.lock, it is generally recommended to include uv.lock in version control.
-#   This is especially recommended for binary packages to ensure reproducibility, and is more
-#   commonly ignored for libraries.
-#uv.lock
-
-# poetry
-#   Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control.
-#   This is especially recommended for binary packages to ensure reproducibility, and is more
-#   commonly ignored for libraries.
-#   https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control
-#poetry.lock
-
-# pdm
-#   Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control.
-#pdm.lock
-#   pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it
-#   in version control.
-#   https://pdm.fming.dev/latest/usage/project/#working-with-version-control
-.pdm.toml
-.pdm-python
-.pdm-build/
-
-# PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm
-__pypackages__/
-
-# Celery stuff
-celerybeat-schedule
-celerybeat.pid
-
-# SageMath parsed files
-*.sage.py
-
-# Environments
-.env
-.venv
-env/
-venv/
-ENV/
-env.bak/
-venv.bak/
-
-# Spyder project settings
-.spyderproject
-.spyproject
-
-# Rope project settings
-.ropeproject
-
-# mkdocs documentation
-/site
-
-# mypy
-.mypy_cache/
-.dmypy.json
-dmypy.json
-
-# Pyre type checker
-.pyre/
-
-# pytype static type analyzer
-.pytype/
-.DS_Store
-# Cython debug symbols
-cython_debug/
-gradio_cache/
-# PyCharm
-#  JetBrains specific template is maintained in a separate JetBrains.gitignore that can
-#  be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore
-#  and can be added to the global gitignore or merged into this file.  For a more nuclear
-#  option (not recommended) you can uncomment the following to ignore the entire idea folder.
-#.idea/

hy3dshape/LICENSE

@@ -1,81 +0,0 @@
-TENCENT HUNYUAN 3D 2.1 COMMUNITY LICENSE AGREEMENT
-Tencent Hunyuan 3D 2.1 Release Date: June 13, 2025
-THIS LICENSE AGREEMENT DOES NOT APPLY IN THE EUROPEAN UNION, UNITED KINGDOM AND SOUTH KOREA AND IS EXPRESSLY LIMITED TO THE TERRITORY, AS DEFINED BELOW.
-By clicking to agree or by using, reproducing, modifying, distributing, performing or displaying any portion or element of the Tencent Hunyuan 3D 2.1 Works, including via any Hosted Service, You will be deemed to have recognized and accepted the content of this Agreement, which is effective immediately.
-1.	DEFINITIONS.
-a.	“Acceptable Use Policy” shall mean the policy made available by Tencent as set forth in the Exhibit A.
-b.	“Agreement” shall mean the terms and conditions for use, reproduction, distribution, modification, performance and displaying of Tencent Hunyuan 3D 2.1 Works or any portion or element thereof set forth herein.
-c.	“Documentation” shall mean the specifications, manuals and documentation for Tencent Hunyuan 3D 2.1 made publicly available by Tencent.
-d.	“Hosted Service” shall mean a hosted service offered via an application programming interface (API), web access, or any other electronic or remote means.
-e.	“Licensee,” “You” or “Your” shall mean a natural person or legal entity exercising the rights granted by this Agreement and/or using the Tencent Hunyuan 3D 2.1 Works for any purpose and in any field of use.
-f.	“Materials” shall mean, collectively, Tencent’s proprietary Tencent Hunyuan 3D 2.1 and Documentation (and any portion thereof) as made available by Tencent under this Agreement.
-g.	“Model Derivatives” shall mean all: (i) modifications to Tencent Hunyuan 3D 2.1 or any Model Derivative of Tencent Hunyuan 3D 2.1; (ii) works based on Tencent Hunyuan 3D 2.1 or any Model Derivative of Tencent Hunyuan 3D 2.1; or (iii) any other machine learning model which is created by transfer of patterns of the weights, parameters, operations, or Output of Tencent Hunyuan 3D 2.1 or any Model Derivative of Tencent Hunyuan 3D 2.1, to that model in order to cause that model to perform similarly to Tencent Hunyuan 3D 2.1 or a Model Derivative of Tencent Hunyuan 3D 2.1, including distillation methods, methods that use intermediate data representations, or methods based on the generation of synthetic data Outputs by Tencent Hunyuan 3D 2.1 or a Model Derivative of Tencent Hunyuan 3D 2.1 for training that model. For clarity, Outputs by themselves are not deemed Model Derivatives.
-h.	“Output” shall mean the information and/or content output of Tencent Hunyuan 3D 2.1 or a Model Derivative that results from operating or otherwise using Tencent Hunyuan 3D 2.1 or a Model Derivative, including via a Hosted Service.
-i.	“Tencent,” “We” or “Us” shall mean THL Q Limited.
-j.	“Tencent Hunyuan 3D 2.1” shall mean the 3D generation models and their software and algorithms, including trained model weights, parameters (including optimizer states), machine-learning model code, inference-enabling code, training-enabling code, fine-tuning enabling code and other elements of the foregoing made publicly available by Us at [ https://github.com/Tencent-Hunyuan/Hunyuan3D-2.1].
-k.	“Tencent Hunyuan 3D 2.1 Works” shall mean: (i) the Materials; (ii) Model Derivatives; and (iii) all derivative works thereof.
-l.	“Territory” shall mean the worldwide territory, excluding the territory of the European Union, United Kingdom and South Korea. 
-m.	“Third Party” or “Third Parties” shall mean individuals or legal entities that are not under common control with Us or You.
-n.	“including” shall mean including but not limited to.
-2.	GRANT OF RIGHTS.
-We grant You, for the Territory only, a non-exclusive, non-transferable and royalty-free limited license under Tencent’s intellectual property or other rights owned by Us embodied in or utilized by the Materials to use, reproduce, distribute, create derivative works of (including Model Derivatives), and make modifications to the Materials, only in accordance with the terms of this Agreement and the Acceptable Use Policy, and You must not violate (or encourage or permit anyone else to violate) any term of this Agreement or the Acceptable Use Policy.
-3.	DISTRIBUTION.
-You may, subject to Your compliance with this Agreement, distribute or make available to Third Parties the Tencent Hunyuan 3D 2.1 Works, exclusively in the Territory, provided that You meet all of the following conditions:
-a.	You must provide all such Third Party recipients of the Tencent Hunyuan 3D 2.1 Works or products or services using them a copy of this Agreement;
-b.	You must cause any modified files to carry prominent notices stating that You changed the files;
-c.	You are encouraged to: (i) publish at least one technology introduction blogpost or one public statement expressing Your experience of using the Tencent Hunyuan 3D 2.1 Works; and (ii) mark the products or services developed by using the Tencent Hunyuan 3D 2.1 Works to indicate that the product/service is “Powered by Tencent Hunyuan”; and
-d.	All distributions to Third Parties (other than through a Hosted Service) must be accompanied by a “Notice” text file that contains the following notice: “Tencent Hunyuan 3D 2.1 is licensed under the Tencent Hunyuan 3D 2.1 Community License Agreement, Copyright © 2025 Tencent. All Rights Reserved. The trademark rights of “Tencent Hunyuan” are owned by Tencent or its affiliate.”
-You may add Your own copyright statement to Your modifications and, except as set forth in this Section and in Section 5, may provide additional or different license terms and conditions for use, reproduction, or distribution of Your modifications, or for any such Model Derivatives as a whole, provided Your use, reproduction, modification, distribution, performance and display of the work otherwise complies with the terms and conditions of this Agreement (including as regards the Territory). If You receive Tencent Hunyuan 3D 2.1 Works from a Licensee as part of an integrated end user product, then this Section 3 of this Agreement will not apply to You.
-4.	ADDITIONAL COMMERCIAL TERMS.
-If, on the Tencent Hunyuan 3D 2.1 version release date, the monthly active users of all products or services made available by or for Licensee is greater than 1 million monthly active users in the preceding calendar month, You must request a license from Tencent, which Tencent may grant to You in its sole discretion, and You are not authorized to exercise any of the rights under this Agreement unless or until Tencent otherwise expressly grants You such rights. 
-Subject to Tencent's written approval, you may request a license for the use of Tencent Hunyuan 3D 2.1 by submitting the following information to [email protected]:
-a.	Your company’s name and associated business sector that plans to use Tencent Hunyuan 3D 2.1.
-b.	Your intended use case and the purpose of using Tencent Hunyuan 3D 2.1.
-c.	Your plans to modify Tencent Hunyuan 3D 2.1 or create Model Derivatives.
-5.	RULES OF USE.
-a.	Your use of the Tencent Hunyuan 3D 2.1 Works must comply with applicable laws and regulations (including trade compliance laws and regulations) and adhere to the Acceptable Use Policy for the Tencent Hunyuan 3D 2.1 Works, which is hereby incorporated by reference into this Agreement. You must include the use restrictions referenced in these Sections 5(a) and 5(b) as an enforceable provision in any agreement (e.g., license agreement, terms of use, etc.) governing the use and/or distribution of Tencent Hunyuan 3D 2.1 Works and You must provide notice to subsequent users to whom You distribute that Tencent Hunyuan 3D 2.1 Works are subject to the use restrictions in these Sections 5(a) and 5(b).
-b.	You must not use the Tencent Hunyuan 3D 2.1 Works or any Output or results of the Tencent Hunyuan 3D 2.1 Works to improve any other AI model (other than Tencent Hunyuan 3D 2.1 or Model Derivatives thereof).
-c.	You must not use, reproduce, modify, distribute, or display the Tencent Hunyuan 3D 2.1 Works, Output or results of the Tencent Hunyuan 3D 2.1 Works outside the Territory. Any such use outside the Territory is unlicensed and unauthorized under this Agreement.
-6.	INTELLECTUAL PROPERTY.
-a.	Subject to Tencent’s ownership of Tencent Hunyuan 3D 2.1 Works made by or for Tencent and intellectual property rights therein, conditioned upon Your compliance with the terms and conditions of this Agreement, as between You and Tencent, You will be the owner of any derivative works and modifications of the Materials and any Model Derivatives that are made by or for You.
-b.	No trademark licenses are granted under this Agreement, and in connection with the Tencent Hunyuan 3D 2.1 Works, Licensee may not use any name or mark owned by or associated with Tencent or any of its affiliates, except as required for reasonable and customary use in describing and distributing the Tencent Hunyuan 3D 2.1 Works. Tencent hereby grants You a license to use “Tencent Hunyuan” (the “Mark”) in the Territory solely as required to comply with the provisions of Section 3(c), provided that You comply with any applicable laws related to trademark protection. All goodwill arising out of Your use of the Mark will inure to the benefit of Tencent.
-c.	If You commence a lawsuit or other proceedings (including a cross-claim or counterclaim in a lawsuit) against Us or any person or entity alleging that the Materials or any Output, or any portion of any of the foregoing, infringe any intellectual property or other right owned or licensable by You, then all licenses granted to You under this Agreement shall terminate as of the date such lawsuit or other proceeding is filed. You will defend, indemnify and hold harmless Us from and against any claim by any Third Party arising out of or related to Your or the Third Party’s use or distribution of the Tencent Hunyuan 3D 2.1 Works.
-d.	Tencent claims no rights in Outputs You generate. You and Your users are solely responsible for Outputs and their subsequent uses.
-7.	DISCLAIMERS OF WARRANTY AND LIMITATIONS OF LIABILITY.
-a.	We are not obligated to support, update, provide training for, or develop any further version of the Tencent Hunyuan 3D 2.1 Works or to grant any license thereto.
-b.	UNLESS AND ONLY TO THE EXTENT REQUIRED BY APPLICABLE LAW, THE TENCENT HUNYUAN 3D 2.1 WORKS AND ANY OUTPUT AND RESULTS THEREFROM ARE PROVIDED “AS IS” WITHOUT ANY EXPRESS OR IMPLIED WARRANTIES OF ANY KIND INCLUDING ANY WARRANTIES OF TITLE, MERCHANTABILITY, NONINFRINGEMENT, COURSE OF DEALING, USAGE OF TRADE, OR FITNESS FOR A PARTICULAR PURPOSE. YOU ARE SOLELY RESPONSIBLE FOR DETERMINING THE APPROPRIATENESS OF USING, REPRODUCING, MODIFYING, PERFORMING, DISPLAYING OR DISTRIBUTING ANY OF THE TENCENT HUNYUAN 3D 2.1 WORKS OR OUTPUTS AND ASSUME ANY AND ALL RISKS ASSOCIATED WITH YOUR OR A THIRD PARTY’S USE OR DISTRIBUTION OF ANY OF THE TENCENT HUNYUAN 3D 2.1 WORKS OR OUTPUTS AND YOUR EXERCISE OF RIGHTS AND PERMISSIONS UNDER THIS AGREEMENT.
-c.	TO THE FULLEST EXTENT PERMITTED BY APPLICABLE LAW, IN NO EVENT SHALL TENCENT OR ITS AFFILIATES BE LIABLE UNDER ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, TORT, NEGLIGENCE, PRODUCTS LIABILITY, OR OTHERWISE, FOR ANY DAMAGES, INCLUDING ANY DIRECT, INDIRECT, SPECIAL, INCIDENTAL, EXEMPLARY, CONSEQUENTIAL OR PUNITIVE DAMAGES, OR LOST PROFITS OF ANY KIND ARISING FROM THIS AGREEMENT OR RELATED TO ANY OF THE TENCENT HUNYUAN 3D 2.1 WORKS OR OUTPUTS, EVEN IF TENCENT OR ITS AFFILIATES HAVE BEEN ADVISED OF THE POSSIBILITY OF ANY OF THE FOREGOING.
-8.	SURVIVAL AND TERMINATION.
-a.	The term of this Agreement shall commence upon Your acceptance of this Agreement or access to the Materials and will continue in full force and effect until terminated in accordance with the terms and conditions herein.
-b.	We may terminate this Agreement if You breach any of the terms or conditions of this Agreement. Upon termination of this Agreement, You must promptly delete and cease use of the Tencent Hunyuan 3D 2.1 Works. Sections 6(a), 6(c), 7 and 9 shall survive the termination of this Agreement.
-9.	GOVERNING LAW AND JURISDICTION.
-a.	This Agreement and any dispute arising out of or relating to it will be governed by the laws of the Hong Kong Special Administrative Region of the People’s Republic of China, without regard to conflict of law principles, and the UN Convention on Contracts for the International Sale of Goods does not apply to this Agreement.
-b.	Exclusive jurisdiction and venue for any dispute arising out of or relating to this Agreement will be a court of competent jurisdiction in the Hong Kong Special Administrative Region of the People’s Republic of China, and Tencent and Licensee consent to the exclusive jurisdiction of such court with respect to any such dispute.
- 
-EXHIBIT A
-ACCEPTABLE USE POLICY
-
-Tencent reserves the right to update this Acceptable Use Policy from time to time.
-Last modified: November 5, 2024
-
-Tencent endeavors to promote safe and fair use of its tools and features, including Tencent Hunyuan 3D 2.1. You agree not to use Tencent Hunyuan 3D 2.1 or Model Derivatives:
-1.	Outside the Territory;
-2.	In any way that violates any applicable national, federal, state, local, international or any other law or regulation;
-3.	To harm Yourself or others;
-4.	To repurpose or distribute output from Tencent Hunyuan 3D 2.1 or any Model Derivatives to harm Yourself or others; 
-5.	To override or circumvent the safety guardrails and safeguards We have put in place;
-6.	For the purpose of exploiting, harming or attempting to exploit or harm minors in any way;
-7.	To generate or disseminate verifiably false information and/or content with the purpose of harming others or influencing elections;
-8.	To generate or facilitate false online engagement, including fake reviews and other means of fake online engagement;
-9.	To intentionally defame, disparage or otherwise harass others;
-10.	To generate and/or disseminate malware (including ransomware) or any other content to be used for the purpose of harming electronic systems;
-11.	To generate or disseminate personal identifiable information with the purpose of harming others;
-12.	To generate or disseminate information (including images, code, posts, articles), and place the information in any public context (including –through the use of bot generated tweets), without expressly and conspicuously identifying that the information and/or content is machine generated;
-13.	To impersonate another individual without consent, authorization, or legal right;
-14.	To make high-stakes automated decisions in domains that affect an individual’s safety, rights or wellbeing (e.g., law enforcement, migration, medicine/health, management of critical infrastructure, safety components of products, essential services, credit, employment, housing, education, social scoring, or insurance);
-15.	In a manner that violates or disrespects the social ethics and moral standards of other countries or regions;
-16.	To perform, facilitate, threaten, incite, plan, promote or encourage violent extremism or terrorism;
-17.	For any use intended to discriminate against or harm individuals or groups based on protected characteristics or categories, online or offline social behavior or known or predicted personal or personality characteristics;
-18.	To intentionally exploit any of the vulnerabilities of a specific group of persons based on their age, social, physical or mental characteristics, in order to materially distort the behavior of a person pertaining to that group in a manner that causes or is likely to cause that person or another person physical or psychological harm;
-19.	For military purposes;
-20.	To engage in the unauthorized or unlicensed practice of any profession including, but not limited to, financial, legal, medical/health, or other professional practices.

hy3dshape/NOTICE

@@ -1,214 +0,0 @@
-Usage and Legal Notices:
-
-Tencent is pleased to support the open source community by making Hunyuan 3D 2.0 available.
-
-Copyright (C) 2025 THL A29 Limited, a Tencent company.  All rights reserved. The below software and/or models in this distribution may have been modified by THL A29 Limited ("Tencent Modifications"). All Tencent Modifications are Copyright (C) THL A29 Limited.
-
-Hunyuan 3D 2.0 is licensed under the TENCENT HUNYUAN 3D 2.0 COMMUNITY LICENSE AGREEMENT except for the third-party components listed below, which is licensed under different terms. Hunyuan 3D 2.0 does not impose any additional limitations beyond what is outlined in the respective licenses of these third-party components. Users must comply with all terms and conditions of original licenses of these third-party components and must ensure that the usage of the third party components adheres to all relevant laws and regulations. 
-
-For avoidance of doubts, Hunyuan 3D 2.0 means inference-enabling code, parameters, and weights of this Model only, which are made publicly available by Tencent in accordance with TENCENT HUNYUAN 3D 2.0 COMMUNITY LICENSE AGREEMENT.
-
-
-Other dependencies and licenses:
-
-
-Open Source Model Licensed under the MIT and CreativeML Open RAIL++-M License:
---------------------------------------------------------------------
-1. Stable Diffusion
-Copyright (c) 2022 Stability AI
-
-
-Terms of the MIT and CreativeML Open RAIL++-M License:
---------------------------------------------------------------------
-Permission is hereby granted, free of charge, to any person obtaining a copy
-of this software and associated documentation files (the "Software"), to deal
-in the Software without restriction, including without limitation the rights
-to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
-copies of the Software, and to permit persons to whom the Software is
-furnished to do so, subject to the following conditions:
-
-The above copyright notice and this permission notice shall be included in all
-copies or substantial portions of the Software.
-
-THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
-IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
-FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
-AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
-LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
-OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
-SOFTWARE.
-
-
-CreativeML Open RAIL++-M License
-dated November 24, 2022
-
-Section I: PREAMBLE
-
-Multimodal generative models are being widely adopted and used, and have the potential to transform the way artists, among other individuals, conceive and benefit from AI or ML technologies as a tool for content creation.
-
-Notwithstanding the current and potential benefits that these artifacts can bring to society at large, there are also concerns about potential misuses of them, either due to their technical limitations or ethical considerations.
-
-In short, this license strives for both the open and responsible downstream use of the accompanying model. When it comes to the open character, we took inspiration from open source permissive licenses regarding the grant of IP rights. Referring to the downstream responsible use, we added use-based restrictions not permitting the use of the Model in very specific scenarios, in order for the licensor to be able to enforce the license in case potential misuses of the Model may occur. At the same time, we strive to promote open and responsible research on generative models for art and content generation.
-
-Even though downstream derivative versions of the model could be released under different licensing terms, the latter will always have to include - at minimum - the same use-based restrictions as the ones in the original license (this license). We believe in the intersection between open and responsible AI development; thus, this License aims to strike a balance between both in order to enable responsible open-science in the field of AI.
-
-This License governs the use of the model (and its derivatives) and is informed by the model card associated with the model.
-
-NOW THEREFORE, You and Licensor agree as follows:
-
-1. Definitions
-
-- "License" means the terms and conditions for use, reproduction, and Distribution as defined in this document.
-- "Data" means a collection of information and/or content extracted from the dataset used with the Model, including to train, pretrain, or otherwise evaluate the Model. The Data is not licensed under this License.
-- "Output" means the results of operating a Model as embodied in informational content resulting therefrom.
-- "Model" means any accompanying machine-learning based assemblies (including checkpoints), consisting of learnt weights, parameters (including optimizer states), corresponding to the model architecture as embodied in the Complementary Material, that have been trained or tuned, in whole or in part on the Data, using the Complementary Material.
-- "Derivatives of the Model" means all modifications to the Model, works based on the Model, or any other model which is created or initialized by transfer of patterns of the weights, parameters, activations or output of the Model, to the other model, in order to cause the other model to perform similarly to the Model, including - but not limited to - distillation methods entailing the use of intermediate data representations or methods based on the generation of synthetic data by the Model for training the other model.
-- "Complementary Material" means the accompanying source code and scripts used to define, run, load, benchmark or evaluate the Model, and used to prepare data for training or evaluation, if any. This includes any accompanying documentation, tutorials, examples, etc, if any.
-- "Distribution" means any transmission, reproduction, publication or other sharing of the Model or Derivatives of the Model to a third party, including providing the Model as a hosted service made available by electronic or other remote means - e.g. API-based or web access.
-- "Licensor" means the copyright owner or entity authorized by the copyright owner that is granting the License, including the persons or entities that may have rights in the Model and/or distributing the Model.
-- "You" (or "Your") means an individual or Legal Entity exercising permissions granted by this License and/or making use of the Model for whichever purpose and in any field of use, including usage of the Model in an end-use application - e.g. chatbot, translator, image generator.
-- "Third Parties" means individuals or legal entities that are not under common control with Licensor or You.
-- "Contribution" means any work of authorship, including the original version of the Model and any modifications or additions to that Model or Derivatives of the Model thereof, that is intentionally submitted to Licensor for inclusion in the Model by the copyright owner or by an individual or Legal Entity authorized to submit on behalf of the copyright owner. For the purposes of this definition, "submitted" means any form of electronic, verbal, or written communication sent to the Licensor or its representatives, including but not limited to communication on electronic mailing lists, source code control systems, and issue tracking systems that are managed by, or on behalf of, the Licensor for the purpose of discussing and improving the Model, but excluding communication that is conspicuously marked or otherwise designated in writing by the copyright owner as "Not a Contribution."
-- "Contributor" means Licensor and any individual or Legal Entity on behalf of whom a Contribution has been received by Licensor and subsequently incorporated within the Model.
-
-Section II: INTELLECTUAL PROPERTY RIGHTS
-
-Both copyright and patent grants apply to the Model, Derivatives of the Model and Complementary Material. The Model and Derivatives of the Model are subject to additional terms as described in Section III.
-
-2. Grant of Copyright License. Subject to the terms and conditions of this License, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license to reproduce, prepare, publicly display, publicly perform, sublicense, and distribute the Complementary Material, the Model, and Derivatives of the Model.
-3. Grant of Patent License. Subject to the terms and conditions of this License and where and as applicable, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable (except as stated in this paragraph) patent license to make, have made, use, offer to sell, sell, import, and otherwise transfer the Model and the Complementary Material, where such license applies only to those patent claims licensable by such Contributor that are necessarily infringed by their Contribution(s) alone or by combination of their Contribution(s) with the Model to which such Contribution(s) was submitted. If You institute patent litigation against any entity (including a cross-claim or counterclaim in a lawsuit) alleging that the Model and/or Complementary Material or a Contribution incorporated within the Model and/or Complementary Material constitutes direct or contributory patent infringement, then any patent licenses granted to You under this License for the Model and/or Work shall terminate as of the date such litigation is asserted or filed.
-
-Section III: CONDITIONS OF USAGE, DISTRIBUTION AND REDISTRIBUTION
-
-4. Distribution and Redistribution. You may host for Third Party remote access purposes (e.g. software-as-a-service), reproduce and distribute copies of the Model or Derivatives of the Model thereof in any medium, with or without modifications, provided that You meet the following conditions:
-Use-based restrictions as referenced in paragraph 5 MUST be included as an enforceable provision by You in any type of legal agreement (e.g. a license) governing the use and/or distribution of the Model or Derivatives of the Model, and You shall give notice to subsequent users You Distribute to, that the Model or Derivatives of the Model are subject to paragraph 5. This provision does not apply to the use of Complementary Material.
-You must give any Third Party recipients of the Model or Derivatives of the Model a copy of this License;
-You must cause any modified files to carry prominent notices stating that You changed the files;
-You must retain all copyright, patent, trademark, and attribution notices excluding those notices that do not pertain to any part of the Model, Derivatives of the Model.
-You may add Your own copyright statement to Your modifications and may provide additional or different license terms and conditions - respecting paragraph 4.a. - for use, reproduction, or Distribution of Your modifications, or for any such Derivatives of the Model as a whole, provided Your use, reproduction, and Distribution of the Model otherwise complies with the conditions stated in this License.
-5. Use-based restrictions. The restrictions set forth in Attachment A are considered Use-based restrictions. Therefore You cannot use the Model and the Derivatives of the Model for the specified restricted uses. You may use the Model subject to this License, including only for lawful purposes and in accordance with the License. Use may include creating any content with, finetuning, updating, running, training, evaluating and/or reparametrizing the Model. You shall require all of Your users who use the Model or a Derivative of the Model to comply with the terms of this paragraph (paragraph 5).
-6. The Output You Generate. Except as set forth herein, Licensor claims no rights in the Output You generate using the Model. You are accountable for the Output you generate and its subsequent uses. No use of the output can contravene any provision as stated in the License.
-
-Section IV: OTHER PROVISIONS
-
-7. Updates and Runtime Restrictions. To the maximum extent permitted by law, Licensor reserves the right to restrict (remotely or otherwise) usage of the Model in violation of this License.
-8. Trademarks and related. Nothing in this License permits You to make use of Licensors’ trademarks, trade names, logos or to otherwise suggest endorsement or misrepresent the relationship between the parties; and any rights not expressly granted herein are reserved by the Licensors.
-9. Disclaimer of Warranty. Unless required by applicable law or agreed to in writing, Licensor provides the Model and the Complementary Material (and each Contributor provides its Contributions) on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied, including, without limitation, any warranties or conditions of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A PARTICULAR PURPOSE. You are solely responsible for determining the appropriateness of using or redistributing the Model, Derivatives of the Model, and the Complementary Material and assume any risks associated with Your exercise of permissions under this License.
-10. Limitation of Liability. In no event and under no legal theory, whether in tort (including negligence), contract, or otherwise, unless required by applicable law (such as deliberate and grossly negligent acts) or agreed to in writing, shall any Contributor be liable to You for damages, including any direct, indirect, special, incidental, or consequential damages of any character arising as a result of this License or out of the use or inability to use the Model and the Complementary Material (including but not limited to damages for loss of goodwill, work stoppage, computer failure or malfunction, or any and all other commercial damages or losses), even if such Contributor has been advised of the possibility of such damages.
-11. Accepting Warranty or Additional Liability. While redistributing the Model, Derivatives of the Model and the Complementary Material thereof, You may choose to offer, and charge a fee for, acceptance of support, warranty, indemnity, or other liability obligations and/or rights consistent with this License. However, in accepting such obligations, You may act only on Your own behalf and on Your sole responsibility, not on behalf of any other Contributor, and only if You agree to indemnify, defend, and hold each Contributor harmless for any liability incurred by, or claims asserted against, such Contributor by reason of your accepting any such warranty or additional liability.
-12. If any provision of this License is held to be invalid, illegal or unenforceable, the remaining provisions shall be unaffected thereby and remain valid as if such provision had not been set forth herein.
-
-END OF TERMS AND CONDITIONS
-
-
-
-
-Attachment A
-
-Use Restrictions
-
-You agree not to use the Model or Derivatives of the Model:
-
-- In any way that violates any applicable national, federal, state, local or international law or regulation;
-- For the purpose of exploiting, harming or attempting to exploit or harm minors in any way;
-- To generate or disseminate verifiably false information and/or content with the purpose of harming others;
-- To generate or disseminate personal identifiable information that can be used to harm an individual;
-- To defame, disparage or otherwise harass others;
-- For fully automated decision making that adversely impacts an individual’s legal rights or otherwise creates or modifies a binding, enforceable obligation;
-- For any use intended to or which has the effect of discriminating against or harming individuals or groups based on online or offline social behavior or known or predicted personal or personality characteristics;
-- To exploit any of the vulnerabilities of a specific group of persons based on their age, social, physical or mental characteristics, in order to materially distort the behavior of a person pertaining to that group in a manner that causes or is likely to cause that person or another person physical or psychological harm;
-- For any use intended to or which has the effect of discriminating against individuals or groups based on legally protected characteristics or categories;
-- To provide medical advice and medical results interpretation;
-- To generate or disseminate information for the purpose to be used for administration of justice, law enforcement, immigration or asylum processes, such as predicting an individual will commit fraud/crime commitment (e.g. by text profiling, drawing causal relationships between assertions made in documents, indiscriminate and arbitrarily-targeted use).
-
-
-
-Open Source Model Licensed under the TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT and Other Licenses of the Third-Party Components therein:
---------------------------------------------------------------------
-1. HunyuanDiT
-Copyright (C) 2024 THL A29 Limited, a Tencent company.  All rights reserved. 
-
-
-Terms of the TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT:
---------------------------------------------------------------------
-TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT
-Tencent Hunyuan Release Date: 2024/5/14
-By clicking to agree or by using, reproducing, modifying, distributing, performing or displaying any portion or element of the Tencent Hunyuan Works, including via any Hosted Service, You will be deemed to have recognized and accepted the content of this Agreement, which is effective immediately.
-1.	DEFINITIONS.
-a.	“Acceptable Use Policy” shall mean the policy made available by Tencent as set forth in the Exhibit A.
-b.	“Agreement” shall mean the terms and conditions for use, reproduction, distribution, modification, performance and displaying of the Hunyuan Works or any portion or element thereof set forth herein.
-c.	“Documentation” shall mean the specifications, manuals and documentation for Tencent Hunyuan made publicly available by Tencent.
-d.	“Hosted Service” shall mean a hosted service offered via an application programming interface (API), web access, or any other electronic or remote means.
-e.	“Licensee,” “You” or “Your” shall mean a natural person or legal entity exercising the rights granted by this Agreement and/or using the Tencent Hunyuan Works for any purpose and in any field of use.
-f.	“Materials” shall mean, collectively, Tencent’s proprietary Tencent Hunyuan and Documentation (and any portion thereof) as made available by Tencent under this Agreement.
-g.	“Model Derivatives” shall mean all: (i) modifications to Tencent Hunyuan or any Model Derivative of Tencent Hunyuan; (ii) works based on Tencent Hunyuan or any Model Derivative of Tencent Hunyuan; or (iii) any other machine learning model which is created by transfer of patterns of the weights, parameters, operations, or Output of Tencent Hunyuan or any Model Derivative of Tencent Hunyuan, to that model in order to cause that model to perform similarly to Tencent Hunyuan or a Model Derivative of Tencent Hunyuan, including distillation methods, methods that use intermediate data representations, or methods based on the generation of synthetic data Outputs by Tencent Hunyuan or a Model Derivative of Tencent Hunyuan for training that model. For clarity, Outputs by themselves are not deemed Model Derivatives.
-h.	“Output” shall mean the information and/or content output of Tencent Hunyuan or a Model Derivative that results from operating or otherwise using Tencent Hunyuan or a Model Derivative, including via a Hosted Service.
-i.	“Tencent,” “We” or “Us” shall mean THL A29 Limited.
-j.	“Tencent Hunyuan” shall mean the large language models, image/video/audio/3D generation models, and multimodal large language models and their software and algorithms, including trained model weights, parameters (including optimizer states), machine-learning model code, inference-enabling code, training-enabling code, fine-tuning enabling code and other elements of the foregoing made publicly available by Us at  https://huggingface.co/Tencent-Hunyuan/HunyuanDiT and https://github.com/Tencent/HunyuanDiT .
-k.	“Tencent Hunyuan Works” shall mean: (i) the Materials; (ii) Model Derivatives; and (iii) all derivative works thereof.
-l.	“Third Party” or “Third Parties” shall mean individuals or legal entities that are not under common control with Us or You.
-m.	“including” shall mean including but not limited to.
-2.	GRANT OF RIGHTS.
-We grant You a non-exclusive, worldwide, non-transferable and royalty-free limited license under Tencent’s intellectual property or other rights owned by Us embodied in or utilized by the Materials to use, reproduce, distribute, create derivative works of (including Model Derivatives), and make modifications to the Materials, only in accordance with the terms of this Agreement and the Acceptable Use Policy, and You must not violate (or encourage or permit anyone else to violate) any term of this Agreement or the Acceptable Use Policy.
-3.	DISTRIBUTION.
-You may, subject to Your compliance with this Agreement, distribute or make available to Third Parties the Tencent Hunyuan Works, provided that You meet all of the following conditions:
-a.	You must provide all such Third Party recipients of the Tencent Hunyuan Works or products or services using them a copy of this Agreement;
-b.	You must cause any modified files to carry prominent notices stating that You changed the files;
-c.	You are encouraged to: (i) publish at least one technology introduction blogpost or one public statement expressing Your experience of using the Tencent Hunyuan Works; and (ii) mark the products or services developed by using the Tencent Hunyuan Works to indicate that the product/service is “Powered by Tencent Hunyuan”; and
-d.	All distributions to Third Parties (other than through a Hosted Service) must be accompanied by a “Notice” text file that contains the following notice: “Tencent Hunyuan is licensed under the Tencent Hunyuan Community License Agreement, Copyright © 2024 Tencent. All Rights Reserved. The trademark rights of “Tencent Hunyuan” are owned by Tencent or its affiliate.”
-You may add Your own copyright statement to Your modifications and, except as set forth in this Section and in Section 5, may provide additional or different license terms and conditions for use, reproduction, or distribution of Your modifications, or for any such Model Derivatives as a whole, provided Your use, reproduction, modification, distribution, performance and display of the work otherwise complies with the terms and conditions of this Agreement. If You receive Tencent Hunyuan Works from a Licensee as part of an integrated end user product, then this Section 3 of this Agreement will not apply to You.
-4.	ADDITIONAL COMMERCIAL TERMS.
-If, on the Tencent Hunyuan version release date, the monthly active users of all products or services made available by or for Licensee is greater than 100 million monthly active users in the preceding calendar month, You must request a license from Tencent, which Tencent may grant to You in its sole discretion, and You are not authorized to exercise any of the rights under this Agreement unless or until Tencent otherwise expressly grants You such rights.
-5.	RULES OF USE.
-a.	Your use of the Tencent Hunyuan Works must comply with applicable laws and regulations (including trade compliance laws and regulations) and adhere to the Acceptable Use Policy for the Tencent Hunyuan Works, which is hereby incorporated by reference into this Agreement. You must include the use restrictions referenced in these Sections 5(a) and 5(b) as an enforceable provision in any agreement (e.g., license agreement, terms of use, etc.) governing the use and/or distribution of Tencent Hunyuan Works and You must provide notice to subsequent users to whom You distribute that Tencent Hunyuan Works are subject to the use restrictions in these Sections 5(a) and 5(b).
-b.	You must not use the Tencent Hunyuan Works or any Output or results of the Tencent Hunyuan Works to improve any other large language model (other than Tencent Hunyuan or Model Derivatives thereof).
-6.	INTELLECTUAL PROPERTY.
-a.	Subject to Tencent’s ownership of Tencent Hunyuan Works made by or for Tencent and intellectual property rights therein, conditioned upon Your compliance with the terms and conditions of this Agreement, as between You and Tencent, You will be the owner of any derivative works and modifications of the Materials and any Model Derivatives that are made by or for You.
-b.	No trademark licenses are granted under this Agreement, and in connection with the Tencent Hunyuan Works, Licensee may not use any name or mark owned by or associated with Tencent or any of its affiliates, except as required for reasonable and customary use in describing and distributing the Tencent Hunyuan Works. Tencent hereby grants You a license to use “Tencent Hunyuan” (the “Mark”) solely as required to comply with the provisions of Section 3(c), provided that You comply with any applicable laws related to trademark protection. All goodwill arising out of Your use of the Mark will inure to the benefit of Tencent.
-c.	If You commence a lawsuit or other proceedings (including a cross-claim or counterclaim in a lawsuit) against Us or any person or entity alleging that the Materials or any Output, or any portion of any of the foregoing, infringe any intellectual property or other right owned or licensable by You, then all licenses granted to You under this Agreement shall terminate as of the date such lawsuit or other proceeding is filed. You will defend, indemnify and hold harmless Us from and against any claim by any Third Party arising out of or related to Your or the Third Party’s use or distribution of the Tencent Hunyuan Works.
-d.	Tencent claims no rights in Outputs You generate. You and Your users are solely responsible for Outputs and their subsequent uses.
-7.	DISCLAIMERS OF WARRANTY AND LIMITATIONS OF LIABILITY.
-a.	We are not obligated to support, update, provide training for, or develop any further version of the Tencent Hunyuan Works or to grant any license thereto.
-b.	UNLESS AND ONLY TO THE EXTENT REQUIRED BY APPLICABLE LAW, THE TENCENT HUNYUAN WORKS AND ANY OUTPUT AND RESULTS THEREFROM ARE PROVIDED “AS IS” WITHOUT ANY EXPRESS OR IMPLIED WARRANTIES OF ANY KIND INCLUDING ANY WARRANTIES OF TITLE, MERCHANTABILITY, NONINFRINGEMENT, COURSE OF DEALING, USAGE OF TRADE, OR FITNESS FOR A PARTICULAR PURPOSE. YOU ARE SOLELY RESPONSIBLE FOR DETERMINING THE APPROPRIATENESS OF USING, REPRODUCING, MODIFYING, PERFORMING, DISPLAYING OR DISTRIBUTING ANY OF THE TENCENT HUNYUAN WORKS OR OUTPUTS AND ASSUME ANY AND ALL RISKS ASSOCIATED WITH YOUR OR A THIRD PARTY’S USE OR DISTRIBUTION OF ANY OF THE TENCENT HUNYUAN WORKS OR OUTPUTS AND YOUR EXERCISE OF RIGHTS AND PERMISSIONS UNDER THIS AGREEMENT.
-c.	TO THE FULLEST EXTENT PERMITTED BY APPLICABLE LAW, IN NO EVENT SHALL TENCENT OR ITS AFFILIATES BE LIABLE UNDER ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, TORT, NEGLIGENCE, PRODUCTS LIABILITY, OR OTHERWISE, FOR ANY DAMAGES, INCLUDING ANY DIRECT, INDIRECT, SPECIAL, INCIDENTAL, EXEMPLARY, CONSEQUENTIAL OR PUNITIVE DAMAGES, OR LOST PROFITS OF ANY KIND ARISING FROM THIS AGREEMENT OR RELATED TO ANY OF THE TENCENT HUNYUAN WORKS OR OUTPUTS, EVEN IF TENCENT OR ITS AFFILIATES HAVE BEEN ADVISED OF THE POSSIBILITY OF ANY OF THE FOREGOING.
-8.	SURVIVAL AND TERMINATION.
-a.	The term of this Agreement shall commence upon Your acceptance of this Agreement or access to the Materials and will continue in full force and effect until terminated in accordance with the terms and conditions herein.
-b.	We may terminate this Agreement if You breach any of the terms or conditions of this Agreement. Upon termination of this Agreement, You must promptly delete and cease use of the Tencent Hunyuan Works. Sections 6(a), 6(c), 7 and 9 shall survive the termination of this Agreement.
-9.	GOVERNING LAW AND JURISDICTION.
-a.	This Agreement and any dispute arising out of or relating to it will be governed by the laws of the Hong Kong Special Administrative Region of the People’s Republic of China, without regard to conflict of law principles, and the UN Convention on Contracts for the International Sale of Goods does not apply to this Agreement.
-b.	Exclusive jurisdiction and venue for any dispute arising out of or relating to this Agreement will be a court of competent jurisdiction in the Hong Kong Special Administrative Region of the People’s Republic of China, and Tencent and Licensee consent to the exclusive jurisdiction of such court with respect to any such dispute.
- 
-
-EXHIBIT A
-ACCEPTABLE USE POLICY
-
-Tencent reserves the right to update this Acceptable Use Policy from time to time.
-Last modified: 2024/5/14
-
-Tencent endeavors to promote safe and fair use of its tools and features, including Tencent Hunyuan. You agree not to use Tencent Hunyuan or Model Derivatives:
-1.	In any way that violates any applicable national, federal, state, local, international or any other law or regulation;
-2.	To harm Yourself or others;
-3.	To repurpose or distribute output from Tencent Hunyuan or any Model Derivatives to harm Yourself or others; 
-4.	To override or circumvent the safety guardrails and safeguards We have put in place;
-5.	For the purpose of exploiting, harming or attempting to exploit or harm minors in any way;
-6.	To generate or disseminate verifiably false information and/or content with the purpose of harming others or influencing elections;
-7.	To generate or facilitate false online engagement, including fake reviews and other means of fake online engagement;
-8.	To intentionally defame, disparage or otherwise harass others;
-9.	To generate and/or disseminate malware (including ransomware) or any other content to be used for the purpose of harming electronic systems;
-10.	To generate or disseminate personal identifiable information with the purpose of harming others;
-11.	To generate or disseminate information (including images, code, posts, articles), and place the information in any public context (including –through the use of bot generated tweets), without expressly and conspicuously identifying that the information and/or content is machine generated;
-12.	To impersonate another individual without consent, authorization, or legal right;
-13.	To make high-stakes automated decisions in domains that affect an individual’s safety, rights or wellbeing (e.g., law enforcement, migration, medicine/health, management of critical infrastructure, safety components of products, essential services, credit, employment, housing, education, social scoring, or insurance);
-14.	In a manner that violates or disrespects the social ethics and moral standards of other countries or regions;
-15.	To perform, facilitate, threaten, incite, plan, promote or encourage violent extremism or terrorism;
-16.	For any use intended to discriminate against or harm individuals or groups based on protected characteristics or categories, online or offline social behavior or known or predicted personal or personality characteristics;
-17.	To intentionally exploit any of the vulnerabilities of a specific group of persons based on their age, social, physical or mental characteristics, in order to materially distort the behavior of a person pertaining to that group in a manner that causes or is likely to cause that person or another person physical or psychological harm;
-18.	For military purposes;
-19.	To engage in the unauthorized or unlicensed practice of any profession including, but not limited to, financial, legal, medical/health, or other professional practices.
-
-For the license of other third party components, please refer to the following URL:
-https://huggingface.co/Tencent-Hunyuan/HunyuanDiT/blob/main/Notice
-
---------------------------------------------------------------------
-
-This Model also incorporates insights from Flux's neural network architechtures (https://github.com/black-forest-labs/flux?tab=readme-ov-file). Credits are given to the orginal authors.

hy3dshape/README-zh.md

@@ -1,47 +0,0 @@
-# Hunyuan3D-2.1-Shape
-
-
-# 训练
-
-我们会展示小数据集上DiT的训练全流程
-
-## 数据预处理
-
-渲染和水密化参考[链接](tools/README.md)，最终得到如下结构
-
-``` yaml
-dataset/preprocessed/{uid}
-├── geo_data
-│   ├── {uid}_sdf.npz
-│   ├── {uid}_surface.npz
-│   └── {uid}_watertight.obj
-└── render_cond
-    ├── 000.png
-    ├── ...
-    ├── 023.png
-    ├── mesh.ply
-    └── transforms.json
-```
-
-我们提供了一个8个case(均来自Objaverse-XL)预处理后的结果在 tools/mini_trainset，可以直接用于过拟合训练
-
-
-
-## 启动训练
-
-我们提供了可供参考的训练配置文件和启动脚本（默认单机8卡deepspeed训练），用户根据需要自行修改。
-
-配置文件
-```
-configs/dit-from-scratch-overfitting-flowmatching-dinog518-bf16-lr1e4-1024.yaml
-```
-启动脚本
-
-```
-export node_num=1
-export node_rank=0
-export master_ip=0.0.0.0 # set your master_ip
-export config='configs/dit-from-scratch-overfitting-flowmatching-dinog518-bf16-lr1e4-1024.yaml'
-export output_dir='output_folder/dit/overfitting'
-bash scripts/train_deepspeed.sh $node_num $node_rank $master_ip $config $output_dir
-```

hy3dshape/README.md

@@ -1,54 +0,0 @@
-# Hunyuan3D-2.1-Shape
-
-## Quick Inference
-
-Given a reference image `image.png`, you can run inference using the following code. The result will be saved as `demo.glb`.
-
-```bash
-python3 minimal_demo.py
-```
-
-**Memory Recommendation:** For we recommend using a GPU with at least **10GB VRAM**.
-
-# Training
-
-Here we demonstrate the complete training workflow of DiT on a small dataset.
-
-## Data Preprocessing
-
-The rendering and watertight mesh generation process is described in detail in [this document](tools/README.md). After preprocessing, the dataset directory structure should look like the following:
-
-```yaml
-dataset/preprocessed/{uid}
-├── geo_data
-│   ├── {uid}_sdf.npz
-│   ├── {uid}_surface.npz
-│   └── {uid}_watertight.obj
-└── render_cond
-    ├── 000.png
-    ├── ...
-    ├── 023.png
-    ├── mesh.ply
-    └── transforms.json
-```
-
-We provide a preprocessed mini_dataset containing 8 cases (all sourced from Objaverse-XL) as `tools/mini_trainset`, which can be used directly for DiT overfitting training experiments.
-
-## Launching Training
-
-We provide example configuration files and launch scripts for reference. By default, the training runs on a single node with 8 GPUs using DeepSpeed. Users can modify the configurations and scripts as needed to suit their environment.
-
-Configuration File
-```
-configs/hunyuandit-mini-overfitting-flowmatching-dinog518-bf16-lr1e4-512.yaml
-```
-Launch Script
-
-```
-export node_num=1
-export node_rank=0
-export master_ip=0.0.0.0 # set your master_ip
-export config=configs/hunyuandit-mini-overfitting-flowmatching-dinog518-bf16-lr1e4-512.yaml
-export output_dir=output_folder/dit/overfitting
-bash scripts/train_deepspeed.sh $node_num $node_rank $master_ip $config $output_dir
-```

hy3dshape/configs/hunyuan3ddit-full-params-finetuning-flowmatching-dinog518-bf16-lr1e5-512.yaml

@@ -1,174 +0,0 @@
-name: "DiT: Flux large flowmatching; VAE: 1024 token length; ImageEncoder: DINO Giant; ImageSize: 518"
-
-training:
-  steps: 10_0000_0000
-  use_amp: true
-  amp_type: "bf16"
-  base_lr: 1.e-5
-  gradient_clip_val: 1.0
-  gradient_clip_algorithm: "norm"
-  every_n_train_steps: 2000 # 5000
-  val_check_interval: 50 # 4096
-  limit_val_batches: 16
-
-dataset:
-  target: hy3dshape.data.dit_asl.AlignedShapeLatentModule
-  params:
-    #! Base setting
-    batch_size: 4
-    num_workers: 8
-    val_num_workers: 4
-
-    # Data 
-    train_data_list: tools/mini_trainset/preprocessed
-    val_data_list: tools/mini_trainset/preprocessed
-
-    #! Image loading
-    cond_stage_key: "image" # image / text / image_text
-    image_size: 518
-    mean: &mean [0.5, 0.5, 0.5]
-    std: &std [0.5, 0.5, 0.5]
-
-    #! Point cloud sampling
-    pc_size: &pc_size 30720
-    pc_sharpedge_size: &pc_sharpedge_size 30720
-    sharpedge_label: &sharpedge_label true
-    return_normal: true
-
-    #! Augmentation
-    padding: true
-
-model:
-  target: hy3dshape.models.diffusion.flow_matching_sit.Diffuser
-  params:
-    first_stage_key: "surface"
-    cond_stage_key: "image"
-    scale_by_std: false
-    z_scale_factor: &z_scale_factor 0.9990943042622529 # 1 / 1.0009065167661184
-    torch_compile: false
-
-    # ema_config:
-    #   ema_model: LitEma
-    #   ema_decay: 0.999
-    #   ema_inference: false
-
-    first_stage_config:
-      target: hy3dshape.models.autoencoders.ShapeVAE
-      from_pretrained: tencent/Hunyuan3D-2.1
-      params:
-        num_latents: &num_latents 512
-        embed_dim: 64
-        num_freqs: 8
-        include_pi: false
-        heads: 16
-        width: 1024
-        point_feats: 4
-        num_decoder_layers: 16
-        pc_size: *pc_size
-        pc_sharpedge_size: *pc_sharpedge_size
-        qkv_bias: false
-        qk_norm: true
-        scale_factor: *z_scale_factor
-        geo_decoder_mlp_expand_ratio: 4
-        geo_decoder_downsample_ratio: 1
-        geo_decoder_ln_post: true
-
-    cond_stage_config:
-      target: hy3dshape.models.conditioner.SingleImageEncoder
-      params:
-        main_image_encoder:
-            type: DinoImageEncoder # dino giant
-            kwargs:
-                config:
-                  attention_probs_dropout_prob: 0.0
-                  drop_path_rate: 0.0
-                  hidden_act: gelu
-                  hidden_dropout_prob: 0.0
-                  hidden_size: 1536
-                  image_size: 518
-                  initializer_range: 0.02
-                  layer_norm_eps: 1.e-6
-                  layerscale_value: 1.0
-                  mlp_ratio: 4
-                  model_type: dinov2
-                  num_attention_heads: 24
-                  num_channels: 3
-                  num_hidden_layers: 40
-                  patch_size: 14
-                  qkv_bias: true
-                  torch_dtype: float32
-                  use_swiglu_ffn: true
-                image_size: 518
-
-    denoiser_cfg:
-      target: hy3dshape.models.denoisers.hunyuan3ddit.Hunyuan3DDiT
-      params:
-        ckpt_path: ~/.cache/hy3dgen/tencent/Hunyuan3D-2-1-Shape/dit/model.fp16.ckpt
-        input_size: *num_latents
-        context_in_dim: 1536
-        hidden_size: 1024
-        mlp_ratio: 4.0
-        num_heads: 16
-        depth: 16
-        depth_single_blocks: 32
-        axes_dim: [64]
-        theta: 10000
-        qkv_bias: true
-        use_pe: false
-        force_norm_fp32: true
-
-    scheduler_cfg:
-      transport:
-        target: hy3dshape.models.diffusion.transport.create_transport
-        params:
-          path_type: Linear
-          prediction: velocity
-      sampler:
-        target: hy3dshape.models.diffusion.transport.Sampler
-        params: {}
-        ode_params:
-          sampling_method: euler # dopri5 ...
-          num_steps: &num_steps 50
-
-    optimizer_cfg:
-      optimizer:
-        target: torch.optim.AdamW
-        params:
-          betas: [0.9, 0.99]
-          eps: 1.e-6
-          weight_decay: 1.e-2
-
-      scheduler:
-        target: hy3dshape.utils.trainings.lr_scheduler.LambdaWarmUpCosineFactorScheduler
-        params:
-          warm_up_steps: 50 # 5000
-          f_start: 1.e-6
-          f_min: 1.e-3
-          f_max: 1.0
-
-    pipeline_cfg:
-      target: hy3dshape.pipelines.Hunyuan3DDiTFlowMatchingPipeline
-
-    image_processor_cfg:
-      target: hy3dshape.preprocessors.ImageProcessorV2
-      params: {}
-
-callbacks:
-    logger:
-      target: hy3dshape.utils.trainings.mesh_log_callback.ImageConditionalASLDiffuserLogger
-      params:
-        step_frequency: 100 # 10000
-        num_samples: 1
-        sample_times: 1
-        mean: *mean
-        std: *std
-        bounds: [-1.01, -1.01, -1.01, 1.01, 1.01, 1.01]
-        octree_depth: 8
-        num_chunks: 50000
-        mc_level: 0.0
-    
-    file_loggers:
-        target: hy3dshape.utils.trainings.mesh_log_callback.ImageConditionalFixASLDiffuserLogger
-        params:
-          step_frequency: 50 # 5000
-          test_data_path: "tools/mini_testset/images.json"

hy3dshape/configs/hunyuan3ddit-mini-overfitting-flowmatching-dinog518-bf16-lr1e4-512.yaml

@@ -1,173 +0,0 @@
-name: "DiT: Flux large flowmatching; VAE: 1024 token length; ImageEncoder: DINO Giant; ImageSize: 518"
-
-training:
-  steps: 10_0000_0000
-  use_amp: true
-  amp_type: "bf16"
-  base_lr: 1e-4
-  gradient_clip_val: 1.0
-  gradient_clip_algorithm: "norm"
-  every_n_train_steps: 2000 # 5000
-  val_check_interval: 50 # 4096
-  limit_val_batches: 16
-
-dataset:
-  target: hy3dshape.data.dit_asl.AlignedShapeLatentModule
-  params:
-    #! Base setting
-    batch_size: 2
-    num_workers: 8
-    val_num_workers: 4
-
-    # Data 
-    train_data_list: tools/mini_trainset/preprocessed
-    val_data_list: tools/mini_trainset/preprocessed
-
-    #! Image loading
-    cond_stage_key: "image" # image / text / image_text
-    image_size: 518
-    mean: &mean [0.5, 0.5, 0.5]
-    std: &std [0.5, 0.5, 0.5]
-
-    #! Point cloud sampling
-    pc_size: &pc_size 10240
-    pc_sharpedge_size: &pc_sharpedge_size 10240
-    sharpedge_label: &sharpedge_label true
-    return_normal: true
-
-    #! Augmentation
-    padding: true
-
-model:
-  target: hy3dshape.models.diffusion.flow_matching_sit.Diffuser
-  params:
-    first_stage_key: "surface"
-    cond_stage_key: "image"
-    scale_by_std: false
-    z_scale_factor: &z_scale_factor 0.9990943042622529 # 1 / 1.0009065167661184
-    torch_compile: false
-
-    # ema_config:
-    #   ema_model: LitEma
-    #   ema_decay: 0.999
-    #   ema_inference: false
-
-    first_stage_config:
-      target: hy3dshape.models.autoencoders.ShapeVAE
-      from_pretrained: tencent/Hunyuan3D-2.1
-      params:
-        num_latents: &num_latents 512
-        embed_dim: 64
-        num_freqs: 8
-        include_pi: false
-        heads: 16
-        width: 1024
-        point_feats: 4
-        num_decoder_layers: 16
-        pc_size: *pc_size
-        pc_sharpedge_size: *pc_sharpedge_size
-        qkv_bias: false
-        qk_norm: true
-        scale_factor: *z_scale_factor
-        geo_decoder_mlp_expand_ratio: 4
-        geo_decoder_downsample_ratio: 1
-        geo_decoder_ln_post: true
-
-    cond_stage_config:
-      target: hy3dshape.models.conditioner.SingleImageEncoder
-      params:
-        main_image_encoder:
-            type: DinoImageEncoder # dino giant
-            kwargs:
-                config:
-                  attention_probs_dropout_prob: 0.0
-                  drop_path_rate: 0.0
-                  hidden_act: gelu
-                  hidden_dropout_prob: 0.0
-                  hidden_size: 1536
-                  image_size: 518
-                  initializer_range: 0.02
-                  layer_norm_eps: 1.e-6
-                  layerscale_value: 1.0
-                  mlp_ratio: 4
-                  model_type: dinov2
-                  num_attention_heads: 24
-                  num_channels: 3
-                  num_hidden_layers: 40
-                  patch_size: 14
-                  qkv_bias: true
-                  torch_dtype: float32
-                  use_swiglu_ffn: true
-                image_size: 518
-
-    denoiser_cfg:
-      target: hy3dshape.models.denoisers.hunyuan3ddit.Hunyuan3DDiT
-      params:
-        input_size: *num_latents
-        context_in_dim: 1536
-        hidden_size: 1024
-        mlp_ratio: 4.0
-        num_heads: 16
-        depth: 8
-        depth_single_blocks: 16
-        axes_dim: [64]
-        theta: 10000
-        qkv_bias: true
-        use_pe: false
-        force_norm_fp32: true
-
-    scheduler_cfg:
-      transport:
-        target: hy3dshape.models.diffusion.transport.create_transport
-        params:
-          path_type: Linear
-          prediction: velocity
-      sampler:
-        target: hy3dshape.models.diffusion.transport.Sampler
-        params: {}
-        ode_params:
-          sampling_method: euler # dopri5 ...
-          num_steps: &num_steps 50
-
-    optimizer_cfg:
-      optimizer:
-        target: torch.optim.AdamW
-        params:
-          betas: [0.9, 0.99]
-          eps: 1.e-6
-          weight_decay: 1.e-2
-
-      scheduler:
-        target: hy3dshape.utils.trainings.lr_scheduler.LambdaWarmUpCosineFactorScheduler
-        params:
-          warm_up_steps: 50 # 5000
-          f_start: 1.e-6
-          f_min: 1.e-3
-          f_max: 1.0
-
-    pipeline_cfg:
-      target: hy3dshape.pipelines.Hunyuan3DDiTFlowMatchingPipeline
-
-    image_processor_cfg:
-      target: hy3dshape.preprocessors.ImageProcessorV2
-      params: {}
-
-callbacks:
-    logger:
-      target: hy3dshape.utils.trainings.mesh_log_callback.ImageConditionalASLDiffuserLogger
-      params:
-        step_frequency: 100 # 10000
-        num_samples: 1
-        sample_times: 1
-        mean: *mean
-        std: *std
-        bounds: [-1.01, -1.01, -1.01, 1.01, 1.01, 1.01]
-        octree_depth: 8
-        num_chunks: 50000
-        mc_level: 0.0
-    
-    file_loggers:
-        target: hy3dshape.utils.trainings.mesh_log_callback.ImageConditionalFixASLDiffuserLogger
-        params:
-          step_frequency: 50 # 5000
-          test_data_path: "tools/mini_testset/images.json"

hy3dshape/configs/hunyuandit-finetuning-flowmatching-dinog518-bf16-lr1e5-4096.yaml

@@ -1,180 +0,0 @@
-name: "DiT: Flux large flowmatching; VAE: 1024 token length; ImageEncoder: DINO Giant; ImageSize: 518"
-
-training:
-  steps: 10_0000_0000
-  use_amp: true
-  amp_type: "bf16"
-  base_lr: 1e-5
-  gradient_clip_val: 1.0
-  gradient_clip_algorithm: "norm"
-  every_n_train_steps: 2000 # 5000
-  val_check_interval: 50 # 4096
-  limit_val_batches: 16
-
-dataset:
-  target: hy3dshape.data.dit_asl.AlignedShapeLatentModule
-  params:
-    #! Base setting
-    batch_size: 4
-    num_workers: 8
-    val_num_workers: 4
-
-    # Data 
-    train_data_list: tools/mini_trainset/preprocessed
-    val_data_list: tools/mini_trainset/preprocessed
-
-    #! Image loading
-    cond_stage_key: "image" # image / text / image_text
-    image_size: 518
-    mean: &mean [0.5, 0.5, 0.5]
-    std: &std [0.5, 0.5, 0.5]
-
-    #! Point cloud sampling
-    pc_size: &pc_size 81920
-    pc_sharpedge_size: &pc_sharpedge_size 0
-    sharpedge_label: &sharpedge_label true
-    return_normal: true
-
-    #! Augmentation
-    padding: true
-
-model:
-  target: hy3dshape.models.diffusion.flow_matching_sit.Diffuser
-  params:
-    first_stage_key: "surface"
-    cond_stage_key: "image"
-    scale_by_std: false
-    z_scale_factor: &z_scale_factor 1.0039506158752403
-    torch_compile: false
-
-    # ema_config:
-    #   ema_model: LitEma
-    #   ema_decay: 0.999
-    #   ema_inference: false
-
-    first_stage_config:
-      target: hy3dshape.models.autoencoders.ShapeVAE
-      from_pretrained: tencent/Hunyuan3D-2.1
-      params:
-        num_latents: &num_latents 4096
-        embed_dim: 64
-        num_freqs: 8
-        include_pi: false
-        heads: 16
-        width: 1024
-        num_encoder_layers: 8
-        num_decoder_layers: 16
-        qkv_bias: false
-        qk_norm: true
-        scale_factor: *z_scale_factor
-        geo_decoder_mlp_expand_ratio: 4
-        geo_decoder_downsample_ratio: 1
-        geo_decoder_ln_post: true
-        point_feats: 4
-        pc_size: *pc_size
-        pc_sharpedge_size: *pc_sharpedge_size
-
-    cond_stage_config:
-      target: hy3dshape.models.conditioner.SingleImageEncoder
-      params:
-        main_image_encoder:
-            type: DinoImageEncoder # dino large
-            kwargs:
-                config:
-                  attention_probs_dropout_prob: 0.0
-                  drop_path_rate: 0.0
-                  hidden_act: gelu
-                  hidden_dropout_prob: 0.0
-                  hidden_size: 1024
-                  image_size: 518
-                  initializer_range: 0.02
-                  layer_norm_eps: 1.e-6
-                  layerscale_value: 1.0
-                  mlp_ratio: 4
-                  model_type: dinov2
-                  num_attention_heads: 16
-                  num_channels: 3
-                  num_hidden_layers: 24
-                  patch_size: 14
-                  qkv_bias: true
-                  torch_dtype: float32
-                  use_swiglu_ffn: false
-                image_size: 518
-                use_cls_token: true
-
-
-    denoiser_cfg:
-      target: hy3dshape.models.denoisers.hunyuandit.HunYuanDiTPlain
-      params:
-        input_size: *num_latents
-        in_channels: 64
-        hidden_size: 2048
-        context_dim: 1024
-        depth: 21
-        num_heads: 16
-        qk_norm: true
-        text_len: 1370
-        with_decoupled_ca: false
-        use_attention_pooling: false
-        qk_norm_type: 'rms'
-        qkv_bias: false
-        use_pos_emb: false
-        num_moe_layers: 6
-        num_experts: 8
-        moe_top_k: 2
-        
-    scheduler_cfg:
-      transport:
-        target: hy3dshape.models.diffusion.transport.create_transport
-        params:
-          path_type: Linear
-          prediction: velocity
-      sampler:
-        target: hy3dshape.models.diffusion.transport.Sampler
-        params: {}
-        ode_params:
-          sampling_method: euler # dopri5 ...
-          num_steps: &num_steps 50
-
-    optimizer_cfg:
-      optimizer:
-        target: torch.optim.AdamW
-        params:
-          betas: [0.9, 0.99]
-          eps: 1.e-6
-          weight_decay: 1.e-2
-
-      scheduler:
-        target: hy3dshape.utils.trainings.lr_scheduler.LambdaWarmUpCosineFactorScheduler
-        params:
-          warm_up_steps: 50 # 5000
-          f_start: 1.e-6
-          f_min: 1.e-3
-          f_max: 1.0
-
-    pipeline_cfg:
-      target: hy3dshape.pipelines.Hunyuan3DDiTFlowMatchingPipeline
-
-    image_processor_cfg:
-      target: hy3dshape.preprocessors.ImageProcessorV2
-      params: {}
-
-callbacks:
-    logger:
-      target: hy3dshape.utils.trainings.mesh_log_callback.ImageConditionalASLDiffuserLogger
-      params:
-        step_frequency: 100 # 10000
-        num_samples: 1
-        sample_times: 1
-        mean: *mean
-        std: *std
-        bounds: [-1.01, -1.01, -1.01, 1.01, 1.01, 1.01]
-        octree_depth: 8
-        num_chunks: 50000
-        mc_level: 0.0
-    
-    file_loggers:
-        target: hy3dshape.utils.trainings.mesh_log_callback.ImageConditionalFixASLDiffuserLogger
-        params:
-          step_frequency: 50 # 5000
-          test_data_path: "tools/mini_testset/images.json"

hy3dshape/configs/hunyuandit-mini-overfitting-flowmatching-dinog518-bf16-lr1e4-4096.yaml

@@ -1,180 +0,0 @@
-name: "DiT: Flux large flowmatching; VAE: 1024 token length; ImageEncoder: DINO Giant; ImageSize: 518"
-
-training:
-  steps: 10_0000_0000
-  use_amp: true
-  amp_type: "bf16"
-  base_lr: 1e-4
-  gradient_clip_val: 1.0
-  gradient_clip_algorithm: "norm"
-  every_n_train_steps: 2000 # 5000
-  val_check_interval: 50 # 4096
-  limit_val_batches: 16
-
-dataset:
-  target: hy3dshape.data.dit_asl.AlignedShapeLatentModule
-  params:
-    #! Base setting
-    batch_size: 2
-    num_workers: 8
-    val_num_workers: 4
-
-    # Data 
-    train_data_list: tools/mini_trainset/preprocessed
-    val_data_list: tools/mini_trainset/preprocessed
-
-    #! Image loading
-    cond_stage_key: "image" # image / text / image_text
-    image_size: 518
-    mean: &mean [0.5, 0.5, 0.5]
-    std: &std [0.5, 0.5, 0.5]
-
-    #! Point cloud sampling
-    pc_size: &pc_size 81920
-    pc_sharpedge_size: &pc_sharpedge_size 0
-    sharpedge_label: &sharpedge_label true
-    return_normal: true
-
-    #! Augmentation
-    padding: true
-
-model:
-  target: hy3dshape.models.diffusion.flow_matching_sit.Diffuser
-  params:
-    first_stage_key: "surface"
-    cond_stage_key: "image"
-    scale_by_std: false
-    z_scale_factor: &z_scale_factor 1.0039506158752403
-    torch_compile: false
-
-    # ema_config:
-    #   ema_model: LitEma
-    #   ema_decay: 0.999
-    #   ema_inference: false
-
-    first_stage_config:
-      target: hy3dshape.models.autoencoders.ShapeVAE
-      from_pretrained: tencent/Hunyuan3D-2.1
-      params:
-        num_latents: &num_latents 4096
-        embed_dim: 64
-        num_freqs: 8
-        include_pi: false
-        heads: 16
-        width: 1024
-        num_encoder_layers: 8
-        num_decoder_layers: 16
-        qkv_bias: false
-        qk_norm: true
-        scale_factor: *z_scale_factor
-        geo_decoder_mlp_expand_ratio: 4
-        geo_decoder_downsample_ratio: 1
-        geo_decoder_ln_post: true
-        point_feats: 4
-        pc_size: *pc_size
-        pc_sharpedge_size: *pc_sharpedge_size
-
-    cond_stage_config:
-      target: hy3dshape.models.conditioner.SingleImageEncoder
-      params:
-        main_image_encoder:
-            type: DinoImageEncoder # dino large
-            kwargs:
-                config:
-                  attention_probs_dropout_prob: 0.0
-                  drop_path_rate: 0.0
-                  hidden_act: gelu
-                  hidden_dropout_prob: 0.0
-                  hidden_size: 1024
-                  image_size: 518
-                  initializer_range: 0.02
-                  layer_norm_eps: 1.e-6
-                  layerscale_value: 1.0
-                  mlp_ratio: 4
-                  model_type: dinov2
-                  num_attention_heads: 16
-                  num_channels: 3
-                  num_hidden_layers: 24
-                  patch_size: 14
-                  qkv_bias: true
-                  torch_dtype: float32
-                  use_swiglu_ffn: false
-                image_size: 518
-                use_cls_token: true
-
-
-    denoiser_cfg:
-      target: hy3dshape.models.denoisers.hunyuandit.HunYuanDiTPlain
-      params:
-        input_size: *num_latents
-        in_channels: 64
-        hidden_size: 2048
-        context_dim: 1024
-        depth: 11
-        num_heads: 16
-        qk_norm: true
-        text_len: 1370
-        with_decoupled_ca: false
-        use_attention_pooling: false
-        qk_norm_type: 'rms'
-        qkv_bias: false
-        use_pos_emb: false
-        num_moe_layers: 6
-        num_experts: 8
-        moe_top_k: 2
-        
-    scheduler_cfg:
-      transport:
-        target: hy3dshape.models.diffusion.transport.create_transport
-        params:
-          path_type: Linear
-          prediction: velocity
-      sampler:
-        target: hy3dshape.models.diffusion.transport.Sampler
-        params: {}
-        ode_params:
-          sampling_method: euler # dopri5 ...
-          num_steps: &num_steps 50
-
-    optimizer_cfg:
-      optimizer:
-        target: torch.optim.AdamW
-        params:
-          betas: [0.9, 0.99]
-          eps: 1.e-6
-          weight_decay: 1.e-2
-
-      scheduler:
-        target: hy3dshape.utils.trainings.lr_scheduler.LambdaWarmUpCosineFactorScheduler
-        params:
-          warm_up_steps: 50 # 5000
-          f_start: 1.e-6
-          f_min: 1.e-3
-          f_max: 1.0
-
-    pipeline_cfg:
-      target: hy3dshape.pipelines.Hunyuan3DDiTFlowMatchingPipeline
-
-    image_processor_cfg:
-      target: hy3dshape.preprocessors.ImageProcessorV2
-      params: {}
-
-callbacks:
-    logger:
-      target: hy3dshape.utils.trainings.mesh_log_callback.ImageConditionalASLDiffuserLogger
-      params:
-        step_frequency: 100 # 10000
-        num_samples: 1
-        sample_times: 1
-        mean: *mean
-        std: *std
-        bounds: [-1.01, -1.01, -1.01, 1.01, 1.01, 1.01]
-        octree_depth: 8
-        num_chunks: 50000
-        mc_level: 0.0
-    
-    file_loggers:
-        target: hy3dshape.utils.trainings.mesh_log_callback.ImageConditionalFixASLDiffuserLogger
-        params:
-          step_frequency: 50 # 5000
-          test_data_path: "tools/mini_testset/images.json"

hy3dshape/configs/hunyuandit-mini-overfitting-flowmatching-dinog518-bf16-lr1e4-512.yaml

@@ -1,180 +0,0 @@
-name: "DiT: Flux large flowmatching; VAE: 1024 token length; ImageEncoder: DINO Giant; ImageSize: 518"
-
-training:
-  steps: 10_0000_0000
-  use_amp: true
-  amp_type: "bf16"
-  base_lr: 1e-4
-  gradient_clip_val: 1.0
-  gradient_clip_algorithm: "norm"
-  every_n_train_steps: 2000 # 5000
-  val_check_interval: 50 # 4096
-  limit_val_batches: 16
-
-dataset:
-  target: hy3dshape.data.dit_asl.AlignedShapeLatentModule
-  params:
-    #! Base setting
-    batch_size: 2
-    num_workers: 8
-    val_num_workers: 4
-
-    # Data 
-    train_data_list: tools/mini_trainset/preprocessed
-    val_data_list: tools/mini_trainset/preprocessed
-
-    #! Image loading
-    cond_stage_key: "image" # image / text / image_text
-    image_size: 518
-    mean: &mean [0.5, 0.5, 0.5]
-    std: &std [0.5, 0.5, 0.5]
-
-    #! Point cloud sampling
-    pc_size: &pc_size 81920
-    pc_sharpedge_size: &pc_sharpedge_size 0
-    sharpedge_label: &sharpedge_label true
-    return_normal: true
-
-    #! Augmentation
-    padding: true
-
-model:
-  target: hy3dshape.models.diffusion.flow_matching_sit.Diffuser
-  params:
-    first_stage_key: "surface"
-    cond_stage_key: "image"
-    scale_by_std: false
-    z_scale_factor: &z_scale_factor 1.0039506158752403
-    torch_compile: false
-
-    # ema_config:
-    #   ema_model: LitEma
-    #   ema_decay: 0.999
-    #   ema_inference: false
-
-    first_stage_config:
-      target: hy3dshape.models.autoencoders.ShapeVAE
-      from_pretrained: tencent/Hunyuan3D-2.1
-      params:
-        num_latents: &num_latents 512
-        embed_dim: 64
-        num_freqs: 8
-        include_pi: false
-        heads: 16
-        width: 1024
-        num_encoder_layers: 8
-        num_decoder_layers: 16
-        qkv_bias: false
-        qk_norm: true
-        scale_factor: *z_scale_factor
-        geo_decoder_mlp_expand_ratio: 4
-        geo_decoder_downsample_ratio: 1
-        geo_decoder_ln_post: true
-        point_feats: 4
-        pc_size: *pc_size
-        pc_sharpedge_size: *pc_sharpedge_size
-
-    cond_stage_config:
-      target: hy3dshape.models.conditioner.SingleImageEncoder
-      params:
-        main_image_encoder:
-            type: DinoImageEncoder # dino large
-            kwargs:
-                config:
-                  attention_probs_dropout_prob: 0.0
-                  drop_path_rate: 0.0
-                  hidden_act: gelu
-                  hidden_dropout_prob: 0.0
-                  hidden_size: 1024
-                  image_size: 518
-                  initializer_range: 0.02
-                  layer_norm_eps: 1.e-6
-                  layerscale_value: 1.0
-                  mlp_ratio: 4
-                  model_type: dinov2
-                  num_attention_heads: 16
-                  num_channels: 3
-                  num_hidden_layers: 24
-                  patch_size: 14
-                  qkv_bias: true
-                  torch_dtype: float32
-                  use_swiglu_ffn: false
-                image_size: 518
-                use_cls_token: true
-
-
-    denoiser_cfg:
-      target: hy3dshape.models.denoisers.hunyuandit.HunYuanDiTPlain
-      params:
-        input_size: *num_latents
-        in_channels: 64
-        hidden_size: 768
-        context_dim: 1024
-        depth: 6
-        num_heads: 12
-        qk_norm: true
-        text_len: 1370
-        with_decoupled_ca: false
-        use_attention_pooling: false
-        qk_norm_type: 'rms'
-        qkv_bias: false
-        use_pos_emb: false
-        num_moe_layers: 3
-        num_experts: 4
-        moe_top_k: 2
-        
-    scheduler_cfg:
-      transport:
-        target: hy3dshape.models.diffusion.transport.create_transport
-        params:
-          path_type: Linear
-          prediction: velocity
-      sampler:
-        target: hy3dshape.models.diffusion.transport.Sampler
-        params: {}
-        ode_params:
-          sampling_method: euler # dopri5 ...
-          num_steps: &num_steps 50
-
-    optimizer_cfg:
-      optimizer:
-        target: torch.optim.AdamW
-        params:
-          betas: [0.9, 0.99]
-          eps: 1.e-6
-          weight_decay: 1.e-2
-
-      scheduler:
-        target: hy3dshape.utils.trainings.lr_scheduler.LambdaWarmUpCosineFactorScheduler
-        params:
-          warm_up_steps: 50 # 5000
-          f_start: 1.e-6
-          f_min: 1.e-3
-          f_max: 1.0
-
-    pipeline_cfg:
-      target: hy3dshape.pipelines.Hunyuan3DDiTFlowMatchingPipeline
-
-    image_processor_cfg:
-      target: hy3dshape.preprocessors.ImageProcessorV2
-      params: {}
-
-callbacks:
-    logger:
-      target: hy3dshape.utils.trainings.mesh_log_callback.ImageConditionalASLDiffuserLogger
-      params:
-        step_frequency: 100 # 10000
-        num_samples: 1
-        sample_times: 1
-        mean: *mean
-        std: *std
-        bounds: [-1.01, -1.01, -1.01, 1.01, 1.01, 1.01]
-        octree_depth: 8
-        num_chunks: 50000
-        mc_level: 0.0
-    
-    file_loggers:
-        target: hy3dshape.utils.trainings.mesh_log_callback.ImageConditionalFixASLDiffuserLogger
-        params:
-          step_frequency: 50 # 5000
-          test_data_path: "tools/mini_testset/images.json"

hy3dshape/hy3dshape/__init__.py

@@ -1,17 +0,0 @@
-# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT
-# except for the third-party components listed below.
-# Hunyuan 3D does not impose any additional limitations beyond what is outlined
-# in the repsective licenses of these third-party components.
-# Users must comply with all terms and conditions of original licenses of these third-party
-# components and must ensure that the usage of the third party components adheres to
-# all relevant laws and regulations.
-
-# For avoidance of doubts, Hunyuan 3D means the large language models and
-# their software and algorithms, including trained model weights, parameters (including
-# optimizer states), machine-learning model code, inference-enabling code, training-enabling code,
-# fine-tuning enabling code and other elements of the foregoing made publicly available
-# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT.
-
-from .pipelines import Hunyuan3DDiTPipeline, Hunyuan3DDiTFlowMatchingPipeline
-from .postprocessors import FaceReducer, FloaterRemover, DegenerateFaceRemover, MeshSimplifier
-from .preprocessors import ImageProcessorV2, IMAGE_PROCESSORS, DEFAULT_IMAGEPROCESSOR

hy3dshape/hy3dshape/models/__init__.py

@@ -1,28 +0,0 @@
-# Open Source Model Licensed under the Apache License Version 2.0
-# and Other Licenses of the Third-Party Components therein:
-# The below Model in this distribution may have been modified by THL A29 Limited
-# ("Tencent Modifications"). All Tencent Modifications are Copyright (C) 2024 THL A29 Limited.
-
-# Copyright (C) 2024 THL A29 Limited, a Tencent company.  All rights reserved.
-# The below software and/or models in this distribution may have been
-# modified by THL A29 Limited ("Tencent Modifications").
-# All Tencent Modifications are Copyright (C) THL A29 Limited.
-
-# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT
-# except for the third-party components listed below.
-# Hunyuan 3D does not impose any additional limitations beyond what is outlined
-# in the repsective licenses of these third-party components.
-# Users must comply with all terms and conditions of original licenses of these third-party
-# components and must ensure that the usage of the third party components adheres to
-# all relevant laws and regulations.
-
-# For avoidance of doubts, Hunyuan 3D means the large language models and
-# their software and algorithms, including trained model weights, parameters (including
-# optimizer states), machine-learning model code, inference-enabling code, training-enabling code,
-# fine-tuning enabling code and other elements of the foregoing made publicly available
-# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT.
-
-
-from .autoencoders import ShapeVAE
-from .conditioner import DualImageEncoder, SingleImageEncoder, DinoImageEncoder, CLIPImageEncoder
-from .denoisers import Hunyuan3DDiT

hy3dshape/hy3dshape/models/autoencoders/__init__.py

@@ -1,20 +0,0 @@
-# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT
-# except for the third-party components listed below.
-# Hunyuan 3D does not impose any additional limitations beyond what is outlined
-# in the repsective licenses of these third-party components.
-# Users must comply with all terms and conditions of original licenses of these third-party
-# components and must ensure that the usage of the third party components adheres to
-# all relevant laws and regulations.
-
-# For avoidance of doubts, Hunyuan 3D means the large language models and
-# their software and algorithms, including trained model weights, parameters (including
-# optimizer states), machine-learning model code, inference-enabling code, training-enabling code,
-# fine-tuning enabling code and other elements of the foregoing made publicly available
-# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT.
-
-from .attention_blocks import CrossAttentionDecoder
-from .attention_processors import FlashVDMCrossAttentionProcessor, CrossAttentionProcessor, \
-    FlashVDMTopMCrossAttentionProcessor
-from .model import ShapeVAE, VectsetVAE
-from .surface_extractors import SurfaceExtractors, MCSurfaceExtractor, DMCSurfaceExtractor, Latent2MeshOutput
-from .volume_decoders import HierarchicalVolumeDecoding, FlashVDMVolumeDecoding, VanillaVolumeDecoder

hy3dshape/hy3dshape/models/autoencoders/attention_blocks.py

@@ -1,716 +0,0 @@
-# Open Source Model Licensed under the Apache License Version 2.0
-# and Other Licenses of the Third-Party Components therein:
-# The below Model in this distribution may have been modified by THL A29 Limited
-# ("Tencent Modifications"). All Tencent Modifications are Copyright (C) 2024 THL A29 Limited.
-
-# Copyright (C) 2024 THL A29 Limited, a Tencent company.  All rights reserved.
-# The below software and/or models in this distribution may have been
-# modified by THL A29 Limited ("Tencent Modifications").
-# All Tencent Modifications are Copyright (C) THL A29 Limited.
-
-# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT
-# except for the third-party components listed below.
-# Hunyuan 3D does not impose any additional limitations beyond what is outlined
-# in the repsective licenses of these third-party components.
-# Users must comply with all terms and conditions of original licenses of these third-party
-# components and must ensure that the usage of the third party components adheres to
-# all relevant laws and regulations.
-
-# For avoidance of doubts, Hunyuan 3D means the large language models and
-# their software and algorithms, including trained model weights, parameters (including
-# optimizer states), machine-learning model code, inference-enabling code, training-enabling code,
-# fine-tuning enabling code and other elements of the foregoing made publicly available
-# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT.
-
-
-import os
-from typing import Optional, Union, List
-
-import torch
-import torch.nn as nn
-from einops import rearrange
-from torch import Tensor
-
-from .attention_processors import CrossAttentionProcessor
-from ...utils import logger
-
-scaled_dot_product_attention = nn.functional.scaled_dot_product_attention
-
-if os.environ.get('USE_SAGEATTN', '0') == '1':
-    try:
-        from sageattention import sageattn
-    except ImportError:
-        raise ImportError('Please install the package "sageattention" to use this USE_SAGEATTN.')
-    scaled_dot_product_attention = sageattn
-
-
-class FourierEmbedder(nn.Module):
-    """The sin/cosine positional embedding. Given an input tensor `x` of shape [n_batch, ..., c_dim], it converts
-    each feature dimension of `x[..., i]` into:
-        [
-            sin(x[..., i]),
-            sin(f_1*x[..., i]),
-            sin(f_2*x[..., i]),
-            ...
-            sin(f_N * x[..., i]),
-            cos(x[..., i]),
-            cos(f_1*x[..., i]),
-            cos(f_2*x[..., i]),
-            ...
-            cos(f_N * x[..., i]),
-            x[..., i]     # only present if include_input is True.
-        ], here f_i is the frequency.
-
-    Denote the space is [0 / num_freqs, 1 / num_freqs, 2 / num_freqs, 3 / num_freqs, ..., (num_freqs - 1) / num_freqs].
-    If logspace is True, then the frequency f_i is [2^(0 / num_freqs), ..., 2^(i / num_freqs), ...];
-    Otherwise, the frequencies are linearly spaced between [1.0, 2^(num_freqs - 1)].
-
-    Args:
-        num_freqs (int): the number of frequencies, default is 6;
-        logspace (bool): If logspace is True, then the frequency f_i is [..., 2^(i / num_freqs), ...],
-            otherwise, the frequencies are linearly spaced between [1.0, 2^(num_freqs - 1)];
-        input_dim (int): the input dimension, default is 3;
-        include_input (bool): include the input tensor or not, default is True.
-
-    Attributes:
-        frequencies (torch.Tensor): If logspace is True, then the frequency f_i is [..., 2^(i / num_freqs), ...],
-                otherwise, the frequencies are linearly spaced between [1.0, 2^(num_freqs - 1);
-
-        out_dim (int): the embedding size, if include_input is True, it is input_dim * (num_freqs * 2 + 1),
-            otherwise, it is input_dim * num_freqs * 2.
-
-    """
-
-    def __init__(self,
-                 num_freqs: int = 6,
-                 logspace: bool = True,
-                 input_dim: int = 3,
-                 include_input: bool = True,
-                 include_pi: bool = True) -> None:
-
-        """The initialization"""
-
-        super().__init__()
-
-        if logspace:
-            frequencies = 2.0 ** torch.arange(
-                num_freqs,
-                dtype=torch.float32
-            )
-        else:
-            frequencies = torch.linspace(
-                1.0,
-                2.0 ** (num_freqs - 1),
-                num_freqs,
-                dtype=torch.float32
-            )
-
-        if include_pi:
-            frequencies *= torch.pi
-
-        self.register_buffer("frequencies", frequencies, persistent=False)
-        self.include_input = include_input
-        self.num_freqs = num_freqs
-
-        self.out_dim = self.get_dims(input_dim)
-
-    def get_dims(self, input_dim):
-        temp = 1 if self.include_input or self.num_freqs == 0 else 0
-        out_dim = input_dim * (self.num_freqs * 2 + temp)
-
-        return out_dim
-
-    def forward(self, x: torch.Tensor) -> torch.Tensor:
-        """ Forward process.
-
-        Args:
-            x: tensor of shape [..., dim]
-
-        Returns:
-            embedding: an embedding of `x` of shape [..., dim * (num_freqs * 2 + temp)]
-                where temp is 1 if include_input is True and 0 otherwise.
-        """
-
-        if self.num_freqs > 0:
-            embed = (x[..., None].contiguous() * self.frequencies).view(*x.shape[:-1], -1)
-            if self.include_input:
-                return torch.cat((x, embed.sin(), embed.cos()), dim=-1)
-            else:
-                return torch.cat((embed.sin(), embed.cos()), dim=-1)
-        else:
-            return x
-
-
-class DropPath(nn.Module):
-    """Drop paths (Stochastic Depth) per sample  (when applied in main path of residual blocks).
-    """
-
-    def __init__(self, drop_prob: float = 0., scale_by_keep: bool = True):
-        super(DropPath, self).__init__()
-        self.drop_prob = drop_prob
-        self.scale_by_keep = scale_by_keep
-
-    def forward(self, x):
-        """Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).
-
-        This is the same as the DropConnect impl I created for EfficientNet, etc networks, however,
-        the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper...
-        See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ... I've opted for
-        changing the layer and argument names to 'drop path' rather than mix DropConnect as a layer name and use
-        'survival rate' as the argument.
-
-        """
-        if self.drop_prob == 0. or not self.training:
-            return x
-        keep_prob = 1 - self.drop_prob
-        shape = (x.shape[0],) + (1,) * (x.ndim - 1)  # work with diff dim tensors, not just 2D ConvNets
-        random_tensor = x.new_empty(shape).bernoulli_(keep_prob)
-        if keep_prob > 0.0 and self.scale_by_keep:
-            random_tensor.div_(keep_prob)
-        return x * random_tensor
-
-    def extra_repr(self):
-        return f'drop_prob={round(self.drop_prob, 3):0.3f}'
-
-
-class MLP(nn.Module):
-    def __init__(
-        self, *,
-        width: int,
-        expand_ratio: int = 4,
-        output_width: int = None,
-        drop_path_rate: float = 0.0
-    ):
-        super().__init__()
-        self.width = width
-        self.c_fc = nn.Linear(width, width * expand_ratio)
-        self.c_proj = nn.Linear(width * expand_ratio, output_width if output_width is not None else width)
-        self.gelu = nn.GELU()
-        self.drop_path = DropPath(drop_path_rate) if drop_path_rate > 0. else nn.Identity()
-
-    def forward(self, x):
-        return self.drop_path(self.c_proj(self.gelu(self.c_fc(x))))
-
-
-class QKVMultiheadCrossAttention(nn.Module):
-    def __init__(
-        self,
-        *,
-        heads: int,
-        n_data: Optional[int] = None,
-        width=None,
-        qk_norm=False,
-        norm_layer=nn.LayerNorm
-    ):
-        super().__init__()
-        self.heads = heads
-        self.n_data = n_data
-        self.q_norm = norm_layer(width // heads, elementwise_affine=True, eps=1e-6) if qk_norm else nn.Identity()
-        self.k_norm = norm_layer(width // heads, elementwise_affine=True, eps=1e-6) if qk_norm else nn.Identity()
-
-        self.attn_processor = CrossAttentionProcessor()
-
-    def forward(self, q, kv):
-        _, n_ctx, _ = q.shape
-        bs, n_data, width = kv.shape
-        attn_ch = width // self.heads // 2
-        q = q.view(bs, n_ctx, self.heads, -1)
-        kv = kv.view(bs, n_data, self.heads, -1)
-        k, v = torch.split(kv, attn_ch, dim=-1)
-
-        q = self.q_norm(q)
-        k = self.k_norm(k)
-        q, k, v = map(lambda t: rearrange(t, 'b n h d -> b h n d', h=self.heads), (q, k, v))
-        out = self.attn_processor(self, q, k, v)
-        out = out.transpose(1, 2).reshape(bs, n_ctx, -1)
-        return out
-
-
-class MultiheadCrossAttention(nn.Module):
-    def __init__(
-        self,
-        *,
-        width: int,
-        heads: int,
-        qkv_bias: bool = True,
-        n_data: Optional[int] = None,
-        data_width: Optional[int] = None,
-        norm_layer=nn.LayerNorm,
-        qk_norm: bool = False,
-        kv_cache: bool = False,
-    ):
-        super().__init__()
-        self.n_data = n_data
-        self.width = width
-        self.heads = heads
-        self.data_width = width if data_width is None else data_width
-        self.c_q = nn.Linear(width, width, bias=qkv_bias)
-        self.c_kv = nn.Linear(self.data_width, width * 2, bias=qkv_bias)
-        self.c_proj = nn.Linear(width, width)
-        self.attention = QKVMultiheadCrossAttention(
-            heads=heads,
-            n_data=n_data,
-            width=width,
-            norm_layer=norm_layer,
-            qk_norm=qk_norm
-        )
-        self.kv_cache = kv_cache
-        self.data = None
-
-    def forward(self, x, data):
-        x = self.c_q(x)
-        if self.kv_cache:
-            if self.data is None:
-                self.data = self.c_kv(data)
-                logger.info('Save kv cache,this should be called only once for one mesh')
-            data = self.data
-        else:
-            data = self.c_kv(data)
-        x = self.attention(x, data)
-        x = self.c_proj(x)
-        return x
-
-
-class ResidualCrossAttentionBlock(nn.Module):
-    def __init__(
-        self,
-        *,
-        n_data: Optional[int] = None,
-        width: int,
-        heads: int,
-        mlp_expand_ratio: int = 4,
-        data_width: Optional[int] = None,
-        qkv_bias: bool = True,
-        norm_layer=nn.LayerNorm,
-        qk_norm: bool = False
-    ):
-        super().__init__()
-
-        if data_width is None:
-            data_width = width
-
-        self.attn = MultiheadCrossAttention(
-            n_data=n_data,
-            width=width,
-            heads=heads,
-            data_width=data_width,
-            qkv_bias=qkv_bias,
-            norm_layer=norm_layer,
-            qk_norm=qk_norm
-        )
-        self.ln_1 = norm_layer(width, elementwise_affine=True, eps=1e-6)
-        self.ln_2 = norm_layer(data_width, elementwise_affine=True, eps=1e-6)
-        self.ln_3 = norm_layer(width, elementwise_affine=True, eps=1e-6)
-        self.mlp = MLP(width=width, expand_ratio=mlp_expand_ratio)
-
-    def forward(self, x: torch.Tensor, data: torch.Tensor):
-        x = x + self.attn(self.ln_1(x), self.ln_2(data))
-        x = x + self.mlp(self.ln_3(x))
-        return x
-
-
-class QKVMultiheadAttention(nn.Module):
-    def __init__(
-        self,
-        *,
-        heads: int,
-        n_ctx: int,
-        width=None,
-        qk_norm=False,
-        norm_layer=nn.LayerNorm
-    ):
-        super().__init__()
-        self.heads = heads
-        self.n_ctx = n_ctx
-        self.q_norm = norm_layer(width // heads, elementwise_affine=True, eps=1e-6) if qk_norm else nn.Identity()
-        self.k_norm = norm_layer(width // heads, elementwise_affine=True, eps=1e-6) if qk_norm else nn.Identity()
-
-    def forward(self, qkv):
-        bs, n_ctx, width = qkv.shape
-        attn_ch = width // self.heads // 3
-        qkv = qkv.view(bs, n_ctx, self.heads, -1)
-        q, k, v = torch.split(qkv, attn_ch, dim=-1)
-
-        q = self.q_norm(q)
-        k = self.k_norm(k)
-
-        q, k, v = map(lambda t: rearrange(t, 'b n h d -> b h n d', h=self.heads), (q, k, v))
-        out = scaled_dot_product_attention(q, k, v).transpose(1, 2).reshape(bs, n_ctx, -1)
-        return out
-
-
-class MultiheadAttention(nn.Module):
-    def __init__(
-        self,
-        *,
-        n_ctx: int,
-        width: int,
-        heads: int,
-        qkv_bias: bool,
-        norm_layer=nn.LayerNorm,
-        qk_norm: bool = False,
-        drop_path_rate: float = 0.0
-    ):
-        super().__init__()
-        self.n_ctx = n_ctx
-        self.width = width
-        self.heads = heads
-        self.c_qkv = nn.Linear(width, width * 3, bias=qkv_bias)
-        self.c_proj = nn.Linear(width, width)
-        self.attention = QKVMultiheadAttention(
-            heads=heads,
-            n_ctx=n_ctx,
-            width=width,
-            norm_layer=norm_layer,
-            qk_norm=qk_norm
-        )
-        self.drop_path = DropPath(drop_path_rate) if drop_path_rate > 0. else nn.Identity()
-
-    def forward(self, x):
-        x = self.c_qkv(x)
-        x = self.attention(x)
-        x = self.drop_path(self.c_proj(x))
-        return x
-
-
-class ResidualAttentionBlock(nn.Module):
-    def __init__(
-        self,
-        *,
-        n_ctx: int,
-        width: int,
-        heads: int,
-        qkv_bias: bool = True,
-        norm_layer=nn.LayerNorm,
-        qk_norm: bool = False,
-        drop_path_rate: float = 0.0,
-    ):
-        super().__init__()
-        self.attn = MultiheadAttention(
-            n_ctx=n_ctx,
-            width=width,
-            heads=heads,
-            qkv_bias=qkv_bias,
-            norm_layer=norm_layer,
-            qk_norm=qk_norm,
-            drop_path_rate=drop_path_rate
-        )
-        self.ln_1 = norm_layer(width, elementwise_affine=True, eps=1e-6)
-        self.mlp = MLP(width=width, drop_path_rate=drop_path_rate)
-        self.ln_2 = norm_layer(width, elementwise_affine=True, eps=1e-6)
-
-    def forward(self, x: torch.Tensor):
-        x = x + self.attn(self.ln_1(x))
-        x = x + self.mlp(self.ln_2(x))
-        return x
-
-
-class Transformer(nn.Module):
-    def __init__(
-        self,
-        *,
-        n_ctx: int,
-        width: int,
-        layers: int,
-        heads: int,
-        qkv_bias: bool = True,
-        norm_layer=nn.LayerNorm,
-        qk_norm: bool = False,
-        drop_path_rate: float = 0.0
-    ):
-        super().__init__()
-        self.n_ctx = n_ctx
-        self.width = width
-        self.layers = layers
-        self.resblocks = nn.ModuleList(
-            [
-                ResidualAttentionBlock(
-                    n_ctx=n_ctx,
-                    width=width,
-                    heads=heads,
-                    qkv_bias=qkv_bias,
-                    norm_layer=norm_layer,
-                    qk_norm=qk_norm,
-                    drop_path_rate=drop_path_rate
-                )
-                for _ in range(layers)
-            ]
-        )
-
-    def forward(self, x: torch.Tensor):
-        for block in self.resblocks:
-            x = block(x)
-        return x
-
-
-class CrossAttentionDecoder(nn.Module):
-
-    def __init__(
-        self,
-        *,
-        num_latents: int,
-        out_channels: int,
-        fourier_embedder: FourierEmbedder,
-        width: int,
-        heads: int,
-        mlp_expand_ratio: int = 4,
-        downsample_ratio: int = 1,
-        enable_ln_post: bool = True,
-        qkv_bias: bool = True,
-        qk_norm: bool = False,
-        label_type: str = "binary"
-    ):
-        super().__init__()
-
-        self.enable_ln_post = enable_ln_post
-        self.fourier_embedder = fourier_embedder
-        self.downsample_ratio = downsample_ratio
-        self.query_proj = nn.Linear(self.fourier_embedder.out_dim, width)
-        if self.downsample_ratio != 1:
-            self.latents_proj = nn.Linear(width * downsample_ratio, width)
-        if self.enable_ln_post == False:
-            qk_norm = False
-        self.cross_attn_decoder = ResidualCrossAttentionBlock(
-            n_data=num_latents,
-            width=width,
-            mlp_expand_ratio=mlp_expand_ratio,
-            heads=heads,
-            qkv_bias=qkv_bias,
-            qk_norm=qk_norm
-        )
-
-        if self.enable_ln_post:
-            self.ln_post = nn.LayerNorm(width)
-        self.output_proj = nn.Linear(width, out_channels)
-        self.label_type = label_type
-        self.count = 0
-
-    def set_cross_attention_processor(self, processor):
-        self.cross_attn_decoder.attn.attention.attn_processor = processor
-
-    def set_default_cross_attention_processor(self):
-        self.cross_attn_decoder.attn.attention.attn_processor = CrossAttentionProcessor
-
-    def forward(self, queries=None, query_embeddings=None, latents=None):
-        if query_embeddings is None:
-            query_embeddings = self.query_proj(self.fourier_embedder(queries).to(latents.dtype))
-        self.count += query_embeddings.shape[1]
-        if self.downsample_ratio != 1:
-            latents = self.latents_proj(latents)
-        x = self.cross_attn_decoder(query_embeddings, latents)
-        if self.enable_ln_post:
-            x = self.ln_post(x)
-        occ = self.output_proj(x)
-        return occ
-
-
-def fps(
-    src: torch.Tensor,
-    batch: Optional[Tensor] = None,
-    ratio: Optional[Union[Tensor, float]] = None,
-    random_start: bool = True,
-    batch_size: Optional[int] = None,
-    ptr: Optional[Union[Tensor, List[int]]] = None,
-):
-    src = src.float()
-    from torch_cluster import fps as fps_fn
-    output = fps_fn(src, batch, ratio, random_start, batch_size, ptr)
-    return output
-
-
-class PointCrossAttentionEncoder(nn.Module):
-
-    def __init__(
-        self, *,
-        num_latents: int,
-        downsample_ratio: float,
-        pc_size: int,
-        pc_sharpedge_size: int,
-        fourier_embedder: FourierEmbedder,
-        point_feats: int,
-        width: int,
-        heads: int,
-        layers: int,
-        normal_pe: bool = False,
-        qkv_bias: bool = True,
-        use_ln_post: bool = False,
-        use_checkpoint: bool = False,
-        qk_norm: bool = False
-    ):
-
-        super().__init__()
-
-        self.use_checkpoint = use_checkpoint
-        self.num_latents = num_latents
-        self.downsample_ratio = downsample_ratio
-        self.point_feats = point_feats
-        self.normal_pe = normal_pe
-
-        if pc_sharpedge_size == 0:
-            print(
-                f'PointCrossAttentionEncoder INFO: pc_sharpedge_size is not given, using pc_size as pc_sharpedge_size')
-        else:
-            print(
-                f'PointCrossAttentionEncoder INFO: pc_sharpedge_size is given, using pc_size={pc_size}, pc_sharpedge_size={pc_sharpedge_size}')
-
-        self.pc_size = pc_size
-        self.pc_sharpedge_size = pc_sharpedge_size
-
-        self.fourier_embedder = fourier_embedder
-
-        self.input_proj = nn.Linear(self.fourier_embedder.out_dim + point_feats, width)
-        self.cross_attn = ResidualCrossAttentionBlock(
-            width=width,
-            heads=heads,
-            qkv_bias=qkv_bias,
-            qk_norm=qk_norm
-        )
-
-        self.self_attn = None
-        if layers > 0:
-            self.self_attn = Transformer(
-                n_ctx=num_latents,
-                width=width,
-                layers=layers,
-                heads=heads,
-                qkv_bias=qkv_bias,
-                qk_norm=qk_norm
-            )
-
-        if use_ln_post:
-            self.ln_post = nn.LayerNorm(width)
-        else:
-            self.ln_post = None
-
-    def sample_points_and_latents(self, pc: torch.FloatTensor, feats: Optional[torch.FloatTensor] = None):
-        B, N, D = pc.shape
-        num_pts = self.num_latents * self.downsample_ratio
-
-        # Compute number of latents
-        num_latents = int(num_pts / self.downsample_ratio)
-
-        # Compute the number of random and sharpedge latents
-        num_random_query = self.pc_size / (self.pc_size + self.pc_sharpedge_size) * num_latents
-        num_sharpedge_query = num_latents - num_random_query
-
-        # Split random and sharpedge surface points
-        random_pc, sharpedge_pc = torch.split(pc, [self.pc_size, self.pc_sharpedge_size], dim=1)
-        assert random_pc.shape[1] <= self.pc_size, "Random surface points size must be less than or equal to pc_size"
-        assert sharpedge_pc.shape[
-                   1] <= self.pc_sharpedge_size, "Sharpedge surface points size must be less than or equal to pc_sharpedge_size"
-
-        # Randomly select random surface points and random query points
-        input_random_pc_size = int(num_random_query * self.downsample_ratio)
-        random_query_ratio = num_random_query / input_random_pc_size
-        idx_random_pc = torch.randperm(random_pc.shape[1], device=random_pc.device)[:input_random_pc_size]
-        input_random_pc = random_pc[:, idx_random_pc, :]
-        flatten_input_random_pc = input_random_pc.view(B * input_random_pc_size, D)
-        N_down = int(flatten_input_random_pc.shape[0] / B)
-        batch_down = torch.arange(B).to(pc.device)
-        batch_down = torch.repeat_interleave(batch_down, N_down)
-        idx_query_random = fps(flatten_input_random_pc, batch_down, ratio=random_query_ratio)
-        query_random_pc = flatten_input_random_pc[idx_query_random].view(B, -1, D)
-
-        # Randomly select sharpedge surface points and sharpedge query points
-        input_sharpedge_pc_size = int(num_sharpedge_query * self.downsample_ratio)
-        if input_sharpedge_pc_size == 0:
-            input_sharpedge_pc = torch.zeros(B, 0, D, dtype=input_random_pc.dtype).to(pc.device)
-            query_sharpedge_pc = torch.zeros(B, 0, D, dtype=query_random_pc.dtype).to(pc.device)
-        else:
-            sharpedge_query_ratio = num_sharpedge_query / input_sharpedge_pc_size
-            idx_sharpedge_pc = torch.randperm(sharpedge_pc.shape[1], device=sharpedge_pc.device)[
-                               :input_sharpedge_pc_size]
-            input_sharpedge_pc = sharpedge_pc[:, idx_sharpedge_pc, :]
-            flatten_input_sharpedge_surface_points = input_sharpedge_pc.view(B * input_sharpedge_pc_size, D)
-            N_down = int(flatten_input_sharpedge_surface_points.shape[0] / B)
-            batch_down = torch.arange(B).to(pc.device)
-            batch_down = torch.repeat_interleave(batch_down, N_down)
-            idx_query_sharpedge = fps(flatten_input_sharpedge_surface_points, batch_down, ratio=sharpedge_query_ratio)
-            query_sharpedge_pc = flatten_input_sharpedge_surface_points[idx_query_sharpedge].view(B, -1, D)
-
-        # Concatenate random and sharpedge surface points and query points
-        query_pc = torch.cat([query_random_pc, query_sharpedge_pc], dim=1)
-        input_pc = torch.cat([input_random_pc, input_sharpedge_pc], dim=1)
-
-        # PE
-        query = self.fourier_embedder(query_pc)
-        data = self.fourier_embedder(input_pc)
-
-        # Concat normal if given
-        if self.point_feats != 0:
-
-            random_surface_feats, sharpedge_surface_feats = torch.split(feats, [self.pc_size, self.pc_sharpedge_size],
-                                                                        dim=1)
-            input_random_surface_feats = random_surface_feats[:, idx_random_pc, :]
-            flatten_input_random_surface_feats = input_random_surface_feats.view(B * input_random_pc_size, -1)
-            query_random_feats = flatten_input_random_surface_feats[idx_query_random].view(B, -1,
-                                                                                           flatten_input_random_surface_feats.shape[
-                                                                                               -1])
-
-            if input_sharpedge_pc_size == 0:
-                input_sharpedge_surface_feats = torch.zeros(B, 0, self.point_feats,
-                                                            dtype=input_random_surface_feats.dtype).to(pc.device)
-                query_sharpedge_feats = torch.zeros(B, 0, self.point_feats, dtype=query_random_feats.dtype).to(
-                    pc.device)
-            else:
-                input_sharpedge_surface_feats = sharpedge_surface_feats[:, idx_sharpedge_pc, :]
-                flatten_input_sharpedge_surface_feats = input_sharpedge_surface_feats.view(B * input_sharpedge_pc_size,
-                                                                                           -1)
-                query_sharpedge_feats = flatten_input_sharpedge_surface_feats[idx_query_sharpedge].view(B, -1,
-                                                                                                        flatten_input_sharpedge_surface_feats.shape[
-                                                                                                            -1])
-
-            query_feats = torch.cat([query_random_feats, query_sharpedge_feats], dim=1)
-            input_feats = torch.cat([input_random_surface_feats, input_sharpedge_surface_feats], dim=1)
-
-            if self.normal_pe:
-                query_normal_pe = self.fourier_embedder(query_feats[..., :3])
-                input_normal_pe = self.fourier_embedder(input_feats[..., :3])
-                query_feats = torch.cat([query_normal_pe, query_feats[..., 3:]], dim=-1)
-                input_feats = torch.cat([input_normal_pe, input_feats[..., 3:]], dim=-1)
-
-            query = torch.cat([query, query_feats], dim=-1)
-            data = torch.cat([data, input_feats], dim=-1)
-
-        if input_sharpedge_pc_size == 0:
-            query_sharpedge_pc = torch.zeros(B, 1, D).to(pc.device)
-            input_sharpedge_pc = torch.zeros(B, 1, D).to(pc.device)
-
-        # print(f'query_pc: {query_pc.shape}')
-        # print(f'input_pc: {input_pc.shape}')
-        # print(f'query_random_pc: {query_random_pc.shape}')
-        # print(f'input_random_pc: {input_random_pc.shape}')
-        # print(f'query_sharpedge_pc: {query_sharpedge_pc.shape}')
-        # print(f'input_sharpedge_pc: {input_sharpedge_pc.shape}')
-
-        return query.view(B, -1, query.shape[-1]), data.view(B, -1, data.shape[-1]), [query_pc, input_pc,
-                                                                                      query_random_pc, input_random_pc,
-                                                                                      query_sharpedge_pc,
-                                                                                      input_sharpedge_pc]
-
-    def forward(self, pc, feats):
-        """
-
-        Args:
-            pc (torch.FloatTensor): [B, N, 3]
-            feats (torch.FloatTensor or None): [B, N, C]
-
-        Returns:
-
-        """
-
-        query, data, pc_infos = self.sample_points_and_latents(pc, feats)
-
-        query = self.input_proj(query)
-        query = query
-        data = self.input_proj(data)
-        data = data
-
-        latents = self.cross_attn(query, data)
-        if self.self_attn is not None:
-            latents = self.self_attn(latents)
-
-        if self.ln_post is not None:
-            latents = self.ln_post(latents)
-
-        return latents, pc_infos

hy3dshape/hy3dshape/models/autoencoders/attention_processors.py

@@ -1,96 +0,0 @@
-# Hunyuan 3D is licensed under the TENCENT HUNYUAN NON-COMMERCIAL LICENSE AGREEMENT
-# except for the third-party components listed below.
-# Hunyuan 3D does not impose any additional limitations beyond what is outlined
-# in the repsective licenses of these third-party components.
-# Users must comply with all terms and conditions of original licenses of these third-party
-# components and must ensure that the usage of the third party components adheres to
-# all relevant laws and regulations.
-
-# For avoidance of doubts, Hunyuan 3D means the large language models and
-# their software and algorithms, including trained model weights, parameters (including
-# optimizer states), machine-learning model code, inference-enabling code, training-enabling code,
-# fine-tuning enabling code and other elements of the foregoing made publicly available
-# by Tencent in accordance with TENCENT HUNYUAN COMMUNITY LICENSE AGREEMENT.
-
-import os
-
-import torch
-import torch.nn.functional as F
-
-scaled_dot_product_attention = F.scaled_dot_product_attention
-if os.environ.get('CA_USE_SAGEATTN', '0') == '1':
-    try:
-        from sageattention import sageattn
-    except ImportError:
-        raise ImportError('Please install the package "sageattention" to use this USE_SAGEATTN.')
-    scaled_dot_product_attention = sageattn
-
-
-class CrossAttentionProcessor:
-    def __call__(self, attn, q, k, v):
-        out = scaled_dot_product_attention(q, k, v)
-        return out
-
-
-class FlashVDMCrossAttentionProcessor:
-    def __init__(self, topk=None):
-        self.topk = topk
-
-    def __call__(self, attn, q, k, v):
-        if k.shape[-2] == 3072:
-            topk = 1024
-        elif k.shape[-2] == 512:
-            topk = 256
-        else:
-            topk = k.shape[-2] // 3
-
-        if self.topk is True:
-            q1 = q[:, :, ::100, :]
-            sim = q1 @ k.transpose(-1, -2)
-            sim = torch.mean(sim, -2)
-            topk_ind = torch.topk(sim, dim=-1, k=topk).indices.squeeze(-2).unsqueeze(-1)
-            topk_ind = topk_ind.expand(-1, -1, -1, v.shape[-1])
-            v0 = torch.gather(v, dim=-2, index=topk_ind)
-            k0 = torch.gather(k, dim=-2, index=topk_ind)
-            out = scaled_dot_product_attention(q, k0, v0)
-        elif self.topk is False:
-            out = scaled_dot_product_attention(q, k, v)
-        else:
-            idx, counts = self.topk
-            start = 0
-            outs = []
-            for grid_coord, count in zip(idx, counts):
-                end = start + count
-                q_chunk = q[:, :, start:end, :]
-                k0, v0 = self.select_topkv(q_chunk, k, v, topk)
-                out = scaled_dot_product_attention(q_chunk, k0, v0)
-                outs.append(out)
-                start += count
-            out = torch.cat(outs, dim=-2)
-        self.topk = False
-        return out
-
-    def select_topkv(self, q_chunk, k, v, topk):
-        q1 = q_chunk[:, :, ::50, :]
-        sim = q1 @ k.transpose(-1, -2)
-        sim = torch.mean(sim, -2)
-        topk_ind = torch.topk(sim, dim=-1, k=topk).indices.squeeze(-2).unsqueeze(-1)
-        topk_ind = topk_ind.expand(-1, -1, -1, v.shape[-1])
-        v0 = torch.gather(v, dim=-2, index=topk_ind)
-        k0 = torch.gather(k, dim=-2, index=topk_ind)
-        return k0, v0
-
-
-class FlashVDMTopMCrossAttentionProcessor(FlashVDMCrossAttentionProcessor):
-    def select_topkv(self, q_chunk, k, v, topk):
-        q1 = q_chunk[:, :, ::30, :]
-        sim = q1 @ k.transpose(-1, -2)
-        # sim = sim.to(torch.float32)
-        sim = sim.softmax(-1)
-        sim = torch.mean(sim, 1)
-        activated_token = torch.where(sim > 1e-6)[2]
-        index = torch.unique(activated_token, return_counts=True)[0].unsqueeze(0).unsqueeze(0).unsqueeze(-1)
-        index = index.expand(-1, v.shape[1], -1, v.shape[-1])
-        v0 = torch.gather(v, dim=-2, index=index)
-        k0 = torch.gather(k, dim=-2, index=index)
-        return k0, v0