Wan 2.2 - Tessy LoRA

LoRA entrenado con Ostris AI Toolkit sobre Wan 2.2 T2V A14B.

Detalles del entrenamiento

  • Trigger word: tessy
  • Steps: 2000
  • Rank / Alpha: 32 / 32
  • Dataset: 46 imágenes extraídas de videos
  • Resolución: 512–1024
  • GPU: A100-80GB en Modal
  • Arquitectura: MOE (Mixture of Experts) - high noise + low noise

Archivos

Este LoRA incluye dos archivos porque Wan 2.2 usa arquitectura MOE:

Archivo Descripción
mi_wan22_tessy_lora_v1_high_noise.safetensors Experto de high noise
mi_wan22_tessy_lora_v1_low_noise.safetensors Experto de low noise

Debes usar AMBOS archivos para que funcione correctamente.

Uso con Diffusers

import torch
from diffusers import WanPipeline
from diffusers.utils import export_to_video

pipe = WanPipeline.from_pretrained(
    "ai-toolkit/Wan2.2-T2V-A14B-Diffusers-bf16",
    torch_dtype=torch.bfloat16
)

# Cargar ambos LoRAs
pipe.load_lora_weights(
    "aztro/wan22-tessy-lora",
    weight_name="mi_wan22_tessy_lora_v1_high_noise.safetensors",
    adapter_name="high_noise"
)
pipe.load_lora_weights(
    "aztro/wan22-tessy-lora",
    weight_name="mi_wan22_tessy_lora_v1_low_noise.safetensors",
    adapter_name="low_noise"
)

pipe.to("cuda")

video = pipe(
    prompt="tessy sitting at a podcast desk, medium shot, studio lighting",
    num_frames=49,
    guidance_scale=3.5,
    num_inference_steps=25
).frames[0]

export_to_video(video, "tessy_podcast.mp4", fps=16)
Downloads last month
34
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for aztro/wan22-tessy-lora

Adapter
(53)
this model

Space using aztro/wan22-tessy-lora 1