--- license: apache-2.0 base_model: Comfy-Org/MiniMax-H3 base_model_relation: adapter pipeline_tag: text-to-video tags: - minimax-h3 - lora - adapter - comfyui - text-to-video - reference-to-video - audio-video - synchronized-audio - few-step - turbo - accelerated-inference - lightx2v - dynamic-rank - svd - safetensors - bfloat16 - pruned-model - curve-form --- # MiniMax-H3 Turbo LoRAs for ComfyUI This repository contains MiniMax-H3 Turbo LoRAs converted and optimized for ComfyUI: - dynamic-rank FL2V and Ref2V LoRAs derived from the official ComfyUI weights in [lightx2v/Minimax-h3-Turbo](https://huggingface.co/lightx2v/Minimax-h3-Turbo/tree/main); - pruned/curve-form compatibility conversions of the Turbo LoRAs released by [larryvrh](https://huggingface.co/larryvrh). These LoRAs accelerate MiniMax-H3 video and synchronized-audio generation by reducing the required number of sampling steps. --- ## **Update — LightX2V FL2V and Ref2V Dynamic-Rank LoRAs** Three dynamically resized BF16 LoRAs are now included. Their source weights were published in the correct MiniMax-H3 ComfyUI layout by: [lightx2v/Minimax-h3-Turbo](https://huggingface.co/lightx2v/Minimax-h3-Turbo/tree/main) ### New files | File | Source | Purpose | | --- | --- | --- | | `minimax_h3_fl2v_turbo_4step_v1.0_768p_comfyui_resized_avg_rank_21_bf16.safetensors` | `minimax_h3_fl2v_turbo_4step_v1.0_768p_comfyui_bf16.safetensors` | Four-step 768p FL2V generation. | | `minimax_h3_fl2v_turbo_8step_v1.0_comfyui_resized_avg_rank_21_bf16.safetensors` | `minimax_h3_fl2v_turbo_8step_v1.0_comfyui_bf16.safetensors` | Eight-step FL2V generation with a greater quality focus. | | `minimax_h3_ref2v_turbo_4step_v0.1_comfyui_resized_avg_rank_21_bf16.safetensors` | `minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors` | Four-step reference-to-video generation. | | `minimax_h3_fl2v_turbo_4step_v1.1_768p_comfyui_resized_avg_rank_64_bf16.safetensors` | `minimax_h3_fl2v_turbo_4step_v1.1_768p_comfyui_bf16.safetensors` | Four-step 768p FL2V generation, high-fidelity dynamic-rank resize. | | `minimax_h3_fl2v_turbo_4step_v1.1_768p_comfyui_resized_avg_rank_28_bf16.safetensors` | `minimax_h3_fl2v_turbo_4step_v1.1_768p_comfyui_bf16.safetensors` | Four-step 768p FL2V generation, compact dynamic-rank resize. | The LightX2V files were already supplied in the correct ComfyUI namespace and QKV layout. No namespace conversion or AdaLN pruning was required. ### How the LoRAs were resized Each LoRA was compressed using an exact compact SVD calculated from its LoRA A/B factors without materializing the full dense weight updates. The process: 1. Treats every attention, MLP and token-refiner projection independently. 2. Separates fused QKV adapters into their original Q, K and V projections. 3. Calculates the exact singular spectrum of each effective `B @ A` update. 4. Selects a dynamic rank independently for every projection. 5. Preserves one additional safety rank beyond the Frobenius-energy threshold. 6. Rebuilds balanced factors as `A = sqrt(S) × Vh` and `B = U × sqrt(S)`. 7. Re-fuses Q, K and V using concatenated A factors and a block-diagonal B matrix. 8. Bakes the source `alpha / rank` scale into `lora_B`. 9. Removes the no-longer-required alpha tensors. 10. Saves all output tensors as BF16. All three files target the same projection-level average rank: - Rank sum: `6778` - Projection count: `312` - Average projection rank: `21.724359` - Filename rank: `floor(21.724359) = 21` Q, K and V are counted separately. The stored rank of a fused QKV adapter is the sum of its three independently selected ranks. ### Resize results | LoRA | Output size | Average rank | Mean Frobenius norm retained | Baked scale | | --- | ---: | ---: | ---: | ---: | | FL2V 4-step v1.0 768p | 284.36 MiB | 21.724359 | 94.95% | 1.0 | | FL2V 8-step v1.0 | 311.89 MiB | 21.724359 | 97.72% | 0.0625 | | Ref2V 4-step v0.1 | 311.79 MiB | 21.724359 | 98.33% | 0.0625 | | FL2V 4-step v1.1 768p (rank 64) | 891.65 MiB | 64.500000 | 98.18% | 1.0 | | FL2V 4-step v1.1 768p (rank 28) | 375.55 MiB | 28.500000 | 93.59% | 1.0 | The Ref2V source was reduced from approximately 1,865.57 MiB to 311.79 MiB, an 83.29% reduction. The v1.1 768p source (1,865.57 MiB) was reduced to 891.65 MiB (−52.21%) at rank 64 and to 375.55 MiB (−79.87%) at rank 28. A separate Frobenius threshold was solved for each LoRA because their singular-value distributions differ. Reusing one fixed threshold would not produce the same average rank. ### Ref2V numerical validation Direct comparison between the original and resized effective LoRA updates produced: | Measurement | Result | | --- | ---: | | Global cosine similarity | 99.9190% | | Global relative L2 error | 4.0234% | | Mean projection Frobenius norm retained | 98.3332% | | Minimum projection Frobenius norm retained | 96.4904% | | Projection rank range | 2–105 | | Size reduction | 83.29% | These are numerical weight-reconstruction measurements, not a perceptual video-quality score. ### FL2V 4-step v1.1 768p numerical validation Direct comparison between the v1.1 source and each resized variant produced: | Measurement | Rank 64 | Rank 28 | | --- | ---: | ---: | | Global cosine similarity | 99.1169% | 97.3279% | | Global relative L2 error | 13.2601% | 22.9625% | | Mean projection Frobenius norm retained | 98.1774% | 93.5945% | | Minimum projection Frobenius norm retained | 98.0817% | 92.6661% | | Projection rank range | 2–120 | 2–99 | | Size reduction | 52.21% | 79.87% | These are numerical weight-reconstruction measurements, not a perceptual video-quality score. ### Structural validation Every resized LightX2V file contains: | Component | Count | | --- | ---: | | BF16 tensors | 416 | | LoRA A/B pairs | 208 | | Independently ranked projections | 312 | | QKV modules | 52 | | Alpha tensors | 0 | Validation confirmed: - all tensors contain finite BF16 values; - every LoRA A tensor has a matching LoRA B tensor; - ComfyUI recognizes all 208 adapters and all 416 tensors; - no incompatible or unresolved tensors remain; - QKV rank columns remain block-diagonal and ordered as Q, K and V; - the original effective LoRA scale is preserved; - the exact average projection rank is 21.724359. --- ## Workflow Example The following output was generated with the recommended pruned-model workflow using the v4 step-600 EMA LoRA. ### Current workflow [Download the current ComfyUI workflow](https://huggingface.co/drbaph/MiniMax-H3-Turbo-Lora-ComfyUI/resolve/main/fl_minimax_h3_turbo_lora_example_workflow.json?download=true) - Workflow: `fl_minimax_h3_turbo_lora_example_workflow.json` - LoRA: `minimax_h3_turbo_v4_step600_ema_pruned_comfyui.safetensors` - Steps: `6–8` - Sampler: `Euler` - Scheduler: `Beta` - LoRA strength: `1.0` - Video sigma shift: `12` - Audio sigma shift: `4–6` The previous workflow remains available here: [Download the previous workflow](https://huggingface.co/drbaph/MiniMax-H3-Turbo-Lora-ComfyUI/resolve/main/fl_minimax_h3_turbo_lora_example_workflow_old.json?download=true) [View the previous workflow output](https://huggingface.co/drbaph/MiniMax-H3-Turbo-Lora-ComfyUI/resolve/main/assets/MiniMax_H3_00060-audio.mp4) --- ## Included Files ### LightX2V dynamic-rank LoRAs | File | Description | | --- | --- | | `minimax_h3_fl2v_turbo_4step_v1.0_768p_comfyui_resized_avg_rank_21_bf16.safetensors` | Dynamic-SVD resize of the official LightX2V four-step 768p FL2V LoRA. | | `minimax_h3_fl2v_turbo_8step_v1.0_comfyui_resized_avg_rank_21_bf16.safetensors` | Dynamic-SVD resize of the official LightX2V eight-step FL2V LoRA. | | `minimax_h3_ref2v_turbo_4step_v0.1_comfyui_resized_avg_rank_21_bf16.safetensors` | Dynamic-SVD resize of the official LightX2V four-step Ref2V LoRA. | ### Pruned/curve-form compatibility conversions | File | Description | | --- | --- | | `minimax_h3_turbo_v4_step600_ema_pruned_comfyui.safetensors` | **Recommended.** Pruned-model conversion of the v4 step-600 EMA checkpoint. | | `minimax_h3_turbo_v4_step600_pruned_comfyui.safetensors` | Non-EMA v4 step-600 variant. | | `minimax_h3_turbo_4step_ema_ckpt850_pruned_comfyui.safetensors` | EMA checkpoint-850 conversion. | | `minimax_h3_turbo_4step_ckpt850_pruned_comfyui.safetensors` | Non-EMA checkpoint-850 conversion. | | `minimax_h3_turbo_4step_ema_ckpt500_pruned_comfyui.safetensors` | EMA checkpoint-500 conversion. | | `minimax_h3_turbo_4step_ckpt500_pruned_comfyui.safetensors` | Non-EMA checkpoint-500 conversion. | | `minimax_h3_turbo_4step_ema_pruned_comfyui.safetensors` | Initial EMA Turbo preview conversion. | | `minimax_h3_turbo_4step_pruned_comfyui.safetensors` | Initial non-EMA Turbo preview conversion. | ### Workflows and assets | File | Description | | --- | --- | | `fl_minimax_h3_turbo_lora_example_workflow.json` | Current recommended ComfyUI workflow. | | `fl_minimax_h3_turbo_lora_example_workflow_old.json` | Previous workflow retained for comparison. | | `assets/MiniMax_H3_audio.mp4` | Current workflow example output. | | `assets/MiniMax_H3_00060-audio.mp4` | Previous workflow example output. | EMA and non-EMA files contain different tensor values and are retained separately. --- ## Pruned/Curve-Form Compatibility Conversions The older full-model Turbo LoRAs use keys such as: `blocks.0.attn.qkv_proj.lora_A.weight` ComfyUI expects: `diffusion_model.blocks.0.attn.qkv_proj.lora_A.weight` The compatibility conversion applies: - `blocks.*` → `diffusion_model.blocks.*` - `token_refiner.*` → `diffusion_model.token_refiner.*` The original full-model LoRAs also contain AdaLN adapters whose dimensions do not match the pruned/curve-form MiniMax-H3 checkpoint. Those incompatible adapters were removed. | Component | Count | | --- | ---: | | Source tensors | 518 | | Source LoRA A/B pairs | 259 | | Retained tensors | 416 | | Retained LoRA A/B pairs | 208 | | Removed AdaLN tensors | 102 | | Removed AdaLN A/B pairs | 51 | The removed adapters consist of 50 AdaLN projection pairs from `blocks.0` through `blocks.49` and one pair from `final_layer.adaln_proj.linear`. These files are partial compatibility conversions and should not be assumed to behave identically to the original full-model LoRA used with its intended complete MiniMax-H3 base model. The LightX2V dynamic-rank LoRAs do not use this pruning process. --- ## Usage Place the LoRA files in: `ComfyUI/models/loras/` Load the selected LoRA with its compatible MiniMax-H3 model and workflow. Recommended starting strength: `1.0` If adjustment is needed, test approximately: `0.8–1.2` When comparing files, keep the following identical: - prompt; - seed; - resolution; - frame count; - base model; - conditioning inputs; - sampler; - scheduler; - LoRA strength; - sigma shifts; - workflow. --- ## Sampling Guidance ### Four steps Maximum speed and useful for rapid previews. Use the matching four-step LightX2V FL2V or Ref2V workflow. ### Six steps A good speed-to-quality compromise for the v4 step-600 EMA pruned conversion. ### Eight steps Recommended when prioritizing quality, motion consistency and synchronized-audio stability. For the existing pruned-model workflow, start with: - Steps: `8` - Sampler: `Euler` - Scheduler: `Beta` - LoRA strength: `1.0` - Video sigma shift: `12` - Audio sigma shift: `4–6` Follow the source project recommendations when using the LightX2V LoRAs: [lightx2v/Minimax-h3-Turbo](https://huggingface.co/lightx2v/Minimax-h3-Turbo/tree/main) --- ## Audio and Scheduler Notes MiniMax-H3 generates video and synchronized audio as separate streams with different scheduling requirements. Incorrect audio scheduling may cause: - severe distortion; - noise-like output; - unstable or blown-out audio; - completely broken audio. If audio is unstable, check: - sampler and scheduler configuration; - audio and video sigma shifts; - ComfyUI version; - MiniMax-H3 sampling implementation; - custom sampler nodes; - whether the workflow matches the selected FL2V or Ref2V LoRA. Broken audio should not automatically be attributed to the LoRA. --- ## Optional Acceleration These LoRAs may be used with compatible ComfyUI acceleration methods such as: - Comfy Kitchen attention; - SageAttention; - Sol Attention; - Spectrum. These methods are optional and are not included in this repository. Some attention backends replace one another rather than stacking, and patch order depends on the implementation. For few-step Turbo workflows, establish a clean baseline before adding approximate attention or trajectory-forecasting methods. --- ## Attribution ### LightX2V LoRAs The dynamically resized FL2V and Ref2V LoRAs are derived from the official ComfyUI-compatible files released by: [lightx2v/Minimax-h3-Turbo](https://huggingface.co/lightx2v/Minimax-h3-Turbo/tree/main) Credit for the source weights, training and release belongs to the LightX2V project and its contributors. This repository provides only the dynamic-rank BF16 resizing, packaging and validation described above. ### Original Turbo LoRAs The older Turbo LoRA weights, training work and distillation method were released by: [larryvrh/MiniMax-H3-Turbo-Lora](https://huggingface.co/larryvrh/MiniMax-H3-Turbo-Lora) Original ComfyUI Turbo custom node: [ComfyUI-MiniMax-H3-Turbo](https://github.com/Larryvrh/ComfyUI-MiniMax-H3-Turbo) This repository does not claim authorship of: - the original LoRA weights; - LoRA training or distillation; - the few-step sampling methods; - MiniMax-H3 video/audio scheduling; - the MiniMax-H3 base model; - the pruned/curve-form checkpoint; - the LightX2V source weights; - the original project documentation or generation code. The modifications provided here are limited to: - pruned/curve-form ComfyUI compatibility conversions; - namespace adjustments; - removal of incompatible AdaLN adapters where required; - exact dynamic-rank SVD resizing of the LightX2V ComfyUI LoRAs; - BF16 safetensors packaging and validation. Please retain attribution and comply with the licenses and terms of the original LoRAs and MiniMax-H3 base model when using or redistributing these files.