H3-Tiny-ComfyUI

H3-Tiny is a structured pruning of MiniMax H3 (simultaneous audio and video generation, original version 33B): AdaLN projections are approximated with rank 8, 8 blocks are removed from the 50 total to yield 17B (without training), then each layer's FFN is narrowed from 14336 to 9792 channels, and distillation using the 33B model restores performance to achieve 14B. The ComfyUI version released on 11/21 includes three single files: a 17B reference-image-to-video model (Ref2VA), a 17B first-and-last-frame / text-to-video model (FL2VA / T2VA), and a 14B reference-image-to-video model, each 28–34 GB; the author tested them on ComfyUI commit afcb984 at 1344×768 resolution with 243 frames, and the output matches the diffusers version. There is also a few-step distilled version with 6 steps, but it is not included in the ComfyUI release.

H3-Tiny is a structured pruning of MiniMax H3 (simultaneous audio and video generation, original version 33B): AdaLN projections are approximated with rank 8, 8 blocks are removed from the 50 total to yield 17B (without training), then each layer's FFN is narrowed from 14336 to 9792 channels, and distillation using the 33B model restores performance to achieve 14B. The ComfyUI version released on 11/21 includes three single files: a 17B reference-image-to-video model (Ref2VA), a 17B first-and-last-frame / text-to-video model (FL2VA / T2VA), and a 14B reference-image-to-video model, each 28–34 GB; the author tested them on ComfyUI commit afcb984 at 1344×768 resolution with 243 frames, and the output matches the diffusers version. There is also a few-step distilled version with 6 steps, but it is not included in the ComfyUI release.

Why It Matters You can run it by simply swapping the model file in ComfyUI's built-in H3 workflow, without needing custom nodes; the author states that the untrained pruned version works directly.

Hugging Face likes 8 · Hugging Face upvotes 8 · Hugging Face downloads 0

License: Conditional commercial use · ComfyUI: Mentioned in model card · VRAM: ≈116 GB estimated

Regional License: The MiniMax H3 open-weight license excludes the EU, UK, South Korea, and the US (Excluded Territories in LICENSE); these four regions require separate authorization from MiniMax.

Reasons

  • The model card clearly specifies the official workflow nodes, sampling steps, and schedulers corresponding to each file, and notes that Comfy-Org's turbo LoRA has not been tested on the pruned version
  • The pruning and inference code is open source (Apache 2.0), allowing you to prune from the 33B model yourself

Usage

  • Place files from diffusion_models/ into ComfyUI/models/diffusion_models, download the text encoder and two VAEs from Comfy-Org/MiniMax-H3, open the official "MiniMax H3 Image to Video" or "Reference to Video" workflow, and select the H3-Tiny file in Load Diffusion Model; use 50 steps, simple scheduler, BasicGuider without CFG, and do not attach the turbo LoRA

Risks

  • Inherits the MiniMax H3 community license, which does not apply to the EU, UK, South Korea, and the US
  • Individual files are still 28–34 GB; the author tested on a single H200; GitHub documentation states that offloading is required if VRAM is below approximately 110 GB. The automatically estimated 116 GB sums all three files together and is inaccurate
  • Performance comparisons are based solely on samples provided by the author; the paper has not yet been published

Open

抱脸 text-to-video 新模型

Explore the latest AI Radar updates