Original title: Lakonik/pi-H3
MiniMax H3 few-step adapters uploaded on 11 12 by Hansheng Chen et al. from Stanford University (part of the work was done at Luma AI). These adapters support text-to-video, first-frame, last-frame, and first-and-last-frame conditioning, while simultaneously generating synchronized audio. The method combines pi-Flow v2 data-free few-step distillation with DWD (Decoding Without a Decoder, paper arXiv 2611.22287): in the final step, the denoising network directly predicts pixels without passing through the VAE decoder; the VAE is used only to encode the first and last frames. There are four adapters in total: for both 2 steps and 4 steps, there are DWD versions and versions that use VAE decoding. The DWD versions also include INT8 ConvRot files exclusively for ComfyUI; the authors recommend the DWD versions.
Why it matters Compresses H3 video generation with synchronized audio down to two to four steps, while eliminating the need for a video decoder; the authors provide ComfyUI workflows and INT8 versions, making it testable on consumer-grade GPUs
Upvotes 55 · Likes 75 · Downloads 0
Reasons
- Speeds provided by the authors in the ComfyUI documentation: using 2 steps with DWD + INT8 + Sol sparse attention, on an RTX 3090 (CUDA 13), generating a 5-second video at 864×480 takes 15.5 seconds, and a 10-second video takes 37.3 seconds; generating a 5-second video at 1344×768 takes 38.5 seconds (assuming the model is loaded and the prompt is encoded)
- The codebase also provides FP8 quantization and Sol sparse attention; there is an online demo Space by the authors on Hugging Face, with quality and speed comparisons available in the authors' blog
- All four adapters support text, first-frame, and last-frame conditioning; switching between 2 steps and 4 steps requires changing only one parameter
Usage
- ComfyUI: Install the ComfyUI-piFlow nodes, drag in pi-H3-DWD.json; use minimax_h3_fl2va_pruned_int8_convrot from Comfy-Org/MiniMax-H3 as the base model, place the adapters in models/loras, and switch between 2 steps or 4 steps on the pi-H3 Adapter node
- diffusers: Install according to the LakonLab v0.3 documentation for pi-H3, and use load_lakonlab_adapter to load pi_h3_4step_dwd_adapter from Lakonik/pi-H3; pass first-frame and last-frame images using image= / last_image=
- Try online: https://huggingface.co/spaces/Lakonik/pi-H3
Risks
- The adapters are derivative models of MiniMax H3, distributed under the MiniMax H3 community license; check the license terms before commercial use
- Requires the authors' LakonLab code or ComfyUI-piFlow custom nodes, not native ComfyUI support; the speed table was measured by the authors on a 3090 with gpu-unlock enabled, and this entry has not been run