Skip to main content
Back

Alibaba's TaoMate-H3 Cuts MiniMax Video Generation 11x in ComfyUI

A ComfyUI-ready conversion of Alibaba's TaoMate-H3 LoRA drops MiniMax H3 video generation to just three sampling steps, with reported 10x+ speedups.

Model, Full Access
  • ComfyUI-compatible port of Alibaba's TaoMate-H3 3-step LoRA for MiniMax H3 video model
  • Single 2.48 GB safetensors file, drops into ComfyUI/models/loras/ with 3-step sampling
  • Reported 11.45x faster pure DiT time, 27.66x faster first chunk latent on 8x H20
  • Peak memory essentially unchanged at ~31 GiB versus stock H3
  • Enables streaming audio-video generation with native stereo audio and long-form continuity
  • Already 26,000+ downloads on Hugging Face; weights untouched, only key layout converted

A community conversion of TaoMate-H3 packages Alibaba’s three-step low-rank adaptation, or LoRA, for standard ComfyUI workflows. The adapter accelerates MiniMax H3 video generation by compressing its denoising schedule, while the conversion removes the need to rewrite the model’s parameter keys manually. The Hugging Face page had recorded more than 26,000 downloads at publication.

Three denoising steps, with caveats

MiniMax H3 weights provide an open-weight, multimodal video model that processes text, images, video, and audio in one context. It generates video with synchronized stereo dialogue, sound effects, and music. The open release supports a 768-pixel short edge, 24 frames per second, clips of about 15 seconds, and output capped at roughly one megapixel.

Alibaba’s upstream adapter distills H3’s diffusion schedule into three Stage 3 denoising intervals per chunk. A LoRA modifies selected behavior in the base model through a comparatively small set of adapter weights, so users still need the full MiniMax H3 model.

Alibaba reports the following results for a 10-second, 480-by-864 clip on a node with eight NVIDIA H20 GPUs:

Alibaba’s TaoMate-H3 benchmark
Metric MiniMax H3 TaoMate-H3 Reported speedup
DiT processing time 169.6 seconds 14.8 seconds 11.45×
First final chunk latent