FastVideo MiniMax H3 — Almost Real-Team Video Generation in ComfyUI (Patreon)

Video : https://youtu.be/Y9J-dXJhz-M

Hey everyone.

New video is up, and this one is about speed. FastVideo, the research lab that has been making fast video generation since Wan 2.1, released a distilled MiniMax H3 that generates video in 4 steps. On their server GPUs that means seconds per clip. On my local machine, it is about 30 seconds for one 5-second clip at 0.7 megapixels, and honestly, that changes how you work. You stop waiting and start iterating.

One thing that surprised me: the setup is not complex at all. People on X were asking if there is some special trick behind the speed, and there is not. You download one 22 GB safetensor file from Kijai's repo, drop it in your diffusion_models folder, and run a plain text-to-video workflow. That is the whole thing.

In the video I also show how I keep two characters consistent across 20 different clips using structured text prompts, and how I used an AI agent to write all 20 prompts for a Japanese anime story in one go.

The workflow I used in the video is attached right here. Grab it, drop in your own model files, and start queueing clips.

RESOURCES:

FastVideo model card: https://huggingface.co/FastVideo/FastVideo-FastH3-4-step-Preview-v1-VSA-DataFree

Kijai MiniMax-H3-experimental (INT8 convrot safetensors): https://huggingface.co/Kijai/MiniMax-H3-experimental/blob/main/minimax_h3_fastvideo_vsa_datafree_1300step_4step_int8_convrot.safetensors

Reactor team example: https://x.com/haozhangml/status/2094188446965547050

Reactor infinite-livestream: https://github.com/reactor-team/infinite-livestream

— Benji