MiniMax H3 FastVideo: ComfyUI setup and batch generation

This tutorial covers FastVideo's distilled four-step MiniMax H3 model in ComfyUI. The presenter uses Kijai's experimental safetensors release, a 22 GB download, with a regular text-to-video workflow. He recommends a GPU with more VRAM but gives no minimum capacity.

On his machine, the presenter reports about 30 seconds to generate a five-second clip at 0.7 megapixels. His chosen settings use eight sampling steps for better visual quality. He says four steps or a lower resolution can shorten generation time at the cost of quality. The queue demonstration starts with the models already loaded into VRAM, so these timings should not be treated as cold-start results or guarantees for other GPUs. The research runtime example uses CUDA on B200 server GPUs.

The demonstrated model version supports text-to-audio-and-video generation. It does not yet support first/last-frame conditioning or reference-to-video. For character consistency, the presenter repeats structured character descriptions across scenes, while acknowledging some variation remains.

Hermes prepares 20 prompts for five-second clips. The presenter manually queues them in ComfyUI and suggests using an AI agent to automate that work. He also discusses Reactor's chat-driven livestream as an application of faster generation, distinct from the speed of his local demo.