MiniMax H3: ComfyUI setup and reference-to-video tests

This tutorial covers MiniMax H3 setup in ComfyUI and tests text-to-video and reference-to-video generation. The presenter describes publicly available open weights and quantized files for a single GPU, with memory needs depending on the selected weights and hardware.

File selection separates FL2VA weights for text and image inputs from Ref2VA weights for reference-to-video and audio tasks. The walkthrough uses a 34 GB INT8 ConvRot diffusion file and discusses FP8 alternatives. The Qwen3-VL 32B text encoder has BF16, INT8 ConvRot and NVFP4 options; audio and video use separate VAEs. The advertised 2K output requires a second H3 Regenerate 2K pass after the 768p base generation. This demonstration focuses on base output.

The presenter uses Hermes with the official prompting guides to expand prompts. Examples include espresso close-ups, facial detail and BMX motion. With 20 sampling steps on a Pro 6000 workstation GPU, the presenter reports about two minutes per five-second clip, rather than a general performance guarantee.

Reference tests combine characters with audio, which the presenter says influences motion and lighting. Fight-scene character replacement proves less reliable: generated action drifts from the source footage, and the presenter rates these attempts below Seedance. The tutorial gives local AI video users both setup guidance and concrete examples of those limits.