MiniMax H3 action workflow: sampling and noise cleanup

This tutorial covers a ComfyUI reference-to-video workflow using MiniMax H3 for fast movement, including running and fighting. The presenter says he tested 161 videos while adapting YZ Jingyu's original workflow. He explicitly limits this version to action shots and recommends other workflows for talking heads and close-up skin.

The setup uses MiniMax H3 Hybrid with fl2va and ref2va, plus the dmad LoRA at 0.8 strength. A speed slider ranges from -8 to 8; the presenter recommends 3 to 4 for fast action. SelfLift two-stage sampling uses the LMS version 0.1 upscaler, which he prefers to the original latent upscaling for these tests, though he says its sharpening can be excessive.

Controls cover duration, aspect ratio, resolution and reference images. The presenter reports generating 15 seconds at 1.8 megapixels. He mentions 2 megapixels as a possibility but says he has not tested it. More reference images increase compute use and generation time.

The default sampling recipe is eight low-resolution steps, four high-resolution steps and one finishing refiner step intended to reduce pixel noise. For backgrounds that look dirty, he suggests six low-resolution steps, five high-resolution steps and the refiner. The resources include a downloadable workflow and model setup script; RunningHub provides a cloud version.