MiniMax H3 in ComfyUI: local video workflow tutorial

This tutorial builds MiniMaxH3 video workflows in ComfyUI on the presenter's own PC. He uses a pruned int8 model on an RTX3090 and discusses other model sizes. The examples demonstrate individual outputs rather than establish general hardware requirements or benchmark results.

The first workflow accepts text, a starting image, an ending image or both endpoints. Prompts describe the intended transition. The presenter also tests a20second clip beyond the stated15second limit; a usable example does not establish a supported maximum.

A separate reference workflow uses images, videos and audio as guidance. Examples combine several input types and reuse character voices in new dialogue. Prompts identify references with labels in angle brackets. Video references take longer in the presenter's reported experience, with runtime dependent on hardware.

The technical section explains model loaders, shared CLIP and VAE requirements, group bypasses, conditioning switches and output settings. Optional frame interpolation is discussed separately. A final section describes the presenter's permission-request process under the model's license. The walkthrough does not establish unrestricted model availability.