MiniMax H3 in ComfyUI: long-video workflow tutorial

Reno explains a MiniMax H3 Motion Context workflow in ComfyUI for joining generated clips into a continuous scene. The opening examples use three 15-second segments. The tutorial pairs Singularity Ref2VA v1.3 with the turbo 4-step v0.1 LoRA, which the presenter says must stay together. He recommends eight sampling steps because, in his tests, audio deteriorates more noticeably than the image over a long sequence.

In the Multi-Track Editor, he selects reference-to-video mode and max quality, then sets later segments to context. He explains that context carries roughly one second of sound and image information from the preceding clip. A shared character reference applies across segments, while a supplied system prompt helps Claude divide a scene into linked prompts. The worked setup uses three eight-second segments.

Reno recommends a single sampling pass. He reports that latent upscaling or a second sampling stage shifts pixels at the joins and exposes seams. Segment numbering starts at one. The metadata distinguishes cloud runs, which render the chain together, from local runs, where saved latents allow individual segments to be redone. RunningHub provides a cloud route before exporting the workflow for local AI use. The final example took three attempts; he suggests starting at 0.5 megapixels with five- to eight-second segments.