MDMZ's MiniMax H3 character-swap workflow for ComfyUI: SAM 3 motion guide, auto-prompting, and where it breaks

MDMZ published this on Oct 22 and it passed 5,000 views in its first hours. The graph uses the MiniMax H3 video model. SAM 3 masks the person in your source clip so H3 copies their motion, not their identity. Two modes: 'background from video' puts the new character into the original scene, and 'background from image' keeps the character in their own scene and only borrows the movement. Either way the framing and camera move follow the source clip. Leave sound on if you want lip sync.

The settings that matter, from the video: H3 takes clips up to 15 seconds, and the graph converts input to 24 fps (MDMZ processes 120 frames, about 5 seconds). Megapixels at 0.5 gives fast previews and 1.2 adds detail. The SAM 3 target defaults to 'person', but a more specific phrase like 'person dancing' helps when there are several people. Raise motion-guide resolution from 0.15 to 0.2 or more for talking heads and facial expressions. H3 needs a long, detailed prompt. ComfyUI's Gemini node writes it for a few cents, or you paste the instruction block into a free chatbot along with the clip and image.

Results and limits he reports: a 5-second clip took about 10 minutes on his machine, and a Turbo LoRA mode trades quality for speed (he renders around 0.7 MP and then upscales). Coloured lighting carries onto the new character, objects in front of the person are handled, and it works across cuts. Lip sync and overall quality hold up best on close-ups, and distant subjects come out pixelated. The workflow file and written guide sit on his Patreon post; there's also a paid hosted option.

Use it on footage you shot or have rights to, and swap in characters you designed or people who agreed to it. Don't use it to put a real person's face into a clip without their consent. We haven't run the graph ourselves.