MiniMax H3 character swap tutorial in ComfyUI

This ComfyUI tutorial explains how to replace a person in footage with the MiniMax H3 Character Swap LoRA. The workflow connects a character sheet and a source video to the ReferenceToVideo node. A dedicated prompt template identifies the person to replace and points to the reference image. The speaker stresses that attaching the files alone does not specify which character the model should swap.

The speaker reports using MiniMax H3 locally and demonstrates a workflow without a SAM 2 or SAM 3 masking step. Examples cover walking footage, a kick and a backflip. In one example, a camera held by the original subject remains in the output. These are results from the speaker's tests, rather than guarantees for other footage.

For sampling, the speaker recommends res multistep, describes Euler as inconsistent and says the beta scheduler gives more texture in these tests. A latent upscaler then enlarges the result by 1.2x to 1.5x.

The main limitation is character count: the speaker says the LoRA was trained for one character and that two-character swaps work inconsistently. Extra prompt text can introduce fire or energy effects, but it can also change the background. Keeping the prompt simple is the speaker's advice for a straightforward replacement.