MiniMax-H3-Character-Swap-LoRA: MiniMax H3 LoRA for Selective Character Replacement in Video

Input a video and a character image to replace a specific person while preserving the original background and camera motion. This experimental LoRA adapter operates within the Ref2VA workflow based on MiniMax H3, focusing on preserving the context of the source video rather than regenerating the entire scene.

It delivers the most stable results in short continuous shots of 4–5 seconds. The final checkpoint, trained for 1,000 steps, is applied at strength 1.0 and can be used directly in tools like ComfyUI without separate Turbo LoRAs or complex attention settings. Accuracy improves when the replacement target and elements to preserve are specifically specified via prompts.

Background preservation has improved over the base model, but composition or timing may drift in longer videos. Close-up expressions and hard cut handling remain unstable, and multi-person replacement was not trained. Audio uses a post-processing method that reuses the original track, so lip-sync synchronization has limitations.

Suitable for video editing tasks requiring single-person style transfer or character-sheet-based replacement. Distributed under the MiniMax H3 Community License, so regional and conditional terms must be verified before commercial use.