An $11 LoRA turns MiniMax H3 into a character-swap tool

Akatz Labs has released a character-swap LoRA for MiniMax H3. Given a source clip and one reference image, the adapter replaces a selected person while attempting to preserve the camera movement, background, lighting, objects, audio, and other people. It targets MiniMax H3’s Ref2VA, or reference-to-video-audio, model.

A LoRA, short for Low-Rank Adaptation, adds a small set of trained parameters to a larger model. Developers can give the base model a specialized behavior without retraining or redistributing its full weights. Akatz trained this adapter overnight on a rented GPU for about $11 and published the weights, dataset, and configuration.

One image in, one performer swapped

The documented workflow accepts a source video through <Video 1> and a full-body reference image or character sheet through <Picture 1>. A prompt identifies the person to replace. The generated clip then follows the source performance with the referenced character, without requiring external pose skeletons, depth maps, rotoscoping, or rigging.

ComfyUI setup

1. Place the final .safetensors checkpoint in ComfyUI/models/loras/.
2. Load MiniMax H3 Ref2VA and apply the adapter at strength 1.0 with a compatible model-only LoRA loader.
3. Connect the source clip as <Video 1> and the replacement image as <Picture 1>.
4. Identify the target person by clothing, position, or another visible attribute. The adapter requires no trained trigger word.
5. Start with a continuous four- to five-second clip at 24 fps. Tests using the full 14-second window showed more drift.

Prompt example

Replace only the man in the purple shirt in <Video 1> with the character in <Picture 1>. Preserve the source video's camera, background, lighting, objects, and all other people.