Akatz Labs trained a character replacement LoRA for MiniMax-H3 multi-reference mode (Ref2VA): Input an original video and a character image to replace the specified person in the video with the character from the image, while following the original video's pose, gaze, expression, and motion, and preserving the scene and other people. V1 was first released on 9 25, V1.5 (recommended version) was released on 10 16, and a native ComfyUI workflow compatible with the author's H3 Relay 2.0 was added on 10 17. It has not been included in the last 7 periods.
Rising score 57 · Likes 704 · Downloads 137211
Reasons
- Solves the need to "preserve performance, swap actors": In the prompt, use <Video 1> to refer to the original video and <Picture 1> to refer to the character image, clearly specifying who to replace (e.g., "the man in the purple shirt"). The author states that no additional trigger words were trained.
- V1.5 was further trained for 1,000 steps starting from V1, incorporating motion video targets and face cropping. The author provided comparison videos for V1 / V1.5, claiming that V1.5 better preserves stylized faces, hairstyles, and character identity, while explicitly noting that these are selected examples, not a comprehensive evaluation.
- The training process is public: The V1 dataset akatz-ai/H3-Character-Swap-v1 and the run-summary.json for V1.5 (rank 16, learning rate 1e-5, starting file hash) are in the repository; the V1.5 file is approximately 155 MB.
Usage
- ComfyUI: Additionally prepare the H3 Ref2VA base model, text encoder, and VAE. Place h3_character_swap_v1.5.safetensors in the LoRA directory, use a model-only LoRA loader, with a strength of 1.0; use V1.5 to replace V1, do not stack both together.
- Recipe tested by the author: Ref2VA INT8 base model + this LoRA at strength 1.0 + a separate minimax_h3_fl2v_turbo_8step_v1.0_768p acceleration LoRA at strength 1.0, 8 steps, CFG 1, sampler er_sde, scheduler normal; 24 fps, start with a short continuous shot (in the author's V1 tests, 4–5 seconds were more stable than 14 seconds).
- If you need masking, long-segment continuation, and preservation of original audio, install the author's H3 Relay 2.0 and use the H3-Character-Swap-Relay-V2 workflow from the repository examples.
Risks
- The author describes this as an experimental model: Character identity, motion timing, expressions, occlusions, and hard cuts may drift; V1 training used only 94 synthesized image editing samples, and multi-person replacement was not specifically trained. This column has not reviewed the comparison videos.
- The original audio in the final video is directly copied back by the tool in post-processing; this does not imply that the LoRA can restore audio or correct lip-sync misalignment.
- Replacing real people involves portrait rights and deepfake risks; only use videos and character images for which you have the right to use. The license is the MiniMax H3 Community License (not Apache); the author warns that its standard authorization does not cover the United States, European Union, United Kingdom, and South Korea. Review the original terms before commercial use or use in these regions.