akatz-ai/MiniMax-H3-Character-Swap-LoRA — new model trending #30 on Hugging Face
Akatz Labs released an experimental MiniMax-H3 character-swap LoRA, trending number 30 on Hugging Face.
Akatz Labs published MiniMax-H3-Character-Swap-LoRA v1, an experimental rank-16 adapter trained for 1,000 updates to replace one character in a source video using an image reference while preserving the scene. It is meant for a MiniMax-H3 Ref2VA runtime such as ComfyUI and is not a standalone model; the base weights are not included. The dataset used for optimization was 76 synthetic image-edit examples plus 32 short video regularization clips. Local tests suggested better background preservation than the base model, but motion timing, facial expressions, hard cuts, and lip-sync remained unreliable.
- Rank-16 LoRA trained for 1,000 updates on MiniMax-H3 Ref2VA.
- Optimized on 76 image edits and 32 short regularization clips.
- Author recommends strength 1.0; it is an adapter, not a base model.
- Expressions, hard cuts, timing, and multi-character swaps remain unreliable.
Full article955 words · extracted from huggingface.co · click to collapse
# MiniMax H3 Character Swap LoRA v1
An experimental character-replacement LoRA trained by Akatz Labs for **1,000 updates**. It aims to replace a selected character using an image reference while preserving the source scene. In our local comparisons it often preserved the background more closely than the base model, but motion timing, facial expressions, and hard cuts remain unreliable.
**Download:** [final 1,000-step LoRA](h3_character_swap_pro4500_1000.safetensors). Only the final 1,000-step checkpoint is published. Start with the final checkpoint at **strength 1.0**. This is an adapter, not a standalone model or a Turbo distillation LoRA.
**Dataset:** [H3 Character Swap v1](https://huggingface.co/datasets/akatz-ai/H3-Character-Swap-v1).
## Use
1. Use an H3 Ref2VA-capable runtime and obtain the base model and VAEs separately.
2. Place the final `.safetensors` file in `ComfyUI/models/loras/` (or your configured shared LoRA directory), and apply it to the H3 model at strength 1.0 using a compatible model-only LoRA loader.
3. Supply the source video as `<Video 1>` and replacement image or character sheet as `<Picture 1>`.
4. Specify the target person in the prompt. No additional trigger word was trained.
Example:
> Replace only the man in the purple shirt in <Video 1> with the character in <Picture 1>. Keep the replacement character's identity, outfit, and art style from <Picture 1>. Preserve the source video's camera, background, lighting, objects, and all other people. Match the target person's position, scale, pose, and movement. Do not show the reference sheet or its background.
The training captions were shorter, for example `Swap the man in the purple shirt in <Video 1> with the character in <Picture 1>.` Preservation instructions helped some local evaluations, but stronger expression instructions sometimes suppressed the swap entirely. Prompt wording is not a guarantee of strict source alignment.
Short, continuous shots of roughly 4–5 seconds were more promising than our full 14-second tests. A precise maximum duration has **not** been established. Use 24 fps and your runtime's supported H3 frame grid. The character-swap LoRA does not require a Turbo LoRA, Spectrum, or Sol attention.
## Base model and tested configurations
Training used `minimax_h3_ref2va_pruned_int8_convrot.safetensors` from [Comfy-Org/MiniMax-H3](https://huggingface.co/Comfy-Org/MiniMax-H3), plus the frozen [Ostris Ref2VA training assistant](https://huggingface.co/ostris/minimax_h3_training_adapter). The assistant and base weights are **not merged into this adapter** and are not distributed here. Model revisions and hashes are in `training/base-model-files.json`.