MiniMax-H3-Image-Character-Swap-LoRA

This is a rank 16 LoRA trained by community author akatz using AI Toolkit. The training data consists of paired samples of "original image + character reference sheet + instruction + edited result." The final round used 140 pairs, with 18 pairs reserved for validation. The workflow is adapted from the official MiniMax H3 Image Edit template: the original image connects to Image 1, and the reference sheet connects to Image 2. Prompts refer to them using <Picture 1> and <Picture 2>. When multiple people are present in the scene, specify the person to be swapped by position or clothing. The model card includes four sets of same-seed comparisons (original image, with LoRA, without LoRA, and reference sheet).

This is a rank 16 LoRA trained by community author akatz using AI Toolkit. The training data consists of paired samples of "original image + character reference sheet + instruction + edited result." The final round used 140 pairs, with 18 pairs reserved for validation. The workflow is adapted from the official MiniMax H3 Image Edit template: the original image connects to Image 1, and the reference sheet connects to Image 2. Prompts refer to them using <Picture 1> and <Picture 2>. When multiple people are present in the scene, specify the person to be swapped by position or clothing. The model card includes four sets of same-seed comparisons (original image, with LoRA, without LoRA, and reference sheet).

Why it matters Preserves the pose, composition, and lighting of the original image while swapping in the face, clothing, and art style of the character from the reference sheet; includes a native ComfyUI node workflow, requiring no custom nodes.

Hugs 14 · Hug Score 13 · Downloads 149

License: Conditional commercial use · ComfyUI: Mentioned in model card · VRAM: Not specified

Regional License: The MiniMax H3 open-weight license excludes the EU, UK, South Korea, and the US (see Excluded Territories in the LICENSE); separate authorization from MiniMax is required in these four regions.

Reasons

  • Test conditions specified: ComfyUI 0.45.0, RTX 4090, INT8 ConvRot base model, LoRA strength 1.0, 6 steps, CFG 1
  • Author lists known issues: may leave remnants of the original clothing, alter details around the target, or swap the wrong person

Usage

  • Download h3_image_character_swap_v1.safetensors and place it in models/loras. Arrange the H3 Image INT8 base model, Qwen3-VL text encoder, and VAE according to the model card table. Open examples/H3-Image-Character-Swap-V1.json. Load the original image and reference sheet into two separate Load Image nodes. Output is approximately 100 pixels, maintaining the original aspect ratio

Risks

  • Follows the MiniMax H3 Image community license, which does not apply to the EU, UK, South Korea, and the US
  • Obtain portrait rights authorization before swapping people in real photos; generated images should be labeled as AI-generated
  • Evaluated only on static images using the INT8 ConvRot base model; other precisions such as BF16 are unverified

Open

抱脸 image-to-image 上升榜

Explore the latest AI Radar updates