H3 Image Character Swap V1

Follow @akatz_ai on X Patreon GitHub Sponsors Ko-fi

Replace a person in a still image with a character from a reference sheet. I trained this LoRA to bring over the character’s face, costume, and style while following the pose, framing, and lighting in the source.

Four matched character swaps in a 2×2 grid. Rain Rosette and Copper Eel Archivist are above Green Reptilian Adventurer and Rattler Rail Marshal. Each comparison shows the source, V1 LoRA, base H3 Image, and character reference.

Each comparison uses the same source, character reference, prompt, seed, and sampling settings. The V1 panel adds the LoRA at strength 1.0. These are selected examples. Source footage for Rain Rosette comes from Blender Institute’s Tears of Steel, © Blender Foundation. Project credits · CC BY 3.0

Source footage for Copper Eel Archivist comes from Lacoste’s Crocodile Inside. Rattler Rail Marshal uses footage from When the potion seller runs out of ingredients. Copper Eel source · Rattler source

Download V1 LoRA · Powered by MiniMax-H3-Image

Use in ComfyUI

  1. Install the files below and open the H3 Image Character Swap V1 workflow in ComfyUI. The V1 LoRA is already connected at strength 1.0.
  2. To replay Rain Rosette, put the source frame and Rain Rosette character sheet in ComfyUI/input/, or upload them through the two Load Image nodes. Queue the workflow with the included prompt and fixed seed.
  3. For your own swap, load the source into Image 1 and the character reference into Image 2. Use <Picture 1> and <Picture 2> in the prompt. If the shot has multiple people, identify the target by position, clothing, or what they are doing. Set the output dimensions to around 1 megapixel and match the source aspect ratio.

The workflow uses native nodes and was tested in ComfyUI 0.45.0 on an RTX 4090 with dynamic VRAM. It is adapted from the official MiniMax H3 Image Edit template. Upstream workflow license

Component File Folder in ComfyUI/models/
H3 Image INT8 ConvRot minimax_h3_image_pruned_int8_convrot.safetensors diffusion_models/
Qwen3-VL text encoder qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors text_encoders/
H3 Image VAE minimax_h3_image_vae_fp16.safetensors vae/
V1 adapter h3_image_character_swap_v1.safetensors loras/
Setting Tested value
LoRA strength 1.0
Steps 6
CFG 1
Sampler euler
Scheduler simple
Denoise 1.0
Negative conditioning ConditioningZeroOut
Prompt enhancer Off

Example prompt

Replace the woman in the dark jacket in <Picture 1> with the character from <Picture 2>. Keep her position, pose, and expression. Preserve the background and everyone else in the scene. Carry over the reference character’s costume and materials, and match the light in the source shot.

Training and limitations

V1 is a rank-16 LoRA trained with AI Toolkit on paired image edits. Each pair uses a source image, a character sheet, an instruction, and the edited target. The earlier public V1 swap dataset is part of the training history. The final training stage used 140 pairs, with 18 separate validation examples. Earlier V1 dataset

The adapter is intended for people and humanoid creatures. Check small or occluded figures, hands near props, and clothing that blends into the background. The edit can preserve pieces of the original outfit, change details around the target, or pick the wrong person. This release has been evaluated on still images with the INT8 ConvRot base. BF16 and other quantizations are unverified.

File identity

Field Value
File h3_image_character_swap_v1.safetensors
Size 155,110,448 bytes
Rank 16
SHA-256 68378574bede5ca9d7d69df8bb8f8d9f7860d07ea45b26d9c67a7babf50d330f

License and attribution

Use and redistribution are subject to the MiniMax-H3-Image Community License Agreement, including its territorial restrictions and separate authorization requirements. See NOTICE for attribution.

💜 Support Akatz Labs

If this helps with your work, you can help fund the next round of model training and ComfyUI workflows. Thanks for trying it.

GitHub Sponsors · Patreon · Ko-fi · Explore Akatz Labs

Downloads last month
149
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for akatz-ai/MiniMax-H3-Image-Character-Swap-LoRA

Adapter
(2)
this model

Dataset used to train akatz-ai/MiniMax-H3-Image-Character-Swap-LoRA