H3-Image Character Swap Merged Version: Swapping Characters in Scene and Character Images

Uploaded by akhaliq on 11 14: Directly merges Akatz Labs' MiniMax-H3 Character Swap LoRA v1.5 (rank 16, covering 50 DiT blocks and 2 token refiner blocks) into the DiT of MiniMaxAI/MiniMax-H3-Image at strength 1.0, eliminating the need to attach a LoRA. Usage involves providing the scene photo as <Picture 1> and the character reference image as <Picture 2>, clearly specifying who to replace; the model swaps the person while preserving the background, lighting, other people, and the original pose. The repository also provides a diffusers version, a ComfyUI layout merged checkpoint (40 GB), as well as the original LoRA (155 MB) and a converted diffusers LoRA (173 MB). The license follows the MiniMax-H3-Image Community License.

Original Title: akhaliq/MiniMax-H3-Image-Character-Swap

Uploaded by akhaliq on 11 14: Directly merges Akatz Labs' MiniMax-H3 Character Swap LoRA v1.5 (rank 16, covering 50 DiT blocks and 2 token refiner blocks) into the DiT of MiniMaxAI/MiniMax-H3-Image at strength 1.0, eliminating the need to attach a LoRA. Usage involves providing the scene photo as <Picture 1> and the character reference image as <Picture 2>, clearly specifying who to replace; the model swaps the person while preserving the background, lighting, other people, and the original pose. The repository also provides a diffusers version, a ComfyUI layout merged checkpoint (40 GB), as well as the original LoRA (155 MB) and a converted diffusers LoRA (173 MB). The license follows the MiniMax-H3-Image Community License.

Why It Matters Cross-style replacements, such as swapping anime characters into real-life profile shots, often fail with the original version; the author claims the merged version wins in most of their comparisons, and ComfyUI users only need to swap a file.

Rising Score 38 · Likes 179 · Downloads 316

Reasons

  • The author states that the original H3-Image retained the original face in all 3 seeds when swapping anime characters onto strict profile photos; in 17 three-way comparisons (original / standalone LoRA / merged version), the merged version was selected 14 times, the standalone LoRA 2 times, and the original 1 times, but only one person provided annotations
  • 8 difficult cases (profile, small face, occlusion, full body, multiple people) and all blind evaluation images are publicly available in the dataset akhaliq/h3-image-swap-tests
  • Before merging, 624/624 keys and shapes were verified, and 208/208 target tensors were confirmed to match the results with the LoRA attached; the merge script merge/merge_and_validate.py is included in the repository

Usage

  • ComfyUI 0.45.0+, template package 0.11.104+: Place minimax_h3_image_character_swap_bf16.safetensors into models/diffusion_models; use the text encoder and VAE from Comfy-Org/MiniMax-H3-Image; open the character swap workflow in comfyui/workflows in the repository (or the official Image Edit template), requiring only the loading of two images
  • Parameters follow the template: euler / simple, 6 steps, CFG 1, denoise 1, starting at approximately 4 MP; in the prompt, describe the scene first, then the character, explicitly identifying the person to be replaced (e.g., "the man in the yellow sweater second from the left"), and specify preserving the face, hairstyle, clothing, and scene
  • For diffusers 0.43.0, use ModularPipeline to load this repository; fully local execution requires approximately 140 GB of VRAM (DiT 66 GB + text encoder 67 GB); you can place the text encoder remotely, keeping only the DiT and VAE on the GPU; to adjust strength, use the standalone LoRA in lora/

Risks

  • No new training was performed; this simply merges the Akatz Character Swap LoRA, which was collected in previous releases, into the image model; the LoRA was originally trained for video and has only been validated on 8 difficult cases for images
  • Swapping real people involves portrait rights and deepfakes; use only images you have the right to use
  • The applicable regions for the MiniMax-H3-Image Community License do not include the EU, UK, South Korea, and the US; if annual revenue exceeds 100 ten thousand USD or if provided as a hosted service, written authorization from MiniMax must be obtained first. This item has not been run

Open

Hugging Face Trending

Explore the latest AI Radar updates