MiniMax-H3-Image-Character-Swap Image Editing Model: 4GB+ VRAM

MiniMax-H3-Image-Character-Swap Image Editing Model: 4GB+ VRAM

Sample outputs are available on the model card.

At a Glance

Item Value
Repository akhaliq/MiniMax-H3-Image-Character-Swap
Family guide MiniMax-H3 guide (14 articles)
Publisher guide MiniMax: models and licenses
Published 2026-11-14
License minimax-h3-image-community-license-agreement
Formats safetensors
Source type Unverified (not confirmed by a primary source)

Values determined by this site’s code when the information was collected. Dates are JST.

Overview

Note: This news has not been officially confirmed and is based on unverified information. The following content is based on the details provided in the Hugging Face model card.

akhaliq/MiniMax-H3-Image-Character-Swap is described as a checkpoint where akatz-ai/MiniMax-H3-Character-Swap-LoRA has been merged directly into the MiniMaxAI/MiniMax-H3-Image image DiT. By inputting a scene photo (), a character reference image (), and text specifying the person to be replaced, the model outputs an image with the character swapped while preserving the background, lighting, other people, and poses (image-to-image). According to the card, it also supports swapping anime-style characters into live-action scenes, which the base model reportedly fails to do.

Specifications

  • Base: MiniMaxAI/MiniMax-H3-Image. The card states this is the first merged version that ports a character-swap LoRA trained for video into an image model.
  • Merge Details: It is stated that h3_character_swap_v1.5.safetensors (rank 16, 312 modules, attention+FFN of 50 DiT blocks and 2 token-refiner blocks) was fused at a strength of 1.0.
  • Base Model DiT (from the original model card): A 33.1B parameter single-stream transformer. It uses the MiniMax-H3 text encoder as-is.
  • Output Resolution: The base model generates at approximately 16MP (4096×4096 to 6272×2688). This merged version’s card recommends a 4MP starting canvas. The examples use width=2720, height=1536 (multiples of 32), and it is reported that 16MP also works.
  • Recommended Settings: euler / simple, 6 steps, CFG 1, denoise 1. As it is a distilled model, no adjustments are required.
  • Input: Two images (scene, then character). It is recommended to specify the person and include instructions in the prompt to “maintain face, hair, and clothing, while preserving background, lighting, camera, and other people."
  • Format: Only the transformer_diffusers/ layout of the modular pipeline is included. The traditional MiniMaxH3ImagePipeline layout is not included.
  • License Terms (Summary from the card):
  • Free for use, including commercial use, in all regions except the EU, UK, South Korea, and the US. – Written permission from MiniMax is required if annual revenue exceeds $1M or if provided as a hosted service to third parties. – No rights are granted under the community license in excluded regions. – MiniMax reportedly does not claim rights to the generated output. – The card itself states this is “not legal advice," so checking the LICENSE file is necessary.

Performance

The following is the publisher’s (akhaliq) own evaluation and has not been verified by a third party.

  • In a 3-way comparison (native, standalone LoRA, merged version) of 6 scenes x 2 characters x 3 seeds (4MP each), 17 scored comparisons were conducted. The merged version was chosen in 14/17 (82%) of cases, the standalone LoRA in 2, and the native in 1.
  • There was only one evaluator, and the card itself states this is “a signal of strong preference, not a benchmark."
  • The comparison dataset is reportedly available at akhaliq/h3-image-swap-tests.
  • Verification was performed on 8 difficult cases (profile, small face, occlusion, full body, group).

The comparison table in the card is primarily image-based and contains no numerical data. The summary is as follows:

Case Native H3-Image This Model
Anime char to profile photo Original face remained (failed in 3/3 seeds) Succeeded with same seed
Live-action char to cafe Skin tone became brighter than reference Maintained identity
Anime char to full-body jogging Pose changed to standing Maintained running pose and clothing

It is reported that the merged version suppresses the skin tone shifts and pose changes seen in the native model. On the other hand, the card lists the following limitations:

  • The LoRA was trained for video, and its porting to images is limited to verification in 8 cases.
  • In extreme profiles or combinations of different styles, the redrawn gaze direction may become unnatural.
  • Deviations may occur in extreme poses, strong occlusions, or very small faces.
  • Results depend on the quality of the reference image.

It is specified that the orientation of the replaced person is adjusted to match the source person.

Strengths and Use Cases

According to the model card, the goal is “image editing to replace an existing person with the face, hair, and clothing of a reference character." These are all based on the publisher’s own description and are unverified.

  • Cross-style swapping, such as replacing anime/illustration-style characters into live-action scenes.
  • Swapping into non-frontal faces, such as profiles. It targets cases where the base MiniMax-H3-Image failed in 3/3 seeds.
  • Editing where you want to keep the background, lighting, other people, and poses, such as conversations in a cafe or full-body jogging.
  • Editing where you specify a person to replace, such as in a group photo.

The original LoRA (akatz-ai/MiniMax-H3-Character-Swap-LoRA) was trained for video-based swapping. This model is a derivative version that ports it to an image model and fuses it into a single checkpoint, aiming for use without additional LoRA files or nodes. The tags are image-to-image and image-editing.

It is stated that a clean multi-view character sheet is best for the reference image. When replacing real people, the card requests consideration for consent and portrait rights.

How It Differs from Similar Models

As an article in the same MiniMax-H3 series, there is “Pika-H3-Flex" Video Generation Model with Audio: List of Distributed Files. This is a LoRA adapter aimed at generating videos with synchronized stereo audio. This model differs in that it specializes in character replacement for still images and has the LoRA already fused into the checkpoint.

“FastH3-Omni-Comfy" Video/Audio Generation Model with Image/Audio Reference: Required VRAM 80GB+ / List of Distributed Files is a Reference-to-Video system that generates video and audio simultaneously using image, video, and audio as references. This model is for editing, taking two images as input and outputting one image, so the use case is different.

Hardware Requirements

Estimated requirements (calculated by Local Model Watch)

Your VRAM Quantization File size Est. memory needed
4GB (laptop iGPU / phone class) Original precision 0.0GB 0.0GB

Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.

Can You Run It Locally?

The publisher distributes this model as safetensors.

License — minimax-h3-image-community-license-agreement: A custom license from the publisher. Check the original terms directly, including whether commercial use is permitted.

Compiled by this site’s code from the published formats and the license field. License summaries are not legal advice — check the publisher’s original terms before relying on them.

Our Own Measurements

Measurements We Did Not Take

We have not confirmed that this site meets the commercial-use terms of this model’s license (minimax-h3-image-community-license-agreement). Because this site carries advertising, we did not run the model (no CPU run, answers, quantization comparison, conversion or generation).

Distributed Files

Weight files published in akhaliq/MiniMax-H3-Image-Character-Swap, listed by this site from the Hugging Face API. Sizes are the actual file sizes.

File Size
comfyui/minimax_h3_image_character_swap_bf16.safetensors 40.23GB
lora/h3_character_swap_v1.5.safetensors 155MB
lora/h3_image_character_swap_diffusers.safetensors 173MB
transformer_diffusers/diffusion_pytorch_model-*-of-00007.safetensors 66.29GB (7 split files)
vae_diffusers/diffusion_pytorch_model.safetensors 2.34GB

How to Get It

  • The distribution format is safetensors. A modular pipeline layout for diffusers (transformer_diffusers/) and a single file for ComfyUI are provided. There is no gated setting requiring license agreement.
  • The repository is akhaliq/MiniMax-H3-Image-Character-Swap.
  • For diffusers, diffusers==0.43.0 is required. The text encoder references the MiniMaxAI/MiniMax-H3-Image side and is loaded automatically.
from diffusers import ModularPipeline
REPO = "akhaliq/MiniMax-H3-Image-Character-Swap"
blocks = ModularPipeline.from_pretrained(REPO).blocks
pipe = blocks.init_pipeline(REPO)
pipe.load_components(dtype=torch.bfloat16)
  • Images are passed in the order image=[scene_image, character_image] (scene, then character). width and height should be multiples of 32.
  • According to the card, running the whole system on a single GPU requires approximately 140GB of VRAM (DiT 66GB + Qwen3-VL encoder 67GB). If the encoder is run remotely, only the DiT and VAE need to be on the GPU.
  • ComfyUI requires 0.45.0 or higher and template pack 0.11.104 or higher. A dedicated workflow JSON is in the repository, and it can also be used with the official “MiniMax-H3-Image: Image Edit" template. Point the diffusion model loader to minimax_h3_image_character_swap_bf16.safetensors. Use the text encoder and VAE from Comfy-Org/MiniMax-H3-Image.
  • According to the card, verification of the ComfyUI version is limited to the same calculations confirmed via diffusers but in a different layout. If there are any ComfyUI-specific issues, you are asked to report them in the discussions.
  • If you want to adjust the strength, the standalone LoRA (under lora/) is also distributed in the same repository.

Related Articles

What to Read Next

Sources

This article contains unverified information. We will append an update note once it is confirmed by a primary source.