MiniMax H3 Image for Portrait Editing: Guide
Plan a MiniMax H3 Image portrait edit with clear references: give your portrait, a scene and an outfit photo one role each, then compare every detail.
To plan a portrait edit with MiniMax H3 Image, give every reference image one job and say what each one must not change. Use your own adult portrait for identity, a scene image for light and framing, and an outfit photo for clothing, then compare the result detail by detail. MiniMax H3 Image is an independent MiniMax model. It is not a Lumatch option, and Lumatch has no integration with it.
By the Lumatch team. Sources read on November 15, 2026. We did not sign in to any MiniMax service, run the model or generate images. This guide contains no results, quality claims or comparisons with other models.
What the official README says
The points below come from the MiniMax H3 Image README on GitHub. They are MiniMax's descriptions, not Lumatch findings.
- Tasks: one checkpoint handles both text-to-image and multi-image editing.
- References: an edit can take up to 16 reference images.
- Output size: the README states that it generates images of about 16 MP, with aspect ratios from 1:3 to 3:1. That is an image-size specification. It does not mean every image will be sharp or that faces and skin will stay true to the person.
- Not yet supported: the README's specification table lists mask-based inpainting, transparent backgrounds, user-specified seeds and preview images as not yet supported.
- Prompting tips: state the task before the style, describe style through visible colour, material and light, and give each reference image a role while saying what changes and what stays.
- Licence: the model is released under the MiniMax-H3-Image Community License Agreement, linked from the README as LICENSE. The README describes open weights, but that does not by itself mean unrestricted free use. Read the current licence. This guide does not offer a legal opinion.
- Access: the README lists MiniMax web, desktop and API services. Whether they are available to you and what they cost depends on those platforms at the time you use them.
Give each reference image one role
You do not need 16 references for a profile photo. Three clear inputs are easier to control, and each extra image is another source of conflict.
| Input | Used for | Keep or leave out |
|---|---|---|
| 1. Your portrait (adult, clear, recent) | Identity only | Keep your features, age, hair, skin texture and body |
| 2. Scene image with no people who have not agreed to be used | Light, framing and setting | Do not take anyone's face, body or text from it |
| 3. Outfit photo you have the right to use | One garment | Take the clothing only, not the wearer's face or build |
Use photos you took yourself or have permission to use. A reference that contains another person, a logo or a recognisable place you have never visited can pull those details into your picture.
One complete instruction
This is an original example that follows the README's advice on roles. It has not been tested on the model, and any model may ignore part of it.
Image 1 is my adult portrait. Use it only for identity, and keep my facial features, age, hairstyle, skin texture and body shape. Image 2 is a scene reference. Use only its soft window light, eye-level framing and plain indoor setting, and do not copy any person or text from it. Image 3 is a clothing reference. Dress me in its grey knit sweater with natural folds that follow my posture, without changing my build. Portrait orientation, head and shoulders in frame. No text, letters or logos in the image.
Notice that it names the task, the layout and the visible light before anything else, and that every image has a role and a list of things it may not change.
Compare the result detail by detail
Put the result next to image 1 at the same size and work through each row.
| Check | Discard the image if |
|---|---|
| Facial features | Your eyes, nose, mouth or jaw differ from image 1 |
| Age | Lines, grey hair or skin texture have been smoothed away |
| Hands | Fingers are extra, missing or fused |
| Fabric | Folds ignore your posture, seams break, or the sweater changes your build |
| Background facts | The setting no longer matches image 2, or new objects or text appear |
If two references disagree, for example the scene light is warm while your portrait is lit by cold daylight, do not add more words. Remove a reference or change one variable at a time: first light, then clothing, never both in the same run.
What multi-image editing cannot do
It is not precise local repair. Because mask-based inpainting is listed as not yet supported, you cannot mark one exact area and expect only that area to change. Any edit may affect the whole image, so check everything, not just the part you asked about.
It does not recover hidden detail. An edit creates new pixels from your instruction and references. It cannot restore what your camera never captured, such as the rest of a cropped shoulder or the real view behind you. Anything it adds is invented, so do not present it as a real place or moment.
It does not make every reference acceptable. Having a slot for a reference image is a technical feature. It does not give you the right to use another person's face, a public figure's likeness or a brand's logo.
How this compares with the Lumatch workflow
Lumatch uses the same idea of separate roles, through its own interface rather than MiniMax. Its studio starts from at least one clear photo of you as an adult and a separate scene reference you have the right to use, with an optional outfit photo. You then set directions such as composition, pose, hairstyle, outfit, retouch and aspect ratio, and review 1–4 candidates per run. The studio requires an account, and generation, reference analysis, visual checks and repairs use credits.
If you only want feedback on an existing photo, the free single-photo review looks at lighting, framing and expression. It does not generate images. Neither tool can promise a particular result or more matches.
Common questions
Is MiniMax H3 Image part of Lumatch?
No. It is an independent model from MiniMax. Lumatch does not use it and has no integration with it.
Do I need 16 reference images?
No. Up to 16 is the limit the README states, not a target. For a profile photo, one portrait plus a scene and perhaps an outfit image is usually easier to check.
Can I use it to fix only one small area?
Not precisely. The README lists mask-based inpainting as not yet supported, so an edit is not limited to a marked area.
Is it free to use commercially?
Do not assume so. The model has its own Community License Agreement. Read the current LICENSE and the terms of any service you use.
Related guides
- AI portrait editing prompts for natural photos: what to keep, what to change and what to check.
- Seedream 5.0 Pro US for profile photos: comparing an AI-edited selfie with the original.
- Fibo 2.0 Dev for portraits: what to check after editing a portrait with an AI model.
Sources
- MiniMax, MiniMax H3 Image README on GitHub (tasks, specifications, prompting tips and licence, read November 15, 2026): https://
github.com/ MiniMax- AI/ MiniMax- H3- Image/ blob/ main/ README.md - MiniMax, MiniMax-H3-Image Community License Agreement (linked for reference, read November 15, 2026): https://
github.com/ MiniMax- AI/ MiniMax- H3- Image/ blob/ main/ LICENSE