MiniMax H3 Image Review: Text, Editing and Realism Tested
Hands-on tests of MiniMax H3 Image: text, exact counts, identity-preserving edits, illustration-to-photo and anatomy. Where it succeeds and where it slips.

How good is MiniMax H3 Image at text, counting, editing and realism?
MiniMax H3 Image is strong at big headline text, exact object counts, scene-level edits and photorealistic detail. It is weaker at small text, short painted names, fine mechanical parts and human anatomy in action poses. In a set of prompt-by-prompt tests run locally, it got the composition right most of the time and missed on small details. Treat it as a capable first-draft tool that still needs a human check.
TL;DR
- Large text works, small text does not. A big neon sign and a poster headline came out clean, while the small line on the poster stayed garbled and faint.
- Counting and layout held up. A prompt for exactly five boats with one fisherman each, lined up diagonally in different colored outfits, was followed correctly.
- Edits keep the scene intact. Swapping a vintage taxi for a tuk tuk and changing a white t-shirt to a poncho left the surroundings and the person's face, curly hair and red lipstick very close to the original.
- Illustration-to-photo was the standout. A flat illustration became a convincing photograph with the same layout, horizon and lighting.
- Anatomy and fine mechanics still slip. A fire-juggler scene had problems, and the over-and-under boat shot was missing its boat details.
- It is heavy to run locally. The tester reported close to 48 GB of VRAM for the setup shown.
What is MiniMax H3 Image?
MiniMax H3 Image is an open-weight image model. "Open-weight" means the trained model files are published, so you can download them and run them on your own hardware instead of only using a hosted service. According to the tester, it is a 33 billion parameter diffusion transformer, distilled down to six sampling steps with no CFG (a setting many diffusion models use to follow prompts more tightly). It generates natively at around 16 megapixels without an upscaler, and it accepts up to 16 reference images, which is what makes editing possible.
In the test setup it ran locally in ComfyUI, a node-based interface for image models, and used close to 48 GB of VRAM (the memory on a graphics card). The tester said you can lower that with fewer steps, but recommended the settings used. Generation took roughly 10 to 15 seconds per image, and one identity-preserving edit finished within about 40 seconds.
One caveat applies to everything below. These are one person's tests, mostly single runs per prompt, judged by eye. They show patterns, not a benchmark.
How well does it render text in images?
It depends on the size of the text.
The first test was a single prompt for a wet city street with a large neon sign, a poster headline, a dog and a bit of a crowd. The neon sign and the poster headline came out clean. The wet street, the dog and the crowd looked natural. The weak spots were the small line of text on the poster, which was still garbled and faint, and the faces in the crowd.
The same pattern showed up in other tests:
- Painted name on a boat. The prompt asked for one short painted name on the foreground boat. It did not appear at all.
- Handwritten city names on a palm. A hand test asked for small handwritten city names and a Michigan mitten outline drawn in marker. The four city names were spelled correctly as far as the tester could tell. The mitten outline was faint and mostly followed the edge of the hand.
- Text after an edit. When the model turned a taxi into a tuk tuk, the requested wording landed on the roof panel instead of the side of the vehicle.
The practical rule: big, prominent text is reliable enough to try. Small print, short text in a specific location, and text inside a busy scene need checking every time.
Can it follow exact counts and layouts?
Mostly yes. One test asked for exactly five boats, each with one fisherman, lined up in a diagonal, each in a different colored outfit, on a misty lake at sunrise in Myanmar's Shan State. The count and layout came out right. Five boats and five fishermen shrank toward the sunrise, and the outfit colors, stilt houses, pagoda and floating gardens were all in place.
The misses were small. The painted name was absent, and the reflections in the water looked a bit artificial, especially the first one. Where foliage met the water, the tester found it looked quite natural.
If you make images for presentations or product mockups, counting is a common failure for image models. A correct five-out-of-five here is a meaningful result, though one run does not prove it will happen every time.
How good are the image edits?
Editing is where the reference-image support matters. Four edit tests stand out.
| Edit test | What worked | What missed |
|---|---|---|
| Taxi to tuk tuk | Candy pink and lime green paint, string lights, wet street reflections, background building and pedestrians all kept | Wording ended up on the roof panel, not the side |
| White t-shirt to woven poncho, background to Machu Picchu | Poncho pattern and color, misty terraces, warm light; face, curly hair and red lipstick very close to original | Background looked a bit plasticky from the sun reflections |
| Drone photo: add five gates, switch to spring, change the man's vest | Gates receded in believable perspective, his raised-arms pose was kept, blossoms, petals, stone lantern and mist came through | Shadow disappeared and the sun angle changed |
| Flat illustration to photograph | Layout, horizon height and lighting carried over; rounded shapes became white Mongolian gers with orange doors; storm sky, rain and a rider in a teal coat looked convincing | Nothing major noted |
The tester called the drone photo the hardest edit, since it required following lens distortion and perspective across a wide-angle shot. The model handled the geometry well. The illustration-to-photo result drew the strongest reaction, including praise for the rain droplets.
Face preservation is the one to watch. Keeping a face, hair and lipstick close to the original while changing clothes and background is what people need for marketing and profile images. The tester noted that likeness is subjective, so check any result against your own source image.
How realistic are the photos?
Realism is a strength in the still, detailed subjects.
- Portrait. The prompt asked for natural skin texture and pores, tiny individual glass beads in geometric patterns, soft golden hour light and a blurred hill background. Skin showed real texture rather than plastic smoothing, each bead was separate, and the rim light and background blur looked like a real lens.
- Ice and light. An Iceland prompt asked for see-through ice with trapped bubbles and cracks, aurora colors bending through it, and a mirror-like wet beach under a long-exposure night sky. The ice was a standout, with believable translucency and green light glowing inside the chunks. The tester doubted the aurora looked fully real and found the overall scene a bit stylized, though dramatic. The star field and a tiny photographer with a red headlamp for scale worked.
- Hand. A hand with exactly five digits, creases and skin texture passed the check.
Where does it fail?
The failures follow a pattern: fine details inside complex scenes.
- Fire juggler. A motion test asked for a dense, frozen, sharp scene with midair glowing fire trails and sparks. The tester said human anatomy, especially hands and feet, is not right every time, and this was a case of that. The elephant looked okay, but the tree behind did not, and the model missed several things.
- Over-and-under shot. A split shot at the waterline in West Papua tested whether the same boat, paddle and light stayed consistent above and below the surface. The waterline was clear and the reef and camera angle were decent. The boat was hard to find, the outrigger was missing, and the school of fish was not vibrant.
- Small text and crowd faces. Covered above.
If your work involves people in motion, mechanical objects with many parts, or physics that must match across a split image, expect to generate several versions and pick the best.
What does this mean for you?
If you make marketing or internal visuals: use a model like this for scenes, backgrounds, mockups and edits where the overall composition matters. Add small text, logos and product labels afterward in a normal design tool. Do not trust any AI image for fine print.
If you edit photos of people: identity-preserving edits looked promising, but always compare to the original and get consent before changing someone's image.
If you are weighing local versus hosted: about 48 GB of VRAM in the tested setup is workstation territory, not a typical work laptop. Most readers will reach a model like this through a hosted service, if at all.
If you evaluate AI tools for your team: copy the method. Write a short set of prompts that test your real needs (a count, a text line, an edit of your own photo), run each one, and score it. Ranking by demo reels is how teams end up disappointed.
A simple scoring template:
Prompt: [what you asked for]
Must-haves: [count / text / layout / identity]
Result: pass / partial / fail
What slipped: [small text, hands, background, etc.]
Fixable by a human in under 5 minutes? yes / no
Could you build your own image-model scorecard?
If your team compares image tools, a shared spreadsheet gets messy fast: prompts in one tab, screenshots in another, scores nobody updates. You could describe a small internal app to Remy in plain English: a form to log each prompt, attach the result, score it against your must-haves and note what slipped, plus a page that ranks tools by pass rate. Remy builds and deploys full-stack apps from that kind of description, with a database and sign-in included, and its model catalog includes image models, so the same app could also generate test images.
Frequently Asked Questions
Is MiniMax H3 Image open source?
The tester describes it as an open-weight model, meaning the model files can be downloaded and run on your own machine. Open-weight is not always the same as fully open source, so check the license on the model card before using it commercially.
How much hardware does MiniMax H3 Image need?
In the tested setup, running locally in ComfyUI used close to 48 GB of VRAM. The tester said using fewer sampling steps can lower that. Generation took about 10 to 15 seconds per image, and edits took longer, around 40 seconds for one example.
Can MiniMax H3 Image edit photos of real people?
It accepts reference images and, in one test, kept a woman's face, curly hair and red lipstick very close to the original while changing her outfit and background. Likeness is subjective, so compare against the source. Only edit images of people you have permission to edit.
Does it get text right?
Large, prominent text such as a neon sign and a poster headline came out clean. Small text was garbled or missing in several tests, including a poster's fine print and a short painted name on a boat. Plan to add small text yourself afterward.
Useful for
How we source, write, and correct our articles. Editorial standards
Related articles

How Does ReShoot 4D Change a Video's Camera Angle With AI?

Tencent's Open World Model: Can an RTX 4090 Run It?

How Much Does an AI Short Film Cost? The $102,000 Breakdown
