Native 4K
Generates 4096×4096 directly, no upscaling, with detail that holds up when you crop.
MiniMax · Image model
MiniMax's new image model: native 4K, up to 16 reference images in one result, and sharper text and detail.
The image version of MiniMax H3 (not the H3 video model), released November 10. Write a prompt, or edit and combine up to 16 reference images, up to 4K.
Opens the generator with MiniMax H3 Image selected. Failed runs are refunded.
3D game art: knight duel
Each image was made with MiniMax H3 from the prompt shown. Like one? Click "Use this prompt" and make your own.
MiniMax H3 · 2K · 3:2
A narrow cobbled street at blue hour just after rain, a cyclist in a yellow rain jacket riding in from the right, warm shop-window light reflected on the wet stones, low camera, 24mm lens, shallow depth of field
Generates 4096×4096 directly, no upscaling, with detail that holds up when you crop.
MiniMax says a 4K image takes under 10 seconds on its own API; here it usually takes 15 to 25 seconds, 4K included.
Give one image each for the person, outfit, product and setting, say "the person from Image 1, the coat from Image 2," and get one consistent scene.
Recolor, add or remove objects, or swap the background. Say what changes and what stays, and the rest holds still.
Headlines, menus and infographic labels come out as written when you put them in quotes.
MiniMax says details look natural, without the overly smooth look of many AI images.
MiniMax says it has the best prompt adherence among open-weight models, so a full brief or product spec can go in as written.
No subscription: each image uses credits, and failed runs are refunded automatically.
Cards 2, 6 and 7 are based on MiniMax's launch post (November 2026); 4K and 16 references are this site's limits.
| Option | Details |
|---|---|
| Mode | Text to image, image edit and multi-image composition |
| Resolution | 1K, 2K, 4K |
| Aspect ratio | Auto (matches Image 1), 1:1, 4:5, 3:4, 2:3, 9:16, 1:3, 5:4, 4:3, 3:2, 16:9, 21:9, 3:1, or a custom width and height |
| Reference images | Up to 16, called Image 1, Image 2… in order |
| Images per run | 1 to 10 |
| Format | JPG, PNG or WebP |
| Compression | Adjustable for JPG and WebP, default 95 |
| Prompt expansion | On by default, can be turned off; always on with references |
| Keep editing | Click "Keep editing" on a result to use it as the reference |
| Credits per image | From 3, up to 25; multiplied by the number of images |
Images marked "Example" were made by us with MiniMax H3. Click any image to zoom in.
Lay out the scene, characters and depth in one go.
How to start: Say where the subject stands first, then the distance and the light.
Concept art for a fantasy exploration game: a traveler in an ochre cloak on a cliff ledge in the lower left, a canyon city carved into sandstone in the distance with rope bridges and hanging gardens between towers, morning sun from the right, haze in the distance
Turn a person, an outfit and a product from separate images into one new scene.
How to start: Upload them in order and write "the person from Image 1, the product from Image 2."
Put the person from Image 2 and the perfume bottle from Image 1 in one scene: the person sits by a seaside café window holding the bottle, keep the bottle's look and label unchanged, afternoon daylight
Headlines, prices and step labels come out as you wrote them.
How to start: Put the exact words in quotes and say where they go and how big.
Overhead shot of an open two-page café menu on white marble, header "HALF MOON CAFÉ", left page "MORNING": Oat latte 4.50, Cinnamon bun 3.80, Shakshuka 12.00; right page "AFTERNOON": Iced hojicha 5.00, Tomato toast 9.00
Keep the product's shape and label and change the background.
How to start: Upload the product shot and say what must not change and what the new background is.
Keep the perfume bottle from Image 1 exactly the same, including its shape, label and text. Replace the background with a bathroom counter in soft morning side light, with a white towel nearby
Lay out parts in an exploded view, or get a character's three views at once.
How to start: Say how the parts line up and what each label says.
Isometric exploded view of a mechanical wristwatch, parts floating along a vertical axis from top to bottom: crystal, bezel, dial, hands, movement, case, strap, off-white background, thin black leader lines to small labels
Make a cinematic still first, then turn it into video.
How to start: Say where the camera is and which third of the frame the subject is in.
A cinematic still to use as a video's first frame: a coastal road at dawn, a red vintage convertible in the right third of the frame, light mist over the sea, warm backlight
Want it to move? Try our Image to GIF tool.
Use this promptClick "Generate with MiniMax H3"; the model is already selected. To edit, upload up to 16 reference images.
Describe what you want: subject, scene and light. For edits, say what changes and what stays. Pick the ratio, resolution and number of images.
Download the result, or click "Keep editing" to change it further.
Start with the goal and who is doing what, then add style in visible terms like "soft side light" or "brushed metal," not just "premium."
Lay out the frame first: how many subjects, where, from what angle and how much empty space. Then add details.
Put on-image words in quotes with position, size and line breaks. For charts, give real numbers and units.
Give each reference a job: "the product from Image 1, colors from Image 2, layout from Image 3."
When editing, say what changes and what stays: "Make the cup on the left blue, keeping its shape, material and position."
For grids, write "2 rows and 3 columns" rather than "2x3," which can flip.
| Feature | MiniMax H3 Image | Seedream 5.5 | GPT Image 2.6 | Nano Banana 2.1 |
|---|---|---|---|---|
| Max resolution | 4K | 2K | 4K | 4K |
| References | Up to 16 | Up to 10 | Up to 4 | Up to 1 |
| Per run | 1–10 | 1 | 1 | 1–4 |
| Transparent background | No | Yes | No | No |
| Best at | 4K, multi-image edits, concept art | Multi-round edits, transparency | In-image text, GIFs, SVGs | Cheap and quick |
| Credits | From 3 | From 3 | From 5 | From 2 |
All of these are in the same generator. MiniMax says H3 Image costs 50–70% of comparable mainstream models; run one prompt on each and compare.
Want cheap quick tries, or xAI's Grok? See the Grok image generator.
Pay per image with credits, no subscription: from 3 at 1K, 4 at 2K and 9 at 4K, a little more with extra references, multiplied by the number of images. The button shows the exact amount before you generate. Packs start at ¥9.9 and never expire; failed runs are refunded automatically.
Using MiniMax's API means your own account, key and token billing. Here you open the page and pay per image.
See credit packsMiniMax released it on November 10, 2026, as the image version of H3. This site uses it through a third-party API and isn't affiliated with MiniMax.
Yes. H3 started as a video model; MiniMax H3 Image, released November 10, makes and edits still images. This page is about the image version.
MiniMax says H3 Image is also on Hailuo AI. Here you use it in the browser with no app, paying per image.
Write a prompt alone for text to image, or upload 1 to 16 reference images to edit and combine them. Refer to them as Image 1, Image 2 in your prompt.
1K, 2K or 4K; 4K is 4096×4096. Aspect ratios run from 1:3 to 3:1.
Your first one can be free: new accounts get 4 credits, enough for one 1K image.
From 3 at 1K, 4 at 2K and 9 at 4K, a little more with references, times the number of images. The button shows the exact amount.
No key needed here; just sign in and generate. If you want the API itself, MiniMax offers it on its developer platform.
MiniMax released the weights on Hugging Face. They're large and need a powerful GPU; if you'd rather not set that up, use it here in the browser.
MiniMax also offers a faster Highspeed version. This site uses the standard version, which MiniMax says is 1.5–2× as fast as mainstream models.
Your credits are refunded automatically.
Yes. Under the terms of MiniMax and our API provider, the images you generate are yours and can be used commercially. Make sure you have the rights to the prompts and reference images you use, you're responsible for content involving other people's likeness or trademarks, and follow the laws where you live.
Start with a sentence, or combine a few reference images. Need it big? Pick 4K.
From 3 credits per image. Failed runs are refunded.