Create and edit still images with MiniMax's new image model, in Standard or the faster Highspeed version. Start from a written brief or up to 16 reference images, combine people, products, and scenes, and run it in your browser with no GPU or local install.
Editing and references
Upload one photo to change a detail, or several to merge them into one picture. The same model handles both, so a draft can move from first idea to final version without switching tools.
One photo
Upload a photo and describe the change, like a new coat colour, a different time of day, or a prop taken out. The rest of the picture stays as close to the original as the model can keep it. Look at the face, hands, and edges before you build on the result.
Several references
Give each reference one job: a room as the setting, a separate photo for each piece of furniture, a palette to follow. MiniMax H3 Image arranges them in one coherent frame. Check the count, scale, and overlaps afterwards, since an item can come out smaller than expected or go missing.
Product and style
Pair a plain product shot with an image of the light and colour you want. The result keeps the product and borrows the mood for a campaign still. Compare the label text and the shape with your original before you publish.
Write plain sentences about the subject, the setting, the light, and the style. When you add references, call them Image 1, Image 2, and so on in the order you uploaded them, and say what each one contributes.
Number the references
Write “Image 1 is the room, Image 2 is the armchair” so every file has a single role in the result.
Say what must stay
In an edit, list what should not move: the face, the pose, the background, the product's shape.
Describe light and framing
For text-only prompts, add the camera distance, the light, and the mood. MiniMax expands short prompts on its own, so the details you name decide what it keeps.
Counting on a mask or seed
There is no mask, seed, or negative prompt. Name the area in words, for example “only the coat”.
Two references with one job
If two images both look like the subject, the model has to guess which one you meant. Drop one, or say which part comes from each.
Many changes in one edit
Changing the outfit, the pose, and the room at once makes a miss hard to trace. Change one thing, check it, then continue.
Example edit prompt: “In Image 1, change the trench coat to cherry red. Keep her face, hair, smile, pose, and the seaside steps behind her.”
A written brief is enough to start the visuals a project needs. Each task below notes what to review before the image goes out.
Describe the product, the surface, and the light to get campaign stills and packaging mockups. Check the label text and proportions before you use them.
Ask for front, side, and back views of one character to plan a costume, a mascot, or a story lead. Compare the panels, because small details can shift between views.
Upload a rough map or layout and describe where the viewer stands to explore a game or story location. Use it for visual exploration, not as an exact reconstruction.
Short titles, prices, and tags can sit inside the design itself. Proofread every word, and keep long copy for your layout tool.
Sizes and frames
Pick 1K, 2K, or 4K before you generate. The tier sets how many pixels you get; the exact width and height follow the aspect ratio you choose.
Resolution
Explore ideas at 1K or 2K, then switch to 4K for the version you will crop, print, or zoom into. At 4K a square image is 4096×4096 pixels and a 16:9 frame is 5120×2880.
Aspect ratio
Fixed ratios run from 9:21 to 21:9. Auto makes a square for text-only prompts. For edits, auto follows the first reference image as long as it sits between 1:3 and 3:1; for a wider panorama, pick a fixed ratio instead.
Both versions take the same prompt, up to 16 reference images, 1K–4K sizes, and the same ratios and formats; they differ in speed and credits. Switch between them in the model menu of the generator at the top of this page.
| Specification | Standard | Highspeed |
|---|---|---|
| 4K text to image, median | 16.0 sMiniMax P50 timing | 10.3 sMiniMax P50 timing |
| 4K edit, 3 references, median | 21.4 sMiniMax P50 timing | 14.4 sMiniMax P50 timing |
| Credits to generate | 6 / 7 / 161K / 2K / 4K | 9 / 10 / 301K / 2K / 4K |
| Credits to edit, 1 reference | 9 / 10 / 201K / 2K / 4K | 13 / 14 / 351K / 2K / 4K |
| Each extra reference | +2 credits | +3 credits |
| Choose it for | Fewer credits per image | Shorter waits per run |
Speed
MiniMax measured median (P50) times on its own platform: Highspeed takes roughly a quarter to a third less time than Standard at every size. On Oeneo AI the wait also includes the queue, so read these as a guide rather than a promise.
Text to image
Highspeed returns a 4K image in a median 10.3 seconds against 16.0 for Standard; at 1K and 2K both stay close to ten seconds or under. The same chart lists other image models from an earlier measurement round, so compare those bars loosely.
Editing
Each extra reference adds a little time to both versions. With three references at 4K, the median is 14.4 seconds on Highspeed and 21.4 on Standard.
Photoreal shots, painted styles, a puzzle board written out in full, and edits from one or ten references, all made by MiniMax H3 Image and shown here in square frames.
Tailoring, fabric texture, and studio light read like a real fashion shoot.
An impossible brief, shot with the lighting and grain of a magazine photo.
A concept-art board with its title lettered into the layout.
A gongbi painting with its calligraphy and red seal set into the composition.
A logic-puzzle board with its clues, stamps, and answer table written out.
One untextured render turned into a lit, photoreal frame with the same layout.
A red arrow on the map sets where the player stands; the map returns as a minimap.
An empty room and nine separate pieces of furniture, composed into one living room.
Each model has its own page and runs in the same workbench on one credit balance.
OpenAIAlso takes up to 16 reference images, with quality levels for cost and detail
View model
GoogleGoogle's editor for up to 14 references, at 1K–4K
View model
New
ByteDanceMulti-round edits, transparent assets, clearer text
View model
NewOpenAIOpenAI's newest: Flare for fast drafts, Sunburst for steady edits
View modelRemove the background, upscale, vectorize or denoise any result, all on the same credit balance.
Transparent PNG or WebP with clean hair and product edges, ready for a listing or a new composition.
2× / 4× / 8× with recovered texture instead of smeared pixels, so a 1K draft can become a print.
Logos, icons and flat illustrations become editable SVG that scales to any size.
Remove grain and compression artifacts from low-light photos and scans while keeping real detail.
MiniMax H3 Image is MiniMax's still-image model for text-to-image generation and image editing, released on November 11, 2026. One model takes a written prompt or up to 16 reference images. MiniMax also publishes the weights under its own community license. Oeneo AI runs both the Standard and the Highspeed version online.
No. Both belong to the MiniMax H3 family, but H3 Image makes still pictures and the video models make clips. To animate an image you made here, download it and use it as the first frame in the Oeneo AI video generator.
No, each run uses credits. New accounts get 10 free credits when they sign up. The price depends on the version, the resolution, and how many reference images you add, and the Generate button shows the exact amount before you start.
Type a prompt in the box at the top of this page, or add reference images first. Pick Standard or Highspeed, the resolution, aspect ratio, and format, check the credits on the Generate button, then generate and download the result.
Up to 16. The upload order sets the numbers: the first file is Image 1, the second is Image 2. Every image needs 128 to 8192 pixels on each side.
No. PNG results keep a solid background. To cut out the subject, run the image through the Oeneo AI Background Remover.
Yes, for short text such as titles, prices, labels, and signs. Put the exact words in quotes in your prompt. Longer paragraphs can come out with mistakes, so proofread every word before you publish.
No. On Oeneo AI the model takes a prompt, optional reference images, and the size, ratio, and format settings. Describe the area you want changed in words, and generate again for another take.
No. It runs online here, so nothing is installed on your computer. MiniMax also offers open weights for local setups such as ComfyUI, which need a large download and a capable GPU. Node settings from the older H3 video workflows do not carry over to this image model.
Highspeed is a separately hosted version of the same model that returns results faster for more credits; it is not a quality setting. Both take the same inputs, sizes, and formats, and the resolution you pick decides the output size in either. Choose one in the model menu above the prompt box.
Write a brief or upload your references, check the credits on the button, and generate.
