Image GenerationActive
MiniMax H3 Image: 4K Generate + Edit
MiniMax H3 Image by MiniMaxAI generates and edits still images up to 20MP. Combine up to 16 references and long prompts for precise layouts and text.
Overview
MiniMax H3 Image is an open-weights image model by MiniMaxAI for text-to-image and instruction-based image editing. It encodes your prompt and optional reference images into one sequence, then a diffusion transformer predicts an image latent that a VAE decodes into the final picture. This design makes it good at high-detail 4K work, multi-reference composition, and edits where most of the frame must stay stable.
What you can build
- Product shots from a written spec, including packaging and set dressing
- Character turnarounds and multi-panel character sheets from a single reference
- Concept art frames you can later animate with image-to-video pipelines
- Instruction-based photo edits like season changes, relighting, or object swaps
- Posters, UI mockups, and labeled diagrams where text placement matters
- Multi-reference composites like “person from image 1, outfit from image 2, room from image 3”
Inputs
- A single image brief, up to 32,000 characters. Write the whole scene: subject, setting, lighting, camera, style, and any exact text to render.
- Optional reference images to edit or combine. You can provide 1 to 16 images as uploads or URLs.
- For multi-reference work, cite images by order in the prompt (for example, “keep <Picture 1> the same, use the color palette from <Picture 2>”).
- An output shape choice. You can pick a fixed aspect ratio, or let it match the first reference image when you provide one.
- An output size choice. You can pick a tier like 1K, 2K, or 4K, or a custom pixel size when the host supports it.
- The number of images to return in one run. Wiro may cap this lower than some hosted APIs.
- An optional seed for repeatable results. Use the same seed and settings to get closer matches across runs.
- A preferred output file format: PNG, JPEG, or WebP.
Outputs
The model returns one or more generated images.
- Each image is delivered as a standard image file (PNG, JPEG, or WebP).
- Results typically include image dimensions and a downloadable file location.
- Some hosts also return helpful metadata such as the expanded prompt the model actually used and per-image safety flags.
Recommended settings
- Use 4K when you need small details, readable labels, or print-like density. MiniMaxAI reports best overall quality around 16 megapixels.
- Use common 16MP sizes for clean results: - 1:1: 4096×4096
- 16:9: 5472×3072
- 9:16: 3072×5472
- 4:3: 4736×3552
- 3:4: 3552×4736
- 21:9: 6272×2688
- Keep prompt expansion on when your prompt is short or sketchy. Turn it off when your prompt already reads like a full art brief.
- For edits, describe what must not change. Name the camera angle, identity, and key objects you want preserved.
Limitations
- Output size is capped by hosts. Common limits are up to 20,000,000 pixels per image, with aspect ratios between 1:3 and 3:1.
- Editing uses references, not paint masks. Mask-based inpainting is not supported on some hosted endpoints.
- Transparent backgrounds are not supported on some hosted endpoints.
- Very long paragraphs of small text are less reliable than short labels and headings. Keep on-image copy tight.
- Low-quality inputs can degrade results. Blurry references, heavy compression, screenshots with artifacts, and scanned images can cause drift or warped details.
- The open-weights release is distilled to a small fixed number of sampling steps. That limits low-level sampling control compared to research pipelines.
Safety & compliance
- Many hosts run a safety checker and may flag or restrict sexual content and other disallowed content.
- The MiniMax-H3-Image community license includes use restrictions. It prohibits using the model or its outputs to improve other AI models.
- The license also includes territory limits for the open-weights release. Review the license terms for your deployment location and distribution plans.
Example prompts
Great starting points for MiniMax-H3-Image.
Keep the street corner in <Picture 1> exactly as it is: the same buildings, the pale blue cafe with its striped awning and the "CAFE LUMEN" sign, the tables, the red bicycle by the lamppost and the same camera angle. Change only the season and the time of day: make it a snowy winter evening at blue hour. Fresh snow covers the rooftops, the awning, the tables, the window boxes and the cobblestones, with a trail of footprints along the pavement. Warm golden light glows from the cafe windows and from the lit lamppost, strings of small warm bulbs hang across the street, and light snow is falling. The sky is a deep twilight blue. A photorealistic photograph with a cozy winter atmosphere.Image Generation
A full-body fashion editorial photograph. The woman from <Picture 1>, with exactly the same face, short curly copper-red hair and freckles, wears the mustard-yellow corduroy jacket from <Picture 2> open over her white t-shirt, keeping the jacket's color, sherpa collar, chest pockets, buttons and stitching exactly as in <Picture 2>, together with her light blue jeans and white sneakers. She stands on the stone steps of the tiled hillside street from <Picture 3>, one hand in her jacket pocket and the other holding a paper coffee cup, smiling at the camera. Warm golden afternoon light from the side, natural shadows on the tiles, shallow depth of field, photorealistic 50mm photograph.Image Generation
A vintage screen-printed travel poster for a fictional seaside town. At the top, large cream-colored Art Deco lettering reads "PORT ALDERA". The center shows a white lighthouse with a red band on a rocky green headland at sunset, with three sailboats on a calm teal sea and seagulls in an orange and coral sky. At the bottom, a navy band holds the text "SUMMER FERRY | DAILY 09:00 - 18:00" in cream capitals, and below it in smaller letters "EST. 1924". Flat limited palette of teal, coral, cream and navy, visible halftone dots, slight ink misregistration and aged paper texture at the edges.Image Generation
An extreme close-up macro photograph of a blue-green dragonfly resting on a thin reed stem at dawn. Its large compound eyes show a fine honeycomb texture, its transparent wings reveal a dense network of dark veins with tiny water droplets clinging to them, and its slender abdomen has a metallic turquoise sheen. Dew drops along the reed reflect the sunrise. The background is a soft out-of-focus wash of pale gold and green from a misty pond. Shallow depth of field, soft natural side light, ultra-detailed 100mm macro photograph.Image Generation
API quick start
Run MiniMax-H3-Image with a single API call.
{
"prompt": "Keep the street corner in <Picture 1> exa…",
"inputImage": "https://your-cdn.com/input.png",
"aspectRatio": "auto",
"resolution": "1K"
}curl -X POST "https://api.wiro.ai/v1/Run/minimaxai/minimax-h3-image" \
-H "Content-Type: application/json" \
-H "x-api-key: YOUR_WIRO_API_KEY" \
--data-binary @- <<'JSON'
{
"prompt": "Keep the street corner in <Picture 1> exa…",
"inputImage": "https://your-cdn.com/input.png",
"aspectRatio": "auto",
"resolution": "1K"
}
JSONOther models you might like
Explore similar image generation models.
