Cinematic game key art
A lone rider pauses beside a desert waystation as a wall of dust approaches from the horizon. Worn leather gear, warm late-afternoon light, distant mesas, strong readable silhouette, cinematic game key art.
Open-weight image generation · Text to image · Reference editing
Create a new image from a detailed brief, or combine visual references to guide the subject, composition, style and spatial details. MiniMax H3 Image supports text-to-image generation, image editing, output up to 4K and as many as 16 reference images.
Create images from text or edit with up to nine reference images. Sign in to use your credits.
Example image preview
Text to image · Reference editing

Illustrative image for composition and lighting inspiration.
Review the official H3-Image inputs, sizing methods, batch controls and output formats before submitting a task.
promptRequiredUp to 32,000 characters describing the complete imagesizeOptionalauto, or custom WIDTHxHEIGHT within the supported pixel-area and ratio limitsaspect_ratioOptionalA ratio from 1:3 to 3:1 when size is autoareaOptional1K, 2K or 4K; the official API defaults to 4KnOptionalGenerate 1 to 10 images; default 1output_formatOptionaljpeg by default, or png / webpdisable_prompt_enhancedOptionalPrompt enhancement is enabled by default for text-to-imagepromptRequiredThe intended result, including what changes and what remains fixedimagesRequired1 to 16 reference images; upload order determines the Picture numbersizeOptionalauto, or custom WIDTHxHEIGHT within the supported limitsaspect_ratioOptionalA ratio from 1:3 to 3:1 when size is autoareaOptional1K, 2K or 4K output-area levelnOptionalGenerate 1 to 10 edited images; default 1output_formatOptionaljpeg by default, or png / webpAvatars, covers and balanced product compositions
Feed posts and editorial portraits
Character cards and print-style layouts
Photography and poster compositions
Stories, Shorts and mobile creative
Mobile hero images and tall displays
The tallest supported canvas ratio
Key art, web headers and presentation frames
Classic frames and technical boards
Camera-style compositions
Product and large-format layouts
Blog headers and campaign banners
The widest supported canvas ratio
Cinematic concept art and title beds
Any integer ratio from 1:3 to 3:1 within API limits
Target area comparable to 1024 × 1024
Target area comparable to 2048 × 2048
Target area comparable to 4096 × 4096
Compact delivery for photographic images
Lossless output for graphics and hard edges
Efficient delivery for modern web experiences
Prompt examples
Use these structures for game key art, character sheets, product concepts and spatial image edits.
A lone rider pauses beside a desert waystation as a wall of dust approaches from the horizon. Worn leather gear, warm late-afternoon light, distant mesas, strong readable silhouette, cinematic game key art.
Create a character sheet for the person in <Picture 1>. Show front, side and rear views plus three facial expressions. Keep the same proportions, costume materials and color palette across every panel.
Create an exploded presentation view of the product in <Picture 1>. Separate the major parts along their assembly axes, keep the materials and proportions accurate, group small fasteners beside the correct component, clean dark studio background.
Move the chair in <Picture 1> between the two mirrors. Keep the camera position, room structure and lighting unchanged. Update both mirror reflections so they match the new chair position.
The open-weight model
MiniMax H3 Image is an open-weight model for generating and editing still images. Start from text alone, or bring reference images for a subject, product, visual style, layout or environment.
The model is designed to follow layered instructions about objects, camera position, composition, lighting and visual relationships. That makes it useful for character development, game worlds, product concepts, spatial scenes and technical illustrations—not only isolated portraits.
The official API supports 1K, 2K and 4K output-area settings, prompts up to 32,000 characters, 1 to 16 reference images and batches of 1 to 10 images. Review every result for identity, text, reflections, object relationships and other details that matter to the final asset.
Image gallery






Official model capabilities
Create from text, combine references, edit spatial relationships and request several visual directions from one structured brief.
| What you want | Supported | What you provide |
|---|---|---|
| Generate from text | Yes | A complete brief describing the subject, composition, environment, light and intended visual style |
| Combine visual references | Yes | Up to 16 images, with a clear role assigned to each subject, product, style or layout reference |
| Edit part of a scene | Yes | The requested change plus explicit instructions for what must remain fixed |
| Build character and concept sheets | Yes | Reference images plus directions for viewpoints, expressions, materials and panel layout |
| Create product and technical visuals | Yes | A source product image plus instructions for cutaways, exploded views or labeled compositions |
| Generate several directions | Yes | A prompt, output settings and a batch size from 1 to 10 images through the official API |
From a prompt alone
Turn a layered visual brief into one or more images at 1K, 2K or 4K, with automatic, ratio-based or custom sizing.
Describe the subject, environment, camera angle, layout, lighting, materials and intended use. H3 Image is designed to carry these connected instructions into one composition.
Choose an output-area level or provide a custom width and height. Use smaller outputs to explore a direction, then move to 4K when the composition is ready for detailed review.
Use automatic sizing, a ratio from 1:3 to 3:1, or supported custom dimensions. The API reports the actual width and height returned for every image.
Request one focused result or a batch of alternatives. Compare the outputs for composition and detail instead of treating batch variations as guaranteed matches.
Multi-reference editing
Blend up to sixteen official-API references and describe which subjects, products, layouts, materials and scene details should change or remain fixed.
Use one image for the subject, another for the product, another for the palette and another for the layout. Tell the model what each reference contributes to the result.
Name the identity cues, proportions, materials, accessories, colors or camera decisions that must remain recognizable after an edit.
Move or replace objects while describing how perspective, shadows and reflections should respond. Also state which parts of the room or composition must stay fixed.
Turn a source into game art, animation, a technical illustration or another clearly named medium while retaining the selected subject and structural details.
Five steps
Choose a workflow, assign references, write the complete brief, set the output and review every result.
Begin with a written brief when you want a new composition. Choose image editing when a subject, product, style, layout or environment should come from reference images.
Upload up to sixteen references through the official workflow and identify what each one controls. For example: use Picture 1 for the subject, Picture 2 for the outfit and Picture 3 for the layout.
Describe the intended result, object relationships, camera or layout, lighting, materials and constraints. For edits, state both what should change and what must remain fixed.
Choose 1K, 2K or 4K, a ratio from 1:3 to 3:1, JPEG / PNG / WebP and one to ten images. Use only one sizing method when calling the official API.
Inspect identity, object placement, reflections, hands, labels and unintended changes. Revise one instruction at a time and retain the version that meets the brief.
Creative and production workflows
Use H3 Image for game art, character development, product concepts, technical illustration, spatial editing and reviewed image-to-video first frames.
| Use case | Endpoint | What it produces |
|---|---|---|
| Game key art | Text to image or references | Characters, environments and atmosphere combined into launch art or store banners |
| Character design | Reference editing | Turnarounds, expressions, costumes and alternate scenes built around one character direction |
| Product concepts | Reference editing | Studio scenes, material studies, exploded compositions and presentation views |
| Technical illustration | Text to image or references | Multi-panel explanations, cutaways, labeled diagrams and visual research boards |
| Spatial scene editing | Reference editing | Repositioned objects with instructions for perspective, shadows and reflections |
| Image-to-video preparation | Text to image or references | Reviewed first frames with settled subject, styling and composition |
Prompt craft
Structure the brief around composition, reference roles, object relationships, fixed details and short readable text.
Name the subject, environment, camera or layout, light, materials and intended visual style so the model has a complete scene to organize.
Explain which picture controls the subject, product, palette, style or layout instead of asking the model to infer why each image was included.
State what is in front, behind, inside, reflected by or attached to something else when spatial accuracy matters.
For edits, preserve specific identity cues, proportions, camera position, architecture, materials or colors instead of only describing the new element.
Use concise titles and labels inside the image. Add long paragraphs later in a design tool, where typography can be checked and edited directly.
| What you get | Why | Fix |
|---|---|---|
| The composition misses part of the brief | The prompt lists objects but not their relationships | Describe placement, scale, viewpoint and which details have priority |
| The wrong reference controls the result | The references were not assigned clear roles | Identify Picture 1, Picture 2 and each additional reference explicitly |
| Unrequested parts of the image changed | The edit describes only the new element | State which subject, camera, layout, colors and materials must remain fixed |
| Reflections or shadows are inconsistent | The prompt changes an object without updating its spatial effects | Describe how the new position should affect reflections, contact shadows and lighting |
| Text is misspelled or too dense | The image contains more copy than can be reliably reviewed | Keep generated text short and add longer copy in a design tool afterward |
Same account, different job
The difference is what you hand over and what you get back. Choose a still for the look, or a clip for motion and sound.
| Workflow | Image editing | Video generation |
|---|---|---|
| What you supply | A prompt and optionally 1–16 reference images | A prompt, first frame or multimodal reference pack |
| What comes back | 1–10 still images through the official API | A timed clip with motion and audio |
| Output | 1K, 2K or 4K area levels | Resolution and duration depend on the selected video model |
| Best for | Concepts, key art, characters, products and reviewed first frames | Motion, camera movement, performance and sound |
| Recommended order | Settle the visual direction first | Animate the approved still after the composition is ready |
A common order of work is to settle the look first, including outfit, styling and location, then take the approved still into the video generator as a first frame.
Frequently asked questions
Official specifications for 4K output, sixteen references, batch generation, sizing, prompt enhancement, formats and open weights.
MiniMax H3 Image is an open-weight model for text-to-image generation, reference-based creation and image editing. It is designed to follow detailed creative instructions involving subjects, objects, composition, lighting and spatial relationships.
The official API supports 1K, 2K and 4K output-area settings. It also accepts supported custom width and height values when the total pixel area and aspect ratio remain within the documented limits.
The official image-editing API accepts 1 to 16 reference images. Upload order determines the Picture number, so prompts can assign a clear role to each reference.
Yes. The official API supports batches of 1 to 10 images per request. Individual results can still differ, so review every output rather than assuming the batch is internally identical.
When using automatic sizing, the official API accepts integer X:Y ratios from 1:3 to 3:1. You can also provide supported custom width and height values instead of an aspect ratio and area level.
The official API accepts prompts up to 32,000 characters. A longer limit is useful for structured briefs, but the prompt should still make priorities, object relationships and constraints easy to identify.
Text-to-image prompt enhancement is enabled by default and can be disabled. Image editing always applies prompt enhancement. The official response can include the revised prompt used for each image.
The official API returns JPEG by default and also supports PNG and WebP. Choose the format according to the final workflow rather than expecting the format alone to improve image detail.
MiniMax describes H3 Image as an open-weight model. Developers should review the official model license and usage restrictions before local deployment, redistribution or commercial integration.
Yes. After reviewing the still for subject, composition and detail, use it as a first frame in an image-to-video workflow and describe the intended subject and camera movement.
Get started
Start from a written idea or bring visual references. Build the composition, review the details and refine the image until it fits your project.