MiniMax H3-Image is an open-weight image generation and editing model from the H3 family. Create from a text brief, work with up to 16 reference images, and choose a high-resolution canvas. Standard and Highspeed modes are available through MiniMax’s image API.
Start with the subject, viewpoint and materials. The farm-table example brings bread, cheese and vegetables into one overhead composition. Describe how objects relate, where the light comes from and what should remain quiet. Prompt enhancement can expand a short text brief; review the expanded description when a detail matters.
The podcast example begins with a real reference supplied in MiniMax’s workflow. The edited result removes the man’s headphones and changes the hosts’ pose into a fist bump. Use a specific instruction, then compare faces, hands and background details. A successful example does not guarantee exact preservation in every edit.
The service accepts one to sixteen reference images for editing. Here, ten inputs—an empty room and nine pieces of furniture—become one furnished interior. Give each reference a clear role and describe the intended placement. This is a ten-input demonstration, rather than a claim that every sixteen-image request will combine perfectly.
A grey lunar-base render establishes the arrangement of habitats, a rover and an astronaut. The paired result adds surface materials, lighting and a photographic treatment. This is image editing from a rendered reference, not the production of a new 3D asset. Check which shapes and spatial relationships survived the change.
A marked position on an illustrated map supplies a different kind of reference. MiniMax’s example turns that location into a first-person scene, with the source map retained as a small overlay. Make the viewing direction explicit. Inspect landmarks and geometry rather than treating the output as a measured reconstruction.
The official examples span photographic scenes, environment concepts and painted illustrations. This pale garden uses washes of color, paper-like space and reflected architecture. Name the medium and the features that define it. For a series, compare framing, palette and texture before committing to a visual direction.
Source: MiniMax official H3-Image examples and input/output panels, cropped for display.
| DIMENSION | MiniMax H3-Image | H3-Image Highspeed |
|---|---|---|
| Tasks | Text-to-image and image editing | Same API tasks and parameter limits |
| Reference inputs | 1–16 images for editing | Same reference-image limit |
| Output controls | Custom dimensions or aspect ratio and area level; up to 20 MP in the API | Same API size controls and limits |
| Positioning | Standard service mode | Speed-focused service mode; actual latency depends on the request |
| Open weights | MiniMax-H3-Image checkpoint is published | The reviewed model card does not list a separate Highspeed checkpoint |
The highlighted column identifies this page’s main model. This is a documented service-mode comparison, not a same-input speed or image-quality benchmark performed by Pexo.
Combine room, furniture or product references into a proposed setting. Check proportions, placement and object details before using the image in a presentation.
Turn a location brief, map or rough render into a visual concept. Use the result to explore atmosphere and composition, with human review of spatial consistency.
Explore a photographic, surreal or illustrated treatment for a story. Compare a small set of directions, then proofread visible text and inspect the final image at full size.
Take the wider visual idea into Pexo’s video creation workflows.
It is MiniMax’s open-weight image generation and editing model, built from the H3 family. The Chinese announcement is dated November 10, 2026; the English announcement is dated November 11. This page covers the image model and its Highspeed service mode, rather than the H3 video models or Pruna’s P-H3-Image release.
No. MiniMax uses a 4K area level of about 16 megapixels: 4096 × 4096 for a square, or 5472 × 3072 in its listed landscape example. The API also accepts custom sizes up to 20 megapixels, subject to its aspect-ratio limits. Inspect the actual returned dimensions instead of assuming every image is 3840 × 2160.
Both modes support the same documented API controls and limits. Highspeed is the speed-focused option. Test the kind of brief and reference set you actually use, then compare latency and the result. The provider’s under-ten-second headline is not a guaranteed completion time for every device, service request or output size.
Text-to-image uses enhancement by default; it can be disabled when you want to control the description directly. Image edits always apply enhancement in the documented API. Review the returned revised prompt when diagnosing an unexpected detail. The local open-weight workflow has its own optional prompt-expansion component.
The reviewed MiniMax API does not support transparent backgrounds or masked edits. Open weights are available under the MiniMax-H3-Image Community License, with separate text-encoder requirements. Check the license and hardware requirements for your intended use; an open-weight release does not imply an unrestricted license or a lightweight installation.
