ChatGPT
Perplexity
Claude
Gemini
Grok
What Is MiniMax H3-Image? Generation, Editing, and Video Handoff
MiniMax H3-Image is an open-weight model for still-image generation and editing; Pexo is a separate option when the finished deliverable is an image-led video. The choice depends on whether you need a generated picture, a reference-guided revision, or a complete story with sound. This guide separates the new image model from the broader MiniMax H3 family and explains what to check before putting an output into production.
What changed with H3-Image?
MiniMax announced H3-Image on November 11, 2026, describing text-to-image, reference-based generation, and editing. Its release post reports a 4096 × 4096 image in under 10 seconds and support for up to 16 references. Treat the speed as a vendor claim, not a timing result from our own test. The announcement lists MiniMax API Platform, MiniMax Code, and Hailuo AI as access routes. MiniMax announcement.
The useful distinction is between a model name and the job assigned to it. An image model creates or revises a still. A production workflow decides where that still belongs, how it changes over time, and what narration should say. A compelling photograph of a product does not establish how the product works. Keep those decisions separate when preparing a campaign.
| Route | Starting point | Intended deliverable | What to check |
|---|---|---|---|
| Pexo | Brief or approved images | Finished video | Story, visual accuracy, sound |
| H3-Image generation | Written image brief | Still image | Subject and composition |
| H3-Image editing | Images plus instruction | Revised image | Preserved details |
| FLUX | Selected Black Forest Labs model | Model-specific output | Exact model and interface |
| Midjourney Editor | Existing or generated image | Edited composition | Area selection and framing |
MiniMax, Pexo, Black Forest Labs, Midjourney, and Ideogram solve overlapping but different tasks. JPEG, PNG, WebP, API, reference images, inpainting, and image-to-video describe different parts of that process. A PNG is a file format, not evidence of transparency; an editing endpoint is not necessarily a brush-based editor. That vocabulary matters when reading a product's feature list.
How do generation and editing differ?
For H3-Image, the documented generation endpoint accepts a prompt, while the edit endpoint takes a prompt with reference images. The API guide lists JPEG, PNG, and WebP output, and explicitly excludes masked edits, transparent backgrounds, and previews. Those constraints belong to the documented API; do not assume every interface has the same controls. MiniMax image guide.
| Task | Better starting material | Example instruction | Review priority |
|---|---|---|---|
| New concept | Written brief | A ceramic cup on a sunlit table | Shape and scene logic |
| Product relocation | Accurate product photograph | Move the cup to a shelf | Product identity |
| Palette exploration | Subject and palette references | Retain the cup; change the room palette | Which reference governs what |
| Campaign composition | Approved product and layout | Leave clear space for a headline | Crop and usable text area |
A reference image needs an explicit role. A room photograph can specify lighting without defining the product. A close-up can specify surface texture without imposing its crop. When two references disagree, decide which controls the disputed detail before generation. Otherwise, a visually plausible blend can be unusable for the intended advertisement.
A practical product brief separates fixed facts from creative freedom. For an imaginary blue ceramic cup, the handle shape and glaze may be fixed, while the tabletop, shadows, and camera position can change. Write those categories down. The example is a review method, not a promise that H3-Image will preserve every requested property.
What does open-weight mean for a team?
MiniMax describes H3-Image as open-weight and links its model repository. Access to weights gives developers a different starting point from a hosted creative app, but it does not settle licensing, hardware, or operational costs. Read the current repository and applicable terms before choosing a deployment path. Official model repository.
For a small design team, the main question may be who maintains the workflow. Hosted use puts less infrastructure in the team's hands; local or managed deployment adds responsibility for installation, updates, resource planning, and output handling. A downloadable model is useful only if those responsibilities fit the project. Do not mistake a download link for a tested production installation.
Keep a reproducible record for any serious evaluation: exact model identifier, interface, input files, prompt, settings, date, returned file, and acceptance decision. Record unsuccessful attempts too. A folder containing only selected successes cannot tell a team how much rework the next campaign will require.
How should you evaluate H3-Image editing?
Start with a narrow edit whose success is easy to judge. For example, request a new background while keeping the product's silhouette and label. Then compare the full image and close details with the original. If the requested change succeeds but the label changes, record a failed product-preservation check rather than calling the result a complete success.
| Check | Acceptance question | Common reason to revise |
|---|---|---|
| Identity | Is this still the intended product? | Changed shape or component |
| Wording | Is every required word correct? | Misspelling or invented claim |
| Composition | Does the planned crop work? | Important content near the edge |
| Consistency | Do selected images belong together? | Conflicting finish or scale |
| Delivery | Is the exported file usable? | Wrong dimensions or format |
Use the same review sheet for each candidate. A team that changes its standards after seeing the result will struggle to compare tools. Define which problems are disqualifying and which are aesthetic preferences. A different wall color may be acceptable; a different product model usually is not. Preserve the approved source as the authority throughout revisions.
For repeated edits, compare every revision to the original reference as well as the previous version. This catches gradual drift. An image can look almost identical to its predecessor while accumulating a visibly different handle, logo, or surface finish over several steps. Avoid using the newest generated version as the sole source of truth.
H3-Image versus FLUX: compare a specific workflow
FLUX is a family rather than one fixed endpoint. Black Forest Labs currently presents hosted API, playground, and open-weight options, with model-specific capabilities. Select the exact model and route before comparing it with H3-Image. A family-level label is too broad for a reproducible conclusion. Black Forest Labs.
A fair comparison starts from the same brief and the same source assets. Keep the intended output size and acceptance criteria comparable, and record any controls that one route cannot reproduce. Count accepted outputs rather than only measuring how quickly the first image appears. A faster first response may still lead to more manual repair.
The most useful result is often conditional: one route may suit a specific product background change, while another gives the designer better revision control. That conclusion is more actionable than an unsupported overall winner. This article does not report a benchmark ranking or claim that either model consistently outperforms the other.
When should an image move into Pexo?
Pexo's distinct role is turning a brief or approved visual into a finished video. Its public workflow describes planning, generation, conversational revisions, and delivery with narration, music, subtitles, and transitions. Pexo also offers image generation and automatic model routing. This does not establish that H3-Image itself is available inside Pexo. Pexo and image generation.
Give the video brief a purpose beyond making the picture move. Specify the message, intended audience, opening view, permitted camera movement, and closing action. For a product reveal, decide whether the viewer needs to understand a feature, recognize a design, or visit a destination. Those goals produce different scene choices even when the source picture is identical.
| If you need | Start with | Finish by checking |
|---|---|---|
| An image-led finished video | Pexo with an approved visual | Complete sequence and audio |
| A fresh still-image concept | H3-Image generation | Composition and factual details |
| A reference-led still revision | H3-Image edit route | Identity preservation |
| A selected-region edit | A documented area editor | Mask boundary and surrounding image |
| Deployment control | A suitable model repository | Terms and tested infrastructure |
Approval of the source image does not approve the animated result. Review motion for changes to the product, inspect text at actual viewing size, and listen to the complete soundtrack. A beautiful frame can become misleading when motion reveals invented geometry. Keep the source and the final video connected in the project record so revisions remain traceable.
