MiniMax H3 Image: 4K Generation and Editing with 16 References
Explore MiniMax H3 Image: up to 16 references, 4K area presets, Highspeed, prompt review with Context-IR, and practical evaluation workflows.
EvoLink TeamProduct TeamOn this page
For teams building product-image tools and marketing-asset workflows, the development worth watching is more control around a single creative task: organize source assets, define the delivery canvas, then generate candidates for review. EvoLink users can use these capabilities as selection criteria alongside the image models already available through a unified API gateway, while checking H3 Image access separately.
MiniMax’s direct APIs use synchronous requests, optional per-image streaming and a default 4K area preset. EvoLink’s route uses asynchronous tasks and defaults to 2K for automatic or ratio-based sizes; this integration covers the standard version. Highspeed, standalone Context-IR and per-image streaming described here are provider capabilities and are not included in this EvoLink integration.
What stands out in the release?
H3 Image can generate a new image from text or accept references and editing instructions. The main controls are:
| Capability | Supported capabilities | Workflow implication |
|---|---|---|
| Multiple references | 1–16 input images for editing | Specify the product, person, scene and visual direction separately |
| Flexible canvas | 1K/2K/4K area presets; custom output up to 20 million pixels | Prepare assets for different placements and delivery sizes |
| Batch generation | 1–10 images per request, with per-image streaming | Review completed candidates while others are still generating |
| Visible prompts | Returns the prompt actually used for each image | Inspect how a short instruction became a visual description |
| Prompt preprocessing | Context-IR can return an enhanced prompt without images | Review the expanded brief before text-to-image generation |
| Model versions | Standard and Highspeed versions | Compare both on the same tasks before choosing a default |
Bring more creative requirements into one image
Up to 16 references give teams room to distinguish the role of each asset. A product advertisement can supply the item, target location and style separately; a character project can specify identity, clothing and environment independently.
For a skincare-product scene, a front-facing product photo can establish packaging and bottle proportions, another image can show the intended space, and a visual reference can express lighting and color. Text instructions specify what must stay and what should change. Each source requirement then has a clear role that reviewers can trace.
For example, replace the cat in the first image with the Shiba Inu in the second, keeping the background. It demonstrates how to specify a source, a change and preserved elements. Product, character and scene combinations can use the same approach, with results validated on your own assets.
For product tasks, check the item first: shape, packaging, labels and proportions. For character tasks, compare identity, clothing and accessories across candidates. Multiple references supply more context; whether the output preserves the right details still needs validation with actual assets.
Adapt the 4K canvas to different placements
| Aspect ratio | 2K preset | 4K preset | Example delivery layout |
|---|---|---|---|
| 1:1 | 2048 × 2048 | 4096 × 4096 | Square product displays and social images |
| 16:9 | 2720 × 1536 | 5472 × 3072 | Landscape ads and video opening frames |
| 9:16 | 1536 × 2720 | 3072 × 5472 | Vertical social assets and video opening frames |
| 3:1 | 3552 × 1184 | 7104 × 2368 | Ultrawide banners |
The layout uses are creative-planning examples.
One area preset can therefore serve different placements, but each placement still needs its own composition. A product-page hero, a banner ad and a vertical social image need different framing and negative space even when their pixel areas are similar.
Output format is also configurable: JPG is the default, with PNG and WEBP available. JPEG and WEBP support compression settings, defaulting to 95; PNG ignores this parameter. It controls delivery-file compression, not a model quality or speed tier. H3 Image currently does not support transparent backgrounds, so choosing PNG does not enable transparency.
Review the expanded prompt before generating
revised_prompt, allowing teams to inspect both the original instruction and its expanded visual description.A brand brief might say only “a warm-toned bathroom product image,” while a reviewer still needs to confirm camera distance, product placement and background elements. Reading the expansion helps reveal whether those interpretations fit the creative direction. It provides a text-level check; the final image still needs review.
The prompt limit is 32,000 characters, but with enhancement disabled, only about the first 4096 tokens take effect. These are different input capacities. Teams maintaining their own complete prompts should still control length when disabling enhancement and state the subject, preserved details and delivery composition clearly.
Generate several candidates and review each as it finishes
A product brief can produce a batch to check the item, copy space and brand suitability of the background before choosing an image for further work. Batch generation provides multiple results in one request. If distinct creative directions are required, write separate briefs; do not assume one batch automatically covers ten predefined concepts.
Which production workflows should you evaluate first?
The following scenarios help evaluate these capabilities; they are not completed model tests.
Product scenes and marketing variants
Product teams can keep a product photo as a fixed input while trying different spaces, lighting and compositions. A campaign often needs a square product display, a landscape banner and a vertical social image. H3 Image's reference editing and canvas settings offer a shared starting point for these deliverables.
Check whether the bottle, shoe shape, packaging or other key structure is preserved before judging the setting. If packaging text or brand marks must be exact, list them as separate acceptance criteria; 4K support does not imply correct typography. Also check product proportions and negative space across formats to determine whether the workflow reduces manual cleanup.
Characters, scenes and concept exploration
For game concepts, character designs or story scenes, teams can supply separate character, clothing and environment references, then describe their relationships. When moving the same character into a new location, specify which appearance details must stay and which lighting or environment can change.
Acceptance centers on relationships between references: is the character holding the correct prop, has clothing borrowed unwanted elements from another image, and are positions correct? Multiple inputs make those conditions easier to express, but identity consistency, complex spatial relationships and repeated revisions still require testing.
Review a still image before video production
For a product clip or character video, H3 Image can be evaluated for opening-frame candidates. Approve the subject, set and composition, then pass the selected image to a separate image-to-video model. This provides a still-image review before investing in video generation.
Standard, Highspeed and open-source developments
MiniMax-H3-Image or MiniMax-H3-Image-Highspeed through the model field. Both versions share the same parameters and limits, so teams can evaluate them with the same assets, canvas and output requirements. The standalone Context-IR service supports only the standard version.Compare Highspeed and the standard version on completion time, acceptance rate and cost per accepted asset before choosing a production default. The Highspeed name alone does not establish a speed multiplier or quality ranking.
Self-hosting requires downloadable weights and a suitable model license; API access alone does not establish either. Check deployment conditions before choosing self-hosting, and confirm account access and charges separately when using an API.
Access and cost details for developers
MiniMax’s direct APIs accept text for generation, or text plus references for editing, and can return completed images individually. Use the request format and response handling of the channel you choose.
For costs, the provider reports input and output token usage, including reference-image usage and text-token equivalents of prompt-enhancement charges; generation billing counts successful images only. For a product-image task, record reference inputs, prompt processing, generation output and additional requests needed to pass acceptance. This is the basis for cost per accepted asset. Token examples do not substitute for actual H3 Image unit prices, and usage from two different tasks does not establish a fixed editing-price multiplier.
Frequently asked questions
Is MiniMax H3 Image the same as H3 image-to-video?
No. H3 Image outputs still images. H3 image-to-video takes image input into a separate video-generation workflow.
Does 4K always mean 4096 pixels wide?
No. 4K specifies target pixel area; dimensions depend on the aspect ratio. At 16:9, the 4K output is 5472 × 3072.
Can H3 Image generate transparent PNGs or masked edits?
H3 Image does not support transparent backgrounds or masked local editing. PNG is an output format option, not transparency support.
How does Highspeed differ from the standard version?
Both versions share the same parameters and limits. Compare latency and image quality on the same brief before choosing a production default; do not assume a fixed speed multiplier or quality advantage.
Does Context-IR generate images?
No. It returns an enhanced prompt. For text-to-image, review it and disable enhancement before submitting generation. Editing always enables enhancement.
