MiniMax Launches H3 Image: 16 References, 4K and Prompts Up to 32,000 Characters

If you have ever made a series of images for a single project, you know how painful it is. The first generation turns out great, but by the third the hero's face has changed, the jacket is a different color, and the background lives a life of its own. This is usually fixed with a combination of references, LoRA, inpainting and manual stitching in an editor. MiniMax offers a different approach: give the model more input data and a long technical brief.

What exactly was released

MiniMax has opened early access to H3 Image on its Hailuo AI platform. It is a model for still images, and, judging by the description, it inherits its "spatial discipline" from the company's video generator. Video models have long been trained to track where things are and how they move. It makes sense that the same logic should help in a single frame too.

According to the stated specifications:

• up to 16 reference images in one request;
• aspect ratios from 21:9 (a cinematic widescreen panorama) to 9:16 (vertical for stories and shorts);
• native export in 1K, 2K and 4K;
• a built-in text encoder with a context of up to 32,000 characters.

Why 16 references

Imagine you are making an ad banner for sneakers. Previously you could give the model one photo of the product and one mood, "like in this picture." Now, in theory, you can submit photos of the sneakers from three angles, the model's face and pose, a lighting reference, the brand palette, a sample background and a logo. All of this goes into a single request rather than a chain of five steps with losses at each one.

For a product manager, this looks like a shorter pipeline. For a developer embedding generation into their service, it is an opportunity to build the interface around a "set of materials" rather than around "one image and a text field." The more the model can keep consistent, the less you have to fix by hand.

A caveat: the stated number of references does not mean the model weighs all 16 equally well. In practice, multi-reference systems often let the first or most prominent images "win," while small details get lost. No one has publicly tested this yet.

Long prompt and 4K

A context window of 32,000 characters is more than a dozen pages of text. For an image, that volume seems excessive, but the idea is clear. You can describe a scene like a screenplay: who stands on the left, who on the right, what is in their hands, what the sign says and in what font. The author of a post in a Telegram channel directly says the model "should parse well" complex actor-placement schemes and typography. But that is an expectation, not a test result.

Native 4K matters too. Without an upscaler, there is no extra layer of "imagined" details, and text and small elements get more pixels to render properly. For print, large screens and cropping from a wide frame, this is significant.

What is not yet known

The source is a short announcement, and many important things are not stated in it:

• there are no benchmarks or comparisons with other image generators;
• it is not said whether there will be an API, or when;
• prices and request limits are not specified;
• it is unclear how the model handles Cyrillic in typography;
• access is still early, so queues, restrictions and instability are possible.

Phrasing like "prompt understanding, it seems, will be very good" is a cautious forecast, not a confirmed fact.

What to do in practice

If you make visual content or a tool for it, it is worth setting up a small test. Take your real task: a hero who must be recognizable across five frames, or a layout with text on a sign. Submit 4–8 references instead of one, and compare the result with what you get now. Separately, test a long structured prompt: does it work better than a short one, or does the model start ignoring the end?

If you are a product manager or founder, watch for the API to appear. It determines whether H3 Image becomes part of your product or remains just another pretty tab in someone else's service. For now, it is a claim of a strong tool for consistent generation, and it should be checked by hand, not judged by the announcement.