Introduction

MiniMax H3-Image is MiniMax's own first-party image model built on its H3 family, and it surfaced on 9 November as an early-access preview on Hailuo. A wave of invited creators started posting results that afternoon, and the Hailuo web app now lists MiniMax H3 Image at the top of its image model picker. It's a closed, hosted model inside Hailuo for now, with no open weights or official API announced yet, so treat this as a Preview: real model, real page, but MiniMax hasn't published a launch post, pricing or a general-availability date.

The headline trick is references. H3-Image takes up to 16 reference images in a single generation and blends characters, products, outfits and locations into one frame, with creators showing output up to 4K.

What We Can Confirm

From the Hailuo app itself

Hailuo's live model config is refreshingly specific:

  • Model id: MiniMax-H3-Image, listed first among Hailuo's image models (ahead of GPT Image 2.5, Nano Banana and Seedream)
  • Mode: image-reference (text plus your own reference images)
  • Max reference images: 16
  • Max prompt length: 32,000 characters, so long art-director briefs are welcome
  • Prompt optimization: switched off, meaning your prompt goes in as written
  • Filter tags: edit, 4K and art

From early-access testers

Creators with access are reporting 1K, 2K and 4K output sizes and say H3-Image is strongest at editing and reference-driven work: character sheets and identity-locked portrait series, product and jewelry composites, exploded-view industrial and automotive concepts, CAD-style architecture, and AAA-style game environments. Several testers also point out it pairs neatly with H3 Video on the same platform, so you can generate a 4K still and animate it as a start frame without leaving Hailuo.

One honest note from the early tests: H3-Image shines with detailed prompts and rich references, while very short, simple prompts can land behind rivals. Write like a director, not a tagline.

How It Fits the H3 Family

If you've followed our H3 coverage, you may remember P-H3-Image, Pruna's open 8-step edit model built on MiniMax H3 open weights. H3-Image is different: it's MiniMax's own hosted image model, not a community fine-tune, and there's no word yet on whether its weights will be released. Until MiniMax says otherwise, assume it stays inside Hailuo.

Prompt to Try

Early testers found that adding the word "snapshot" pushes H3-Image toward candid, unretouched photorealism. Try it with two or three reference photos of the same person:

An authentic handheld snapshot from 2019 of the woman from the reference images, laughing while she shelters under a café awning during a sudden evening downpour, her denim jacket damp at the shoulders, natural skin texture with visible pores and a few freckles, no retouching, warm tungsten light spilling from the café window mixing with cool blue street light, rain streaks and soft bokeh from passing headlights, slightly off-center framing, shot on a phone camera

Then reuse the same references for a second shot, swapping the location and mood, to see how tightly H3-Image holds the identity across a set.

Original Source

Conclusion

H3-Image looks like MiniMax's answer to the multi-reference race: 16 inputs, 4K output and a model that clearly prefers long, specific briefs. If you're on Hailuo, check the image model picker for early access. Everyone else should watch for an official launch post with pricing and an API, and we'll update you the moment it lands.

Want to play with MiniMax H3 today? Try this on Gen → https://artrealmai.com/gen?utm_source=magazine&utm_campaign=gen&utm_content=minimax-h3-image-early-access-16-reference-4k

—Aurelia ♡