Microsoft Launches MAI-Picture-2 Textual content-to-Picture Mannequin—And It's Higher Than Anticipated – Decrypt

Microsoft has been quietly constructing its personal picture generator. Introduced Thursday by the corporate’s AI Superintelligence workforce, MAI-Picture-2 has already landed at #3 on the Enviornment.ai leaderboard—behind solely the fashions from Google and OpenAI—making Microsoft a legit participant in an area it had beforehand outsourced to its companions.

That final half is price sitting with. Microsoft has been paying OpenAI billions to energy Copilot and Bing Picture Creator. Constructing a competing picture mannequin in-house is an attention-grabbing enterprise transfer.

MAI-Picture-2 is on the market now within the MAI Playground, with a gradual rollout to Copilot and Bing Picture Creator underway. API entry is at present restricted to pick out enterprise prospects, with broader availability on Microsoft Foundry coming quickly.

The workforce says it constructed the mannequin by speaking on to photographers, designers, and visible storytellers. Three issues got here out of these conversations: improved photorealism, extra dependable in-image textual content technology, and stronger capability for detailed, imaginative scene building. Whether or not or not that course of translated right into a genuinely great tool is a special query.

Testing MAI-Picture-2

The very first thing you discover once you open the MAI Playground is how understated it’s. The interface is minimal and clear, visually someplace between Claude and Hume, with not one of the maximalist dashboard vitality you get from Midjourney or the chatbot expertise you get from Gemini.

The photographs themselves are genuinely fairly robust. Photorealism is an actual power right here—the mannequin has a stable grasp of pure gentle, floor texture, and spatial relationships. It would not fairly hit the extent of Google’s Nano Banana Professional, which nonetheless guidelines the leaderboard for a purpose, however in some realism assessments it comes surprisingly shut.

Higher prompting seemingly pushes it additional; our preliminary outcomes improved noticeably as we dialed in our descriptions.

Even complicated, unrealistic scenes with parameters that defied logic had been correctly dealt with by the mannequin, beating different fashions in particulars just like the physique proportions, limb place, depth, and spatial positioning.

For instance, this picture of a canine driving a motorbike in the midst of the ocean is arguably essentially the most correct one we’ve produced in zero-shot assessments.

Textual content technology is a legit spotlight. MAI-Picture-2 dealt with complicated typography with way more consistency than we anticipated—giant blocks of textual content in pictures, posters, signage—with out the standard garbling you see from most fashions.

We even pushed it towards multilingual textual content: It managed to generate some hanzi Chinese language characters, although the accuracy wasn’t good. Nonetheless, the truth that it tried and obtained partway there’s notable.

The mannequin understands creative fashion effectively, shifting between photographic realism, graphic design aesthetics, and illustrated kinds with out a lot friction. It reads prompts fastidiously, together with stylistic directions, and delivers one thing coherent on the opposite finish. For a broad vary of visible duties, it is versatile.

Now for the more durable truths.

MAI-Picture-2 is aggressively filtered—extra so than Google Imagen, and extra so than OpenAI’s DALL-E. We ran our common check of a cartoon drawing of a spider chasing a girl, and obtained a flat refusal. Once more, that is a drawing—of a spider. The content material moderation right here is tuned to a stage that can frustrate anybody doing inventive work in grey areas, horror illustration, or something that reads as remotely tense.

The utilization limits are equally restrictive. Every technology triggers a 30-second cooldown. After 15 pictures, you are locked out for twenty-four hours. For informal experimentation, that is manageable. For any sort of manufacturing workflow, it is a dealbreaker within the native UI.

There’s additionally just one decision: 1:1. No panorama, no portrait, no customized ratios. In 2026, that is a major limitation—notably for social media content material, which is exactly the place Microsoft presumably desires this embedded in Copilot.

And talking of Copilot: MAI-Picture-2 is not there but. The rollout is going on, however as of as we speak, the product you’d really need it in would not have it.

Yet one more lacking piece: That is purely a text-to-image device. No image-to-image, no inpainting, no outpainting, no reference picture assist. For customers anticipating something near Firefly or Midjourney’s enhancing capabilities, this may really feel half-finished.

Our take

MAI-Picture-2 performs higher than its leaderboard rating suggests. In our hands-on assessments, it beat GPT-Picture on picture high quality and textual content rendering, which is attention-grabbing provided that GPT-Picture sits above it on Enviornment.ai’s leaderboard. Benchmark positions do not all the time inform the complete story.

The strategic logic behind constructing that is clear. Microsoft has been licensing OpenAI’s picture fashions for Copilot whereas concurrently funding OpenAI’s greatest competitor, Anthropic. Having a succesful in-house mannequin reduces dependency, cuts prices at scale, and offers Microsoft one thing to iterate on with out asking for permission.

From that angle, MAI-Picture-2 would not must beat Nano Banana. It simply must be adequate—and it’s.

The issue is the product constraints. The technology caps, the strict content material coverage, the 1:1-only output, the lacking enhancing options, and many others; these are the sorts of limitations that put a ceiling on real-world utility. A mannequin this succesful deserves infrastructure that matches it.

MAI-Picture-2 is a powerful technical basis hamstrung by conservative product choices. As soon as Microsoft loosens the restrictions, this turns into a severe contender. Proper now, it is a promising preview of what Microsoft’s picture stack might really develop into.