MAI-Image-2.5 - Generate and edit images with precise scene control

MAI-Image-2.5 is Microsoft AI's new image generation and editing model, now ranked No. 2 on Arena's image-edit leaderboard and No. 3 for text-to-image as of June 1, 2026.

Most image models treat editing as regeneration swap one thing, risk breaking everything else. MAI-Image-2.5 approaches it differently: localized edits that understand scene context, so changing a background, replacing text, or adding an object doesn't degrade what you didn't touch.

🖼 Text-to-image generation with stronger prompt adherence and text rendering so your prompts produce what you actually described, not an interpretation of it

✂️ Localized edits across objects, backgrounds, and text without affecting the surrounding image

🧑 Face and identity consistency preserved across pose, expression, and viewpoint changes useful for product, portrait, and commercial workflows

⚡ MAI-Image-2.5-Flash for high-throughput, cost-sensitive pipelines at roughly half the output token cost

Built for developers and ML teams embedding image generation or editing into production apps where controllability, identity handling, and cost-performance tradeoffs all matter.

Try it in the MAI Playground or access it via Azure Foundry and OpenRouter today.

P.S. I hunt the latest and greatest launches in tech, SaaS and AI, follow to be notified → @rohanrecommends

Looks impressive, congrats on the launch! I have a question: can I feed it a reference image and have it reproduce or closely match that image, rather than just editing parts of an existing one? Trying to understand whether the identity preservation works from a supplied reference, or only within an image I'm already editing.

The localized edit approach is what actually matters here, most models treat the whole image as fair game when you change one thing and you end up regenerating half the scene the identity preservation across pose changes is the interesting one for product and commercial workflows, that's where consistency usually breaks.

One question though, how does it handle complex prompts with multiple simultaneous edits? does it prioritize or does everything run in parallel?

The "precise scene control" angle is what caught my eye. I make all my own brand graphics, and the hard part is never one good image, it's keeping a set consistent. Does the scene control help hold style and layout steady image-to-image, or is it more about nailing a single composition?

Text rendering is the thing I check first with any new image model - most still mangle anything past a short headline. How does MAI-Image-2.5 handle multi-line text in non-Latin scripts, or is the benchmark mostly English?