As of 22 Aug 2026, Microsoft AI released MAI-Image-2.5 on 2 June 2026. It ranks No. 2 on Arena Image Edit and No. 3 on text-to-image, with scene-aware local edits that change one region and leave the rest.
Upload a photo and describe the change — swap a background, replace text, or add an object. The live ChatImg generator opens on this page. Free to start.
MAI-Image-2.5's public claim is that you can change one named region and keep identity, nearby pixels, and layout. That is the same job ChatImg already sells as a chat instruction on an uploaded photo.
Microsoft called MAI-Image-2.5 its strongest image model yet. The public claim is a leaderboard jump plus a specific editing behaviour: change the named part of a picture without wrecking identity, layout, or nearby pixels.
Microsoft's 2 June 2026 post puts MAI-Image-2.5 at No. 2 on Arena's Image Edit leaderboard, ahead of Nano Banana 2, and ahead of GPT-Image-1.5 on the same board.
The same post ranks it No. 3 on Arena text-to-image, above GPT-Image-1.5 and Nano Banana Pro 2K in Microsoft's snapshot.
Product Hunt's listing: swap a background, replace text, or add an object while identity, nearby pixels, and layout stay put. That is the job, not a full re-roll of the picture.
ChatImg does not list MAI-Image-2.5 in the model picker. It does let you pick Qwen Image Edit or GPT-Image-2, upload a photo, and describe the change — background, text, or an extra object — without rebuilding the whole frame.
Start from a product shot, poster, or portrait. Name the edit in one sentence. The rest of the picture is the thing you keep.
Open the ChatImg model list and choose Qwen Image Edit for instruction-led local edits, or GPT-Image-2 when you want a fresh cutout or restyle. No Discord, no new subscription to start.
Try the generator from the daily free quota. The same chat is localized across ChatImg's 15 locales, so the instruction can be in the language you already write.
No API key, no Discord, no new subscription to start.
Use the editor on this page, or the deep link with Qwen Image Edit already selected.
"Swap the background to a studio grey." "Replace the sign text with OPEN." "Add a coffee cup on the table."
Check that faces, products, and layout still match the original. If they do, you are done.
These are the task verbs people actually search — not a model name.
Ecommerce SKUs and creator thumbnails. Keep the subject; drop in a studio, a street, or a solid brand color.
Posters, menus, packaging mockups. Change the words without redrawing the layout around them.
A product on a table, a prop in a portrait. The new object should sit in the same light, not look pasted.
Stay on the generate-and-edit path — each of these is a live page, not a dead end.
Text-to-image and image-to-image in one chat.
The ChatImg landing for instruction-led local edits.
Generate cutouts with a real alpha channel.
Repair scratches and fading from one upload.
Photo-first generate and edit, no Discord.
Instruction-based image editing changes an existing picture from a written instruction such as "make the shirt blue". The original image is kept and only the named part is altered. InstructPix2Pix (Brooks, Holynski, Efros, 2022) showed that no mask is required — the model infers the region from the words.
Scene-aware means the model treats the rest of the picture as a scene that should still make sense: identity, lighting, and nearby pixels stay consistent after the named change. A background swap that leaves a halo, or new text that ignores perspective, is the failure mode this claim is meant to avoid.
Microsoft AI, 2 June 2026: MAI-Image-2.5 is its strongest image model yet, ranking No. 2 on Arena Image Edit (ahead of Nano Banana 2) and No. 3 on text-to-image.
Microsoft AI — Introducing MAI-Image-2.5 ↗Product Hunt lists MAI-Image-2.5 as a text-to-image and image-editing model for localized edits, identity preservation, and text rendering, with developer access via Foundry and OpenRouter.
Product Hunt — MAI-Image-2.5 ↗InstructPix2Pix (Brooks, Holynski, Efros, 2022) established instruction-following image editing: one written instruction, no user-supplied mask, edits in a single forward pass.
InstructPix2Pix — arXiv:2211.09800 ↗FAQ'S
Ask us anything!
Upload a photo. Name the change. Keep the rest. Qwen Image Edit and GPT-Image-2 are in the model list.
