Microsoft’s New MAI-Image-2.7 AI Model Improves Image Editing Capabilities

Microsoft’s MAI-Image-2.7 Preview will let people alter existing pictures in the company’s web playground, with preference-test scores close to OpenAI’s leading image models. Independent evaluator Artificial Analysis reports the largest gains over the preceding 2.6 model in reframing images, editing text and restyling scenes. Users supply an image and instructions for what to change.

More Control Over Existing Pictures

Image editing already belonged to Microsoft’s MAI family. Its MAI-Image-2.5 models supported editing since June, and the subsequent 2.6 series supported using multiple reference images. The new preview’s advance is the stronger measured preference for its edits over 2.6.

Microsoft pitches 2.7 around control: changing a scene’s lighting or visual style, revising words and symbols inside an image, and improving or restoring an existing asset. Its examples include moving a product into a new setting and changing individual objects while keeping the wider scene coherent. It also highlights reasoning-based edits, which account for relationships between objects and their surroundings.

Those tasks place different demands on an editor. A lighting change affects the look of a whole scene; a text correction needs the requested wording to fit inside an otherwise stable picture. In its analysis of the release, XenoSpectrum illustrates the distinction with a product photograph: replacing the background still requires preserving the product’s shape, texture and color.

Artificial Analysis reports improvement over 2.6 in all seven editing actions it measures. Its Elo ratings reflect how often evaluators prefer an edited image. Composition and framing gained 57 Elo points, text or symbol editing gained 49, and scene and style editing gained 48. The evaluator also places 2.7 first in four actions: scene and style, text or symbols, reasoning-based editing, and enhancement and restoration.

Microsoft separately reports gains over 2.6 across all 10 evaluated use cases and says its scene-and-style score is 24 Elo ahead of the next-best model.

Close Scores at the Top

The AA-Image-Editing v2.1 ranking compares outputs made from the same original picture and editing instruction. Evaluators choose between results without seeing the model names.

On October 28, 2.7’s overall rating was 42 Elo above 2.6. It sat between two separate OpenAI GPT Image 2.5 models: Flare and Sunburst, both tested at their max quality setting. OpenAI’s Flare and Sunburst models also accept text and images and return images as output. OpenAI positions Flare for speed and everyday generation, and Sunburst for detailed instruction following and editing precision.

The preview’s uncertainty interval overlaps both OpenAI models’ intervals, while its interval and 2.6’s are separate. The stronger predecessor result is therefore clearer than the fine ordering among the top three.

On October 27, Arena reported third place in its Image Edit Arena, at 1470 points, 45 points above 2.6 on its own rating scale.

Playground Editing and Developer Access

The live MAI Playground offers 2.7 for editing while retaining 2.6 in its image-generation selector. Microsoft plans text-to-image generation for 2.7 later, without giving a date.

Microsoft’s September 2.6 release brought the earlier series to developers in a Foundry public preview. Foundry is Microsoft’s service for building with AI models. For 2.7, Artificial Analysis lists no developer API, and Microsoft’s announcement gives neither pricing nor a Foundry launch date.