MAI-Image-2.7-Preview Operational

MAI-Image-2.7-Preview is Microsoft's image editing model. It ranks second globally with 1,193 Elo on the Artificial Analysis image editing leaderboard, behind only GPT Image 2.5 Sunburst, and leads the next-best model by 24 Elo in scene and style editing. MAI-Image-2.7-Preview excels at complex tasks such as changing styles, relighting, editing text and symbols, restoration, and enhancement. It is currently available on platforms such as MAI Playground. Its weights are not open, and it remains a Preview release.

Scene and style transformation: Redraw the style, color tone, and atmosphere of an entire scene in one step. This is the model's most significant lead, outperforming the next-best model by 24 Elo.
Intelligent relighting: Adjust lighting and mood so that shadow direction and material reflections follow physical logic.
Text and symbol editing: Modify text and symbols in an image while integrating them with the original lighting, perspective, and material.
Restoration and enhancement: Remove noise, restore old assets, and improve image quality while consistently preserving the original composition.
Reasoning-based editing: Understand the relationships between objects, lighting, and the environment to handle complex instructions that require reasoning, such as moving a product to a new scene.

Diffusion-based generative architecture: The model uses a diffusion-based generation method and learns the continuous transformation between noise and data distributions through a flow-matching loss, ensuring stable and efficient training.
Multimodal editing reconstruction: It follows an approach of rebuilding from scratch for multimodal editing workflows, while performing cross-dimensional reasoning over objects, scene structure, lighting, scale, and spatial position. It can produce semantically consistent edits even from ambiguous instructions.
Multi-constraint collaborative editing: The model can simultaneously handle multiple constraints, including layout preservation, object changes, text updates, and contextual adaptation. It supports object removal, replacement, attribute modification, and inpainting without disrupting the overall composition, while maintaining visual consistency across multiple editing iterations.
Safety alignment mechanism: The model filters data and applies additional mitigation measures during deployment, with the goal of preventing the generation of harmful images involving violence, gore, pornography, and similar content, even when requested by users.

Visit the official website: Go to the MAI-Image-2.7-Preview official website at https://playground.microsoft.ai/chat.
Sign in: Click the sign-in button in the upper-right corner and use a Microsoft account to sign in. The model is available to try for free.
Upload an original image: Select the MAI-Image-2.7-Preview entry in the Images/image editing section, then click upload or drag in an existing photo, design, or screenshot as the editing base image.
Enter instructions and generate: Write a clear modification instruction in English in the input box, then click the generate button and wait a few seconds for the result.
Refine and download: Review the generated result. If it is unsatisfactory, revise the instruction and iterate again. Once satisfied, click download to save it locally.

Second place globally: It ranks second globally with 1,193 Elo on the Artificial Analysis image editing leaderboard, behind only GPT Image 2.5 Sunburst (max), outperforming offerings from Google, Meta, xAI, and GPT-Image-2.5-Flare.
Strong lead in scene and style editing: It leads the next-best model by 24 Elo in the scene and style editing category, its largest margin across all capabilities.
全面超越自家上代：相较 MAI-Image-2.6，在 Artificial Analysis 全部 10 个用例和 7 类编辑动作上均有提升，无短板。
Four capabilities reach the category frontier: Restoration and enhancement, reasoning-based editing, scene and style editing, and text and symbol editing all reach category-leading levels.

Project website: https://microsoft.ai/news/mai-image-2-7-debut/

E-commerce product image transformation: Move product photos to new scenes in one step while automatically matching lighting and perspective, quickly producing product hero images and marketing assets in multiple styles.
Design and UI iteration: Apply style transformations, change theme colors, and explore layout variations in designs and screenshots, enabling designers to test multiple directions at low cost before finalizing a design.
Old photo and asset restoration: Repair noise and scratches in old photos, improve the quality of low-resolution assets, and consistently preserve the original composition. This is suitable for archive digitization and asset reuse.
Marketing material localization: Directly modify text and symbols in posters, billboards, and packaging while integrating them with the original perspective and lighting, eliminating the need to re typeset the material.