Microsoft Launches MAI-Image-2, Its Second-Generation AI Image Model

Microsoft has shipped MAI-Image-2, the second generation of its in-house AI image model. The model is rolling out now across Copilot and Bing Image Creator, Microsoft’s two primary consumer surfaces for AI image generation. According to Microsoft, MAI-Image-2 delivers “enhanced photorealism” and significantly more reliable text generation within images — a perennial weak spot for diffusion-based models.

Why text-in-image matters

Getting readable, correctly spelled text to appear inside an AI-generated image has been a consistent failure point for every major model since Stable Diffusion. Midjourney, DALL·E, and Adobe Firefly have all iterated heavily on this capability. Reliable in-image text unlocks real use cases: marketing mockups, presentation slides, social media graphics with accurate captions. If MAI-Image-2 has genuinely solved this to a production-quality level, it’s a meaningful differentiator for the Copilot suite.

Microsoft’s image model history

MAI-Image-2 is the successor to MAI-Image-1, Microsoft’s first proprietary image generation model. Until recently, Bing Image Creator ran on OpenAI’s DALL·E 3. The shift to an in-house model gives Microsoft more control over pricing, moderation, and iteration speed — the same strategic logic that pushed Google to move Google Images to its own Imagen series and Meta to its Emu family.

The model was developed under Microsoft AI (MAI), the research and product group that has been growing rapidly as Microsoft integrates AI across its entire product surface.

Where it shows up

MAI-Image-2 is live in two places:

• Microsoft Copilot — the AI assistant embedded across Windows, Microsoft 365, and the web
• Bing Image Creator — the standalone image generation tool accessible at bing.com/images/create

This means the model is immediately available to a massive installed base. Microsoft 365 alone has hundreds of millions of users, and Copilot’s integration across Word, PowerPoint, and Teams means MAI-Image-2 has a direct path to enterprise workflows — particularly slide decks and document design.

What Microsoft hasn’t said

The announcement from Microsoft is characteristically sparse on benchmarks. There are no side-by-side comparisons against DALL·E 3, Midjourney v7, or Stable Diffusion 3, and no published evaluation scores on photorealism or text accuracy. “Enhanced photorealism” and “more reliable text generation” are marketing claims until demonstrated.

The actual image samples Microsoft released — a set of photorealistic tiles — look strong on surface texture and lighting. Whether that holds up across diverse prompts, edge cases, and the kinds of requests real users actually make is something that will be tested by the millions of Copilot users now hitting the model.

The competitive read

Microsoft is in an interesting position. It has a deep partnership with OpenAI and still uses OpenAI models across much of its product stack. But developing MAI-Image-2 in-house signals that Microsoft doesn’t intend to be permanently dependent on OpenAI for generative media. This is the same pattern playing out across the industry: everyone who can build in-house image models is doing so, because the capability is becoming core infrastructure.