微软 MAI-Image-2.5 在基准测试中追平谷歌 Nano Banana 2

Microsoft's MAI-Image-2.5 pulls even with Google's Nano Banana 2 on benchmarks

AI 导读

微软的 MAI-Image-2.5 在 Arena 文本生成图像排行榜上排名第三，与谷歌的 Nano Banana 2 持平，但仍落后于 OpenAI 的 Image-2。该模型相比其前代产品有显著提升，特别是在渲染图像内文字和商业视觉效果方面。

正文 · 原文

Microsoft shipped an update to its MAI image model. MAI-Image-2.5 currently sits at number three on Arena's text-to-image leaderboard, according to the MAI team. That puts it on par with Google's Nano Banana 2, though still a clear step behind OpenAI's Image-2.

Microsoft calls it the strongest MAI image model yet, with major gains over MAI-Image-2 in text rendering, stylized illustrations, and commercial visuals. It's also said to follow prompts more closely and produce more consistent lighting, depth, and spatial relationships. Microsoft is pitching it for professional use cases like product photography and brand design.

Radardiagramm „Arena Model Rankings – Text to Image", das MAI-Image-1, MAI-Image-2 und MAI-Image-2.5 über acht Kategorien vergleicht. MAI-Image-2.5 übertrifft beide Vorgänger in allen Kategorien deutlich, mit den größten Zugewinnen bei Text Rendering, Portraits und Product/Branding/Commercial. — According to the Arena ranking, MAI-Image-2.5 clearly outperforms its predecessors in all eight categories, especially in text rendering, portraits, and commercial motifs. | Image: Microsoft

MAI-Image-2.5 is available now on Arena and should roll out to the MAI Playground and Foundry within two weeks.