Microsoft announced MAI-Image-1, its first image generation model developed entirely in-house. The model will become available on Copilot and Bing Image Creator “very soon” and can currently be tested on the LMArena platform, where it was initially benchmarked.
In developing MAI-Image-1, Microsoft stated its team focused on avoiding repetitive or generically stylized outputs. “For example, we prioritised rigorous data selection and nuanced evaluation focused on tasks that closely mirror real-world creative use cases,” a company statement explained. This development process also incorporated direct feedback from professionals working within creative industries to inform the model’s capabilities and overall refinement. LMArena, the platform used for testing, operates by having users pose queries to two anonymous chatbots and then vote for the superior response until a winner is determined.
The model is reported to excel at generating landscapes and photorealistic imagery. Its specific strengths include the accurate capture of intricate details related to lighting, shadows, and reflections within a generated scene. Microsoft noted that this high level of performance is particularly evident “when compared to many larger, slower models,” indicating an emphasis on computational efficiency in its design. This capability positions it as a tool for creating detailed and realistic visual content.
On the LMArena text-to-image leaderboard, MAI-Image-1 achieved a rank of #9 with a score of 1,096 points. For comparison, Google’s Gemini-2.5-Flash, also known as Nano-Banana, secured the #2 rank with 1,154 points, while OpenAI’s model was positioned at #7 with 1,123 points. The leaderboard is currently led by Hunyuan-image-3.0, an AI model developed by the Chinese technology company Hunyuan.
The creation of MAI-Image-1 is part of a wider in-house AI initiative at Microsoft. The company has also developed other proprietary models, including MAI-Voice-1 for natural speech generation and the Phi series of small language models, which are designed for efficient performance in reasoning tasks. This internal development occurs alongside the company’s continued support for OpenAI, which includes providing both financial backing and essential infrastructure for its separate model development efforts.
The AI image generation field is experiencing a period of intense activity. OpenAI’s model previously gained viral attention for its striking imitation of the Studio Ghibli art style, while Google’s Nano-Banana set a new benchmark with its powerful AI editing capabilities. Using LMArena, AIM conducted a direct comparison of Microsoft’s MAI-Image-1, Google’s Gemini-2.5-Flash, and OpenAI’s GPT-image-1. The models were tested with a prompt depicting “two people in a café by a window during late afternoon.” This specific test was designed to evaluate how well each model handled mixed lighting, reflections, and shadow realism. Users can submit similar prompts on the LMArena platform to test these models themselves.
