Microsoft unveils MAI-Image-1, its first in-house developed image generation model

For the past few years, Microsoft has heavily relied on OpenAI’s AI models to power its products and services. Except for the Phi family of small language models, Microsoft has not developed any major models on its own. In August this year, Microsoft surprised everyone by announcing its first two in-house AI models: MAI-Voice-1 and MAI-1-preview.

Today, Microsoft announced its third in-house AI model, MAI-Image-1. MAI-Image-1 is an image generation model that will compete with OpenAI's gpt-image-1 and Google's Gemini/Imagen models.

Microsoft stated that it trained the new MAI-Image-1 model with the goal of delivering genuine value for creators. Specifically, the team focused on preventing the model from producing repetitive or overly stylized outputs. Microsoft achieved this by prioritizing rigorous data selection and nuanced, task-focused evaluation. The company also claims that MAI-Image-1 performs better at generating photorealistic imagery, landscapes, and more.

Microsoft further highlighted that this new model is significantly faster compared to other large models on the market. The company believes users will be able to generate their desired images more quickly and iterate through them faster.

In LMArena's Image Arena, an open platform for community-driven AI benchmarking, Microsoft's MAI-Image-1 debuted at the #9 position. While that may not sound remarkable, it represents a solid achievement for a first-generation model developed entirely in-house. ByteDance, OpenAI, Google, and Tencent are currently ahead of Microsoft on the leaderboard.

Microsoft AI CEO Mustafa Suleyman mentioned on X that they will continue refining the model to climb higher on the leaderboard.

As Microsoft continues to refine the model, the creative community will be watching closely to see if it can challenge the current industry leaders.