Microsoft unveils MAIImage1, its first in-house image generator

Microsoft this week introduced MAIImage1, its first image-generation model developed entirely in-house, saying the text-to-image system is designed to produce photorealistic images and will be rolled into Copilot and the Bing Image Creator after a test period. The company said MAIImage1 is currently available to select testers on the LMArena benchmarking platform, where it entered the top 10 of text-to-image models, and that it aims to deliver faster, more natural results for creators.

Microsoft described MAIImage1 as especially capable at rendering natural lighting, landscapes, reflections, shadows and other fine scene details. In a blog post, the company said it “trained this model with the goal of delivering genuine value for creators, and we put a lot of care into avoiding repetitive or generically-stylized outputs.” The team said it used curated training data and iterative evaluation with feedback from creative professionals to steer outputs away from the repetitive aesthetics that have characterized some earlier image generators.

On LMArena’s leaderboard, MAIImage1 ranked ninth with a score of 1,096 points. That placement followed a direct comparison with competing...