MAI-Image-2 is a flagship text-to-image generation model developed by the Microsoft AI Superintelligence team. Introduced in early 2026 as a successor to MAI-Image-1, the model is designed to produce high-fidelity photorealistic imagery with a focus on natural lighting, accurate skin tones, and environmental detail. It was developed in collaboration with photographers and visual storytellers to ensure outputs meet the requirements of professional creative workflows, particularly for campaign-ready assets.
Technically, MAI-Image-2 is built on a flow-matching diffusion architecture containing between 10 billion and 50 billion non-embedding parameters. This approach enables a continuous transformation from noise to data, which improves training stability and alignment between complex text prompts and visual outputs. The model supports an extensive text context length of up to 32,000 tokens and generates images at resolutions up to 1024x1024 pixels, maintaining a maximum total pixel count of approximately 1.05 million across various supported aspect ratios.
Beyond realism, the model features enhanced in-image text rendering, allowing for the generation of legible typography in contexts such as signage, product packaging, and diagrams. At its release, MAI-Image-2 achieved a top-three ranking on the Arena.ai global image generation leaderboard. To support enterprise scale, Microsoft also offers MAI-Image-2-Efficient, a production-optimized variant that provides higher throughput and lower costs for high-volume tasks like e-commerce photography and UI mockups.