Introducing MAI-Image-2: for limitless creativity

TL;DR

Microsoft AI’s launch post for MAI-Image-2, a closed text-to-image model from the MAI Superintelligence team, positioned as the #3 model family on the Arena.ai text-to-image leaderboard at announcement. The post emphasizes three creative-work axes the model was tuned for — photorealism, in-image text generation, and rich/cinematic scene generation — based on consultation with photographers, designers, and visual storytellers. Distribution: a public MAI Playground preview, beginning rollout in Copilot and Bing Image Creator, and API access for select customers (with broader availability on Microsoft Foundry “soon”). No technical report, no weights, no benchmark numbers beyond the Arena placement claim.

Key claims

MAI-Image-2 ranks #3 in the model-family standings on the Arena.ai text-to-image leaderboard at announcement [§Introducing].
The model targets three product-defined creative axes: enhanced photorealism (natural light, accurate skin tones, lived-in environments), reliable in-image text generation (posters, infographics, slides, diagrams), and rich/cinematic scene generation (surreal, ornate, hyper-detailed compositions) [§Built with creatives].
Distribution surfaces at launch are the MAI Playground (preview), Copilot, Bing Image Creator (beginning rollout), and API access for select Microsoft customers — broader Foundry availability is promised but unspecified [§Make something today].
The post discloses no architecture, parameter count, training data, or quantitative benchmark beyond the Arena.ai leaderboard placement [§throughout].
The next-generation GB200 cluster is operational at MAI [§Build the future].

Method

Not disclosed. The post is a product announcement: no model size, no architecture family, no training-data description, no objective function, no eval protocol beyond a single Arena.ai leaderboard reference. The three product-axis improvements (photorealism, in-image text, cinematic scene generation) are described qualitatively and illustrated only by inviting users to try the Playground.

Results

A single quantitative claim: #3 on the Arena.ai text-to-image labs leaderboard at the time of announcement [§Introducing]. No FID, no GenEval, no DPG-Bench, no per-axis A/B against named competitors, no human-preference numbers from the consulted photographers/designers. The qualitative claims (photorealism, in-image text, cinematic scenes) are presented as design intents rather than as measured deltas.

Why it’s interesting

Adds another datapoint to the closed-but-API-accessible cohort the Open foundation-model releases page tracks as a comparison baseline — alongside Veo 3, Sora 2, Seedream 4.0, Nano Banana, GPT-Image, and (as a multimodal-embedding analogue) Gemini Embedding 2 (Gemini Embedding 2: SOTA multimodal embedding model (Google product announcement)). Sits in the same product slot as the open HunyuanImage 3.0 Technical Report release (80B-total / 13B-activated MoE T2I with full tech report and weights) and the closed Nano Banana 2 / Gemini 3.1 Flash Image (Google product announcement) Nano Banana 2 announcement — useful for tracking how much of the “frontier-T2I” surface is now closed-API vs open-weights. Contrast with the more disclosure-heavy product blog Introducing Recraft V4: Design Taste Meets Image Generation is informative: Recraft V4 ships a similar three-axis pitch (taste / typography / control) but with named benchmarks; MAI-Image-2 ships only a leaderboard rank.