Microsoft
Microsoft's MAI-Image 2.5 did not just fail the arena because Azure ResponsibleAI blocked 10 benign standard-set prompts. The accepted images also showed enough basic scene-logic, framing, object-continuity, and rendering failures that the model is hard to recommend even before the guardrails enter the conversation.
Microsoft / 1K / 1:1 / microsoft/mai-image-2.5
The images
Same standard set, same framing, same model comparison surface.
I · Through the ages
Same eleven eras, two life stages, side by side: a group of fresh-faced nineteen-year-olds doing what the young did — hunts, dances, mosh pits, raves — next to middle-aged forty-year-olds doing what the settled did — harvests, workshops, offices, backyard parties. Different people, different lives, one timeline. Watch how each model handles youth vs. age across history.
II · Fashion & portrait
Editorial polish, neon glamour, old-Hollywood and couture, and a 1950s centerfold — faces, skin, fabric, and styling.
III · Product
Reflections, materials, and one duck modeling everything — the commercial-shot test, with a surreal closer.
IV · Pets
Small animals, soft light — the “make-it-adorable” test.
V · Food
A six-patty monster, two models eating it, and a humble popsicle dressed up like luxury perfume.
VI · Worlds & abstract
A neon-cyberpunk country town and a geometric color explosion that melts — imagination over realism.
The model
Microsoft's MAI-Image 2.5 did not just fail the arena because Azure ResponsibleAI blocked 10 benign standard-set prompts. The accepted images also showed enough basic scene-logic, framing, object-continuity, and rendering failures that the model is hard to recommend even before the guardrails enter the conversation.
The guardrails are the loudest failure, but they are not the only one. In the accepted set, 1960 · 19 uses strange colored framing instead of a convincing beach-bonfire scene, 1980 · 40 and 2000 · 40 both place the computer so it faces away from the people using it, and 2025 · 19 includes phone and pet handling that does not make physical sense.
Several images have the surface polish of a modern generator without the underlying object logic. The luxury popsicle reads bizarrely over-sharpened, and one structural/abstract output has parts that do not connect to what appears to be holding them up. The model can make a sharp-looking square, but too often the scene does not survive inspection.
The blocked prompts were not fringe requests. MAI-Image 2.5 refused harmless social, portrait, fashion, product-continuity, and food-ad prompts, including a 1950s sock hop, an 1980s roller disco, the whole original fashion/portrait section, the duck product-continuity test, and both burger-eating ad portraits.
OpenRouter surfaced the failures as Azure ResponsibleAI content_safety_violation blocks. That lines up with Microsoft's stacked safety architecture: prompt-side filtering, output-side filtering, and Azure policy layers can each stop a request. The practical problem is not that unsafe prompts were blocked; it is that ordinary creator prompts crossed an opaque threshold with little explanation.
Neutral, comma-separated rewrites fixed the 1950 sock-hop and 1980 roller-disco prompts on the first attempt. Replacing “pinup centerfold” with a “tasteful vintage fashion magazine page” passed on the second attempt. Editorial and couture only passed after removing the person and using a mannequin. Neon, Old Hollywood, duck/waterfowl product continuity, and both burger-with-person variants kept blocking after multiple safer rewrites.
For solo creators, MAI-Image 2.5 is a hard do-not-use in its current form. The guardrails repeatedly refuse normal commercial prompts, and the accepted outputs still miss enough basic composition, object continuity, and physical logic that this is not a model to build a production workflow around.