Microsoft (MS) recently released its first image generation model, "MAI-Image-1 (MAI)." MS said it developed MAI with the goal of generating photo-level realistic images. The current image generation model market is led by OpenAI's ChatGPT, and Google's Nano Banana is catching up. Attention is on whether MS, as a latecomer, can draw users' interest.
On Oct. 13 (local time), MS said on its website that it would unveil MAI. MAI is the first image generation model released by MS. MS explained, "We trained MAI with the goal of delivering true value to creators and took careful measures to avoid repetitive or generalized style outputs." MS also said MAI excels at generating realistic images such as lighting like reflections and landscapes.
◇ Excellent at background depiction but falls short on accuracy
The artificial intelligence (AI) model comparison site LMArena compared MS MAI, OpenAI ChatGPT, and Google Nano Banana (Gemini 2.5 Flash Image). All three models were asked to "depict Son Heung-min of Los Angeles FC celebrating after scoring a goal." They were also instructed to stage teammates rushing over to celebrate together.
However, MAI produced an image of an East Asian male player who could not be considered Son Heung-min. As MS described, the background elements such as lighting, spectators, and the grass field were more elaborate than competitors, but it fell short in accuracy. ChatGPT also drew a figure completely different from Son. In the case of the image generated by Nano Banana, the background was rougher than MAI's, but it rendered a figure closer to Son than the other models.
MAI successfully depicted the appearance of a foreign celebrity. The three models were asked to "draw Brad Pitt, an American actor, acting on the set of an F1 movie." MAI not only portrayed Brad Pitt realistically, but also arranged filming equipment such as racing cars and cameras around him to match the features of a set. In terms of bloom from studio lights and shadow expression, MAI outperformed competitors.
MAI rejected requests for certain subjects such as political figures. MS restricts responses about political figures under its own artificial intelligence (AI) ethics policy. All three models were instructed to "draw U.S. President Donald Trump giving a speech." ChatGPT and Nano Banana provided images after about 30 seconds, but MAI refused to generate an image, saying there was "an error." MS said about this, "We are building an ethical and safe AI model."
Unlike ChatGPT or Nano Banana, MAI was also unable to use AI to transform a desired image into another image. When LMArena instructed MAI to render a cat photo as if it were a 3D (dimensional) figure toy, a notice appeared saying, "The selected image generation model does not offer this feature." ChatGPT and Nano Banana, by contrast, provided the requested images.
◇ "MS continues to keep its distance from OpenAI with MAI launch"
MS, an early investor in OpenAI, maintained a close partnership, using the ChatGPT search engine in its own Copilot. However, as OpenAI demanded more computing resources and funds from MS and began selling AI products to corporate clients, the two companies reportedly drifted apart. Recently, MS also adopted AI models from Anthropic, a competitor to OpenAI, and "Grok" from xAI, an AI startup led by Elon Musk.
It is uncertain whether MS, a latecomer in image generation models, can catch up with OpenAI, which dominates the market. According to market research firm DemandSage, ChatGPT holds a 81.13% share of the generative AI market, ranking No. 1 in the industry.
IT outlet The Verge said, "MS is making large-scale investments in training its own AI models like MAI," adding, "The relationship between the two companies (MS and OpenAI) is becoming increasingly complicated."