Wait, Has Microsoft Changed?
Honestly, from an ordinary consumer's point of view, Microsoft's AI tech was the definition of lame. MS Copilot felt like a ChatGPT knockoff, and the hallucinations were so bad that if you got into a serious conversation, you'd get fooled before you knew it.
The Copilot built into Office just rambled nonsense and kept repeating, "I can't do that." And the image generation services Microsoft offered, MS Designer and Bing Image Creator, didn't get good reviews either, to be honest.
You've probably been to those sites, and you probably thought, "This is just a place for kids to play with drawing," and then moved on to some other image generation site. At least, if you were really trying to make images seriously.
✦ Side Note
Each AI's specialty
In December 2024, one Microsoft Designer user posted a complaint that the image generation results had changed from before. Faces and clothing details were mushy, the whole frame had gone way too bright, and blurry text was even floating in front of the image.
The user griped that they'd entered the exact same old prompt and the results got worse. Less than a month later, the same kind of thing blew up with Bing Image Creator.
Contrary to the announcement that the new model was twice as fast with better image quality, users protested that the pictures had turned flat and lost detail. In the end, Microsoft decided to roll back to the previous model.
But in October 2026, Microsoft announced its own in-house image model, and MAI Image 2.7 is now fighting GPT Image 2.5 for the top spots in editing evaluations. Pretty amazing.
Whose Technology Was Microsoft's Image Generation, Anyway?
Bing Image Creator first appeared in March 2023. Microsoft built generation features into Bing and Edge, but the engine actually making the images was OpenAI's DALL·E. After that, the Image Creator in Designer, which came into Copilot, also used DALL·E 3.
In April 2025, OpenAI's GPT-4o image generator went into the Create feature of Microsoft 365 Copilot. Then from December 2025 through March 2026, the image generation in Microsoft 365 Copilot and Copilot Chat was swapped from GPT-4o to GPT-Image-1.5.
In other words, the image generation services Microsoft has offered so far were almost all OpenAI image models integrated into its products. Even when a user made an image in a Microsoft service, the model actually generating it was OpenAI technology.
✦ Microsoft Image Generation and Editing Model Timeline
From OpenAI integration to in-house models
On December 18, 2024, a PR16 build of OpenAI's DALL·E 3 started rolling out to Bing Image Creator. Microsoft was confident that image generation would get twice as fast and the quality would go up. But users reacted in exactly the opposite way.
Comparison images started showing up on X and Reddit, saying results looked less realistic than before, fabric and lighting had gone flat, and the output had turned cartoonish.
This didn't end as a simple complaint. Jordi Ribas, Microsoft's head of search, confirmed some of the problems and said they would pull PR16 and go back to the earlier version, PR13.
An update meant to boost speed and quality got rolled back because of user reaction. The benchmarks for an image model may have improved, but if people can't get the pictures they want, it's a failure.
The complaints about MS Designer kept coming too. In March 2026, a Microsoft 365 Premium subscriber posted, "I used to get four usable images from one prompt, but now I wait a long time and only get one lower-quality image."
The comments also piled up with stories of results too poor for professional work.
The aspect ratio was a pain point too. In April 2026, questions about Bing Image Creator included requests for a 21:9 ultrawide ratio and proper image-to-image editing.
That's because the image size choices in Microsoft 365 Copilot Create had only three presets: square 1024×1024, portrait 1024×1536 and landscape 1536×1024.
Microsoft: We Make Our Own Images Now! Teehee!
Microsoft changed course in October 2025. The MAI-Image-1 it unveiled then was the first image generation model it had developed entirely in-house.
In November, MAI-Image-1 went into Bing Image Creator's model selection menu right alongside DALL·E 3 and GPT-4o.
Next, MAI-Image-2, released in March 2026, started rolling out to Copilot and Bing Image Creator too. In September, MAI-Image-2.6 was released for developers on Microsoft Foundry.
It came with the ability to edit by referencing several images, make images that reflect web information, and adjust the aspect ratio to fit the screen layout. It supported resolutions up to 1.5K.
And finally, on October 27, MAI Image 2.7 arrived. This model is focused on how precisely it can edit existing images, more than on making images from scratch.
✦ MAI Image 2.7 Editing Features
Fine-tuning images that already exist
Typical jobs include changing the lighting and mood of an entire scene, moving a product onto a different background, editing text inside an image, and restoring old images.
How Does MAI Image 2.7 Perform?
The most eye-catching result is the image editing leaderboard at Artificial Analysis. As of October 31, 2026, GPT Image 2.5 Sunburst was first at 1,199 Elo, the MAI Image 2.7 Preview second at 1,193, and GPT Image 2.5 Flare third at 1,182.
Considering that in actual use it's hard to feel a big difference between the Sunburst tier and the Flare tier of GPT Image 2.5, MAI has already caught up to GPT Image 2.5 on benchmark scores.
✦ Artificial Analysis Editing Leaderboard
Checked October 31, 2026, Elo scale
The evaluation samples were 4,891 for MAI, 18,250 for Sunburst and 19,136 for Flare. The 95% confidence intervals overlap, at 1,178 to 1,208 for MAI and 1,188 to 1,210 for Sunburst.
So you can't say MAI clearly beat the top GPT Image 2.5 variant, but it's true that it has climbed to a score that can compete in the same top tier.
Microsoft said it came out ahead of the other models especially in the editing categories that change scenes and styles.
In Arena's single-image editing ranking, MAI Image 2.7 came in third as well, at 1,471 points. GPT Image 2.5 Sunburst scored 1,517 and Flare scored 1,482.
✦ Arena Single-Image Editing Ranking
Shown October 23, 2026, a separate scale from Artificial Analysis
MAI got 3,402 votes, while the two GPT models logged 90,545 and 87,941 votes. It didn't get many votes for its score, and my guess is that MAI is still unfamiliar to everyday users, so it still hasn't gotten attention that matches its performance.
Independent creator DrSadek fed the same image and prompts to MAI Image 2.7, GPT Image 2.5 Sunburst (max) and Nano Banana 2.1 and compared five edits.
On jobs like changing a coat's color, turning a daytime scene into evening, swapping a perfume bottle for a ceramic vase and fixing the wording on a poster, the three models mostly did what was asked.
All of them spelled the poster text right; only the line breaks differed.
✦ The Chair Color That Split the Results
When three requests were applied in order
The difference showed up when the edits were chained. Three requests were applied in order: first turn the cushion blue, then turn the collar red, then change the background to a recording studio.
Along the way, MAI and Nano Banana changed the chair color even though nobody asked them to. GPT Image preserved the chair's color well.
This scene shows the tricky part of image editing well. Changing what you want is one thing, but leaving everything else alone matters just as much.
If you're fine-tuning a poster or a product image, a small change like this can have a big effect on the final result.
Hoping MAI-Image Becomes a Real GPT Image Rival
MAI Image 2.7 is still just an image editing preview. Right now you can only test its editing performance on MAI Playground, Arena and Artificial Analysis, and the feature for making new images from text alone will be released later.
So the editing ability is pretty impressive, but whether it has caught up with GPT Image in pure image generation needs more verification.
Personally, I get the feeling that GPT Image, the best image editing model right now, gets more and more likely with each version to ignore your instructions and keep drawing only the style it wants. Is it just me?
Back in the day, changing the prompt got you tons of really creative pictures, right? But more and more it feels like a distilled version of a local image generation model, tending to spit out only the "pretty images" people like to look at.
I'd love for MAI Image to rise up as a new alternative and give consumers a choice.
Anyway, once MAI-Image's image generation ability is released, let's take another look at it then.
That's all for this post. See you next time.
Related posts
- Midjourney Thinking Mode: Image AI That Fixes Its Own Errors
- Medical AI MAI-DxO Beat Doctors. Will Doctors Disappear?
- World Models Arrive: Tencent Hy-Nomas and Google Genie 4
🌐 Korean version: 허접했던 MS의 이미지 편집 기술... MAI Image 2.7이 GPT Image 2.5를 따라잡았다!
