Illustration by Haein Jeong / The Verge
Microsoft is rolling back a model upgrade to its AI-powered Bing Image Creator, reports TechCrunch. The rollback came after weeks of complaints by users that the tool just didn’t work as well after Microsoft “upgraded” to a new version of the DALL-E 3 model on December 18th.
Microsoft declined to comment on its decision to roll things back or offer specifics on what may be causing the gap between user’s expectations and its output.
Today, Microsoft’s head of search, Jordi Ribas, tweeted that they could reproduce “some of the issues reported” and are reverting to an older version of the DALL-E model for now, although it could take a few weeks until it’s complete.
Since the launch of Bing Image Creator last spring, users have generated billions of images with text prompts. I’m pleased to share our latest updates to enhance your creative experience. Starting today, we’re rolling out the latest DALL-E 3 model PR16, which will create images… pic.twitter.com/3p9HsYMlr6
— Jordi Ribas (@JordiRib1) December 18, 2024
Thanks again for the feedback and patience. We’ve been able to repro some of the issues reported and plan to revert to PR13 until we can fix them. All Pro users and about 25% of the requests using boosts are now on PR13. The deployment process is very slow unfortunately. It…
— Jordi Ribas (@JordiRib1) January 8, 2025
As soon as Ribas posted about the change in December, there were complaints that Bing Image Creator was producing less-detailed results or images that didn’t accurately reflect their prompts. In his initial replies, Ribas said the model’s output quality “should be a bit better on average” than before.
It was the same story in posts and comments on Reddit and OpenAI’s community forums. On OpenAI’s forums, a person complained about the model’s handling of fabric on an anime-style character’s dress. The person who posted the below images says the one on the left is “perfect quality” while the one on the right is “over-lit.”
Screenshot: OpenAI community forum
Another didn’t like the way Bing placed starburst effects:
Screenshot: OpenAI community forum
All of these things are subjective, and I can’t claim to think any of them look better than any others. If anything, it feels like an indication that Microsoft doesn’t just have to deal with complaints about bugs or people upset about feature changes — now it has to deal with AI art critics comparing the machine’s output to what they imagine it should create. Perhaps they should ask the artists whose work the generators were trained on for tips about managing a client’s expectations?