Problem Solutions
ChatGPT Image Generation Got Worse Since October 29, Users Report: Fewer Images, Ignored Edits and False Refusals
ChatGPT users say image generation has got noticeably worse since October 29. The complaints, gathered in OpenAI's own developer community forum, describe the same pattern across different accounts: fewer images per request, lower quality than before, edits that ignore explicit instructions, and harmless prompts refused for no clear reason.
The main thread, posted on November 1, is specific about what changed. One user asked for ten images and received one. Quality was described as subpar compared with earlier output. Instructions such as “do not change this” were ignored, with details not preserved. Repeated edit requests returned the same image. Character sheets and reference images were seemingly ignored. The user's summary was blunt: the tool “isn't trying as hard anymore.”
Other users confirmed the same timeframe. Reports of the deterioration cluster around October 29 across multiple accounts, which is what makes this more than one person's bad session. Some users speculated that people are being pushed toward Pro mode or that generation is being throttled. That is speculation, not confirmed fact, and it should be treated as such — but the shared date is a real pattern in the reports.
A second thread, posted October 30 by a Plus subscriber, describes seven days of image failures in sharper detail. Harmless edits were rejected as sexual content, including shortening a chin and emptying a swimming pool. The model added unrequested clothing, backgrounds and compositions. Instructions from previous images leaked into new ones. Support replies, the user said, read like templates, with no confirmed engineering investigation.
There is older context too. An October 23 thread complained that editing an image degraded the whole picture, with quality compounding downward with each edit. That is a separate, earlier complaint, but it shows editing quality has been a sore point before this latest round — and it makes the October 29 reports harder to dismiss as a one-day glitch.
What has not been established matters here. OpenAI has not confirmed any change to image generation, and no evidence in these threads ties the problems to a specific model update. The cause could be throttling, routing, safety tuning or a genuine regression — nobody outside OpenAI knows, and the forum threads do not prove any of those explanations. This article therefore reports user complaints, not a confirmed OpenAI change, and readers should hold that distinction.
If you are hitting these problems, a few workarounds are circulating among users, with no guarantee attached. Generate one image per request instead of asking for several at once. Start a fresh chat for a new edit rather than chaining edits on one image. Restate every constraint in full each time instead of relying on the model to remember earlier instructions. And describe the complete image you want rather than asking for a small diff on a previous one.
It is also worth reporting the problem properly if it affects you: use the feedback controls in ChatGPT and note the date, your plan type and the exact failure. If OpenAI investigates, dated, specific reports are what engineers can act on, and the October 29 cluster is already specific enough to check against.
If you want to test whether you are affected, keep the experiment clean. Run the same prompt in a fresh chat and in a chat with a long edit history, and compare the results side by side. Try one image per request before concluding the model has degraded, since the “ten images, got one” failure looks different from a quality failure. And note the date and your plan type: the forum pattern clusters around October 29 across accounts, so a dated report helps OpenAI more than a general complaint.
Some context on what sits behind the tool: ChatGPT's image generation is powered by OpenAI's GPT Image model family, which generates and edits images natively rather than bolting an image head onto a text model. That native design is why it handles text in images and chained edits better than older tools — and why a regression in editing behaviour is so noticeable to regular users. When the model stops listening to “do not change this,” it breaks the exact workflow that made it useful for design iteration.
The bottom line is simple. If your ChatGPT images got worse around October 29, you are not imagining it and you are not alone — but until OpenAI says something, the cause remains unconfirmed, and the fixes are workarounds, not solutions.