ChatGPT Users Argue Over Image Generation Getting Worse After October 29

⚡ The gist of the dispute

A thread author under the nickname 1184195305 noticed changes in ChatGPT's image generator starting October 29 (the post text says 2026) and listed specific failures: instead of ten images only one arrives, edits aren't applied, and things they asked not to touch get changed anyway. They ask whether others are experiencing the same and remind that they pay for a subscription. In the OpenAI Community thread (12 comments, 0 votes, 469 views) several people with similar experiences turned up — but not everyone's picture is the same.

🗣 What participants are saying

The author himself describes three separate symptoms. The first is quantity: "I usually ask for 10 images to be given back to me yet I only receive 1". The second is quality: the images are "subpar compared to what it has usually given me in the past". The third is controllability: the request "do not change this" is ignored, and an edit request returns the same image. His conclusion is blunt: "It isn't trying as hard anymore and doesn't listen to prompts".

A participant under the nickname stgore212 confirms the decline, but tested it more broadly — across several accounts. According to him, the same prompts and the same references now yield noticeably worse results, repeated attempts don't change the picture, and the effect reproduces from run to run. He also looks for the cause: asking whether there were "recent changes to image generation or reference-image p…" (the comment text is cut off), and mentions seeing similar complaints from other users around the same dates.

battlecat044 briefly joins the observation: long-standing characters stopped assembling properly for him, and that's "in the most basic conditions". themaggielange reacts more emotionally and more broadly: she feels both services have degraded, "Even gemini gone bad", and reminds that she came to ChatGPT in the spring precisely because of this.

showaemergency puts forward a working hypothesis: it may be about the thinking level. He asks whether Pro mode makes a difference, and admits his Pro generations are already used up for the week, though he saw the same behavioral change the previous evening. His main complaint is "character sheets / character references are seemingly being ignore[d]", and he suggests users are being "pushed towards using Pro mode more and more".

Later the same showaemergency returns with an important correction. On his own material — realistic frames styled after vintage film — the model has "actually improved" over the past week. Something else broke: multi-step sequences like a storyboard, where with character sheets attached each frame ends up with only one character instead of several.

themaggielange adds a second line — moderation. According to her, it now triggers on harmless content: "even non-sexual content get „violate" stuff now", and after the October update characters look "creepy looking". Her takeaway is practical: "I cant continue my work, since its so bad".

⚖️ Where they agree and where they diverge

The common denominator is narrow but persistent: several people independently tied the changes to late October, and all complain about the same thing — references and character sheets aren't being held, and edits aren't applied. At the same time, no one provides model versions, request IDs, or logs — only descriptions of their own workflow.

The real point of conflict is that the "regression" isn't universal. showaemergency explicitly says that single realistic generation improved, while serial generation broke. Plus, three different complaints live under one headline in the thread: image count, adherence to references, and filters. The hypothesis about pushing Pro mode is one participant's guess, backed by nothing; there's no response from OpenAI in the comments. So there's no consensus even on the diagnosis.

🧪 What to check yourself

• Compare the same prompt in regular and Pro mode (if you have quota): look not at "beauty" but at the number of images delivered — that's a direct test of the Pro hypothesis.
• Ask for a series of 5–10 images and count how many actually came back; repeat the request twice and compare.
• Take an existing character sheet and order a 3–4 frame storyboard: count how many different characters appear in each frame.
• Make a control pair "reference + one edit" and check whether a new file arrived or the same one; log the date and settings.
• Separately run a neutral query to test for false filter triggers and record the wording that triggered it.

🔗 Sources

OpenAI Community — ChatGPT Image Generation Has Gotten Significantly Worse Since October 29th — Is Anyone Else Experiencing This?

Disclaimer: this is a summary of a discussion, not facts verified by the editorial team. Opinions belong to the participants; the numbers belong to those who cited them.