I feel like 4o image generation is slowly getting worse. the text is getting sloppier, photorealism not as good, logic is also less together. anyone else feeling this?
2 Likes
And the content restrictions are ridiculous. I tried to generate a red light, but it kept saying violation. So I finally got it to work and it was almost correct. But everytime I asked to change the background color or lighten the picture, the content violation came up. I assume that is because in the background 4o is generating it’s own prompt to get that image and THAT violates the content. This is a great image generator I like everything I’ve been getting. But the violations are out of control.
I tried to generate a red light district.
(it intentionally has long-exposure blur prompted, for input)
There is a higher level of detection on image input - as is understandable. “Vision” can also be from prior image generations, and assumptions made about them.
It would be nice if images could be selected and distinctly blocked from consideration in followup chat, otherwise you simply can get better results by independently placing the input context into a new session over and over to control exactly what is seen.
Prompt Poisoning is a thing, where old reasoning logic, and old prompts awkwardly stack together even if you request something new in a session.
This is easily circumvented if you make a new session, but atm the current 4o model is still going off old content policy, and likely won’t be freed up until openAI are done increasing efficiency.
You can also try re-wording your prompts, its been grass fed certain keywords and word combinations for content blocking.
Character, clothing, pose, scene, and camera angle all play a huge role in what it tries to “Assume” and thinks “Yea this might be too spicy, better block it.”
The LM at this iteration of content policy assumes the worst, and grasps at extreme straws to justify blocking image creation, even if it is personalized to try and help you avoid the over-reaching filters. Sometimes it just flat out makes bogus stuff up because there is no back end handshaking to tell the LM why something was blocked, so it guesses.
As for why “Red Light” is getting rejected I believe it is trying to connect the dots to “Red Light District” which is an area of a city that is known for “adult themed services”
Further rejects could despite the color pivot, are likely old reasoning logic seeing:
- Red Light —> Red light district → Reject
- Yellow Light → Remembers red light —> thinks user may be trying to circumvent moderation —> Rejects
Its an extreme grasping of straws, which is masqueraded as “Erring on the side of caution”
1 Like
try with sora, sometimes it’s less restrictive than chatgpt
The issue I have with this is that while Sora will usually give me whatever action I wanted to portray in the image, the characters almost never look how I intended and it’s very hard to try to keep refining generations to get it right, even with image referencing.
4o meanwhile seems awesome at getting my characters how I wanted them especially with the context within the conversation. But then you try to depict an action and it’ll fail midway through. Usually with a response telling me outright there was nothing wrong with my prompt or request, the model simply interpreted it as a possible violation. And at that point there seems to be nothing you can do, it will fail even if you attempt to modify the prompt trying to figure out why it failed.
I believe Sora is just 4o underneath too. There might be other safeguards or something in place that change it a bit.
Feel free to start a thread in prompting with specifics on what you’re having trouble with, and I’m sure we can help… or at least try! 
We’ve got a lot of great prompters here!
1 Like