slim fable
Hello everyone,
We are working on using gpt-image-2 to generate previews of the user wearing some wardrobe.
Unfortunately almost all requests are rejected because of safety_voliations=[sexual]
We supply 2 images, one of the dress and one of the user selfie, and they are not explicit in any way.
I already tried setting moderation to low, but it didn't help
Any suggestions or workaround for this?
Interestingly doing similar in the chatgpt app works fine and produces desired output.
scenic solar
Hello everyone,
We are working on using gpt-image-2 to generate previews of the...
Hi, there. Did you try asking ChatGPT for a prompt? Or to analyze your use case and troubleshoot?
slim fable
Yes, I run codex to iterate on the prompt in a loop without acceptable results. It's either failing or modifying dress to unacceptable levels
scenic solar
I understand. But did you ask ChatGPT? You mentioned that you had no issues with ChatGPT, so I'm guessing it's doing something that's helping avoid the safety violation.
slim fable
The thing is that in chatgpt generated images didn't need any guidance to be generated and they were IMO looked more "sensitive" like for a mini skirt attempt (maybe not that mini but definitely shortest from examples that I tried). So for me it looks like chatgpt uses much more relaxed filtering that is just not available over api.
scenic solar
I can't speak to different sensitivity levels but ChatGPT adds a layer of interpretation to user prompt before it's sent to Image 2.0. I was trying to see if you can get it to explain how to avoid triggering the filter.
slim fable
As far as I understand this is the same as using responses API for this. I tried this with chat-latest, gpt-5.5 all reasoning levels and all permutations of available options.
It looks like API natively (js and python libs) does not support setting moderation param, but codex stuffed it as an extra param. Maybe it's ignored on the backend?
It is present in docs
scenic solar
Makes sense Codex image gen is the same as API. ChatG is definitely different. Have you tried it via API?
slim fable
I mainly tested via API. I gave the codex API key to experiment. Codex claims that it cannot edit images, only generate them
steady crystal
just ask gpt do it again until is work and when i start to argue open a new chat and repeat it i will finish to work since your request isn't illegal at all