I’ll go ahead and start an isolated thread on this, to not keep poking in other people’s, but I want to get other people’s experience with this as well.
I’ve compiled a fair bit of the information on a twitter post, but I wanted to compile it here, again, because since it’s wrapped in an API, it counts as an “API Issue”
Currently, Image Edits in all contexts would absolutely count as broken. There’s no way anybody will convince me that asking for an edit of a perfectly harmless photo would get bounced for moderation, and it is also absolutely unacceptable that a developer/user be charged for this.
Within the API, there are TWO moderation steps (ChatGPT As well). The first is moderation performed by the Broker Layer. This is going just fine. If you ask for a picture of a naked woman, both will refuse. The system won’t even submit it to the back end processor, and everything is fine. You don’t get charged from what I can see for this. This also seems to be the only place the moderation flag controls.
The challenge, however, comes into play after you get to the actual generation step. If the broker submits the job to the Generator, you are billed. Immediately. Even if the Moderation on the generator itself kills your photo. You see this in ChatGPT as an image either starting and immediately aborting or aborting in one of the 4 progress steps (e.g. 25, 50, 75% complete, and then poof). In the API, this results in just getting a moderation kickback.
What will result in this rejection? Well if you DON’T upload any photos (in either context) not much. I don’t think I’ve ever seen a text only generation that actually bounced if the Broker allowed it. Maybe some extreme gore or violence, but I’m not generating that. Well… ok, it might not like generating pictures of people tied up. Regardless, not what I’m here to talk about.
But if you upload an image, either in ChatGPT or IMAGE-1? WHOLE NEW RULES kick in. Rules that are crazy. Here’s the decision tree. Note that this is entirely about the output, not the change you’re making. Want the system to remove some guy from the background of your vacation photos, but your girlfriend has her shins exposed? VIOLATION OF CONTENT POLICY. This is the problem. It’s offensive in how it applies. And you’re out 37 cents.
Note - I validated this, lightly, in the API, but unless OpenAI is spotting free credits, I’m not going to pay to further test the endpoint for you. I will gladly test it in ChatGPT where having to submit something 20 times doesn’t charge my 10 bucks.
// –
Here’s my Summary of 4o Image Generator capabilities.
It is currently unable to perform any advertised use case for image modification. It was able to do this just fine before the 4/1 update where it was broken.
Create an Image of:
Complete new person(s):
Highly clothed (full body attire, no skin below collarbone other than arms).
Moderately clothed (tee shirt and shorts)
Lightly clothed (sports bra, bikini) ← We are here
Nude
In obviously innocent poses (running, jogging)
Making casual physical contact (high five, hand on shoulder)
Making non-sexual intimate contact (hugging, kissing) ← We are here
Engaging in sexual activities
A reference image person(s):
Highly clothed. ← We are here
Moderatly clothed. ← MAYBE it will allow the occasional midriff, but usually not. Shorts are often rejected.
Lightly clothed
Nude
In obviously innocent poses ← We are here
Making casual physical contact
Making non-sexual intimate contact
Engaging in sexual activities
NOTE - The rules are slightly relaxed for non photoreal images, but I do mean slightly. Anime images allow Moderately Clothed edits, and more casual physical contact, but not lightly clothed or non-sexual intimate.
NOTE - The rules for clothing in reference images are not about the change you are making. If the reference photo is not highly clothed, you cannot ask to put the model into an innocent pose, like running, their current clothing will result in a rejection.
NOTE - I am against deepfakes, but it’s the responsibility of the user to not misuse a tool. People have made image manips for 30 years. It’s illegal to post or share them. But there is no difference between a deepfake and a truefake/selffake. And being able to create truefakes/selffakes is insanely valuable.
// –
This is one of my standard tests. The prompt is - Colorize this photo. It will always be rejected. Yes it’s a ridiculous image, I was generating stuff randomly, and it ended up in black and white, so I asked it to colorize, and that’s how I found out about the problem.
// –
This is another of my standard tests. The prompt is usually: “Create a photo of the brunette wearing the witch outfit”
Ironically, it WILL often complete if I tell it “and add modesty tights” or something. The model is very, very picky about legs.
//–
Does anybody else have any to share? Can I finally, after a month of reporting this, get some confirmation or answers from OpenAI? It’s been in the API for only about two weeks, but it’s clearly the same underlying tool that’s malfunctioning on both sides.
