Tell HN: OpenAI added heavy censorship on GPT4o image generation yesterday | Hacker News Reader

Tell HN: OpenAI added heavy censorship on GPT4o image generation yesterday

7 pointsaurareturn12 comments

It won't let you use famous people anymore.

It won't even let you use any real people. If you upload a picture of yourself and tell it to generate a Studio Ghibli style picture, it'll tell you it can't use real people.

This is the response I see:

"I wasn’t able to generate the image you requested because it didn’t follow our content policy. Specifically, generating images of real people—even in a stylized form like Studio Ghibli—is not allowed."

View on HN

> our content policy

This is probably why commercial offering will fail apart from some niches they might be able to fill.

Not that generating people would be something many would use, but the act of lobotomizing the AI will have an effect on any other content they will generate.

I just tested this with a picture of me sitting at a restaurant and told it to draw a picture of Ice Cube (the rapper) sitting behind me.

What it won’t do is draw a realistic picture of me in a superhero costume and says it violates the policy. I didn’t name the hero, but I described the costume.

Second anecdote: Earlier today, I sent a friend a ChatGPT modified picture of him wearing a MAGA cap and standing by Kanye West with Kanye’s arm around him.

I tried modifying my own picture with a MAGA hat and it wouldn’t let me. I could add another random hat with a logo.

I don't understand why this is varying so widely person to person. I was just able to use a photo of a famous athlete and a photo of myself and get them to shake hands in a photo of a location where neither of us have ever been.

Some people I work with are having similar experiences to you though, but it's wildly inconsistent.

Perhaps the person who can't create good results is using the free (basic) ChatGPT.

I thought this might be the case but checked it and nope, everybody paid doing the exact same thing and some getting different reasons why it won’t work and some getting actual output.

I’ll admit we didn’t test exact same word for word prompts though.

Should it be really that simple to create pictures of other people without any kind of concent just by uploading couple of images?

Imaging if you're actually famous / politically exposed person and some company publishes a tool that can generate any video, voice and pictures with you or sell some courses narrated by you.

I think OpenAI is doing the right thing as far as reputational risk.

But on the other hand, the cat is out of the bag. Someone will release an unconstrained model and eventually this will run on affordable consumer hardware.

Having a non-profit org run statical analysis on terrabytes of pirated content and then sell itself to a big corporation doesn't sound like a right thing to me.

This is where you ask ChatGPT to describe the Studio Ghibli art style and then use that to create a prompt. It's an annoying extra step but it can get good results.

It's not about the art style. It's that it detects real people in the photo and refuses to do anything with it even if the people are not famous.

Trying to make them in Simpson style results in possible copyright infringement. I gave up after waiting for many minutes and seeing constant rejections.

Gooooood.

Reply on news.ycombinator.com