My guess is it can see that fulfilling your prompt would mean imagining some name in the search field, and it's not allowed to identify people in images like that. Different class of disallowed behavior vs. reproducing celebrity likeness given their name as text
@goodside
-
ChatGPT 4.5 image output and conversation behavior
By
–
You're using ChatGPT 4.5, which doesn't have native image output at all Also my screenshots are two separate chats; initial refusals make it more likely to refuse in the future so you want to start a new one
-
Generative prompting yields mixed celebrity image matches
By
–
As in, vs. closely memorized real photos? I think they’re all a bit of a mix. I tried the GIS prompt for a bunch of celebs and many results are clearly “trying to be” specific photos like album covers but details are often wrong.
-
ChatGPT 4o’s quirky behavior imagining real people
By
–
ChatGPT 4o isn’t quite willing to imagine Harry Styles from a text prompt but it doesn’t quite know it isn’t willing to imagine Harry Styles from a text prompt so if you ask it to imagine being asked to imagine Harry Styles from a text prompt it imagines Harry Styles:
-
Model trained to generate images privately for spatial reasoning
By
–
I’m sure either can do “Draw an owl” style diagrams, but what I mean is a model trained to generate images privately for reasons other than showing them to the user, e.g. to improve spatial reasoning answers
-
Commentary on AI capability frontier and maze generation
By
–
No, they aren’t very good, and they aren’t reliably right either. But the quoted post I’m responding to says, “It absolutely cannot generate valid mazes.” I don’t mean this to be useful, I mean it to show the difficulty of knowing the shape of the true capability frontier
-
RAG and LangChain relevance for updated guide
By
–
I’d guess most of the content I was referring to (RAG, the grab-bag of things LangChain implemented early on) isn’t widely useful now but the guide might still be good where it’s updated
-
Improving prompts and AI topic categorization
By
–
I think you’re right that’s probably important for this prompt working but with one more step you can make it work for paths harder to generate top-to-bottom:
-
Workflow tip for improving AI diagram generation
By
–
The main bottleneck is that it can’t leave things alone reliably so incremental progress on a large diagram is hard. For complex diagrams try having it design smaller components in isolation, then upload all of the best curated output to a new session and ask it to combine.
-
Image-domain chains of thought and ‘chain of images’
By
–
Examples like this are a big part of what I like about 4o. It feels like we’re repeating the themes of the GPT-3 era with shaping chains of thought, just now in the image domain. Eventually we’ll have “chain of images” — “Let’s visualize step-by-step.”