←back to thread

67 points arduinomancer | 3 comments | | HN request time: 0s | source

I’ve noticed a lot of the time you can tell an image is AI generated because it has a shiny/glossy lighting look to it.

Has anyone figured out why this is the case?

Show context
osigurdson ◴[] No.41263216[source]
I tend to agree. However, I tried to continue to prompt ChatGPT make make the picture less "AI like" and it actually ended up doing a really good job after 5 or 6 attempts. I'm not sure why it took so much prompting. Further prompting just made it worse.
replies(1): >>41266161 #
1. DemocracyFTW2 ◴[] No.41266161[source]
Be it said that ChatGPT is not ideally suited for this kind of inquisitive work because it will re-write your prompt as it sees fit so without checking back what the actual prompt was (e.g. on the Bing Image Creator page) you can never be sure what it added, took away or modified from your input before it hit image generation proper.
replies(1): >>41266195 #
2. osigurdson ◴[] No.41266195[source]
Are you sure? I know from using the OpenAI API, that the entire chat history is part of the request. I suspect the ChatGPT UI likely does some processing for efficiency but the core model is stateless so the entire list of messages needs to be passed.
replies(1): >>41266580 #
3. DemocracyFTW2 ◴[] No.41266580[source]
I'm not sure whether we're on the same page—when you ask the chat interface to generate an image, the language model will / may / can rewrite your wording, making it even more difficult to link cause and effect with that in the pipeline. IOW the verbiage you use in the chat interface is not the input the image generator gets to work on. That's how far I know the system, please correct me.