That's because the text part and the image part are two separate systems. ChatGPT can't actually 'see' the pictures. It passes it off to a different system, like Stable Diffusion.

Imagine you are talking to a blind guy in a box, but you don't know he's blind, and you ask him to find you a picture of whatnots. He says 'sure, I can do that, of course' and then looks up in his braile dictionary what whatnots are. After this he knocks on the box next to him where his friend is, who can see, and describes whatnots.

His friend passes him the picture and he passes it to you. It's actually a picture of whatchamacallits.

"This isn't a whatnot. This picture has a dongle, it's supposed to have a dangle."

"Oh, of course, you're so right to call this picture out, let me try again."

And so on. That's why it constantly prompts you to describe more so it can pass fine tuned descriptors to the imager, but it has no idea what the picture is you want or if what he gives you is correct.