As more and more blog posts I see cram in some generative AI to make their point or keep the prose bit colourful, I was wondering if it would be possible to give a model an image and retrieve the prompt back so that when that same prompt is passed to the corresponding model, it generates that image, pixel perfect.

I am wondering because that's how whole generative AI probably started with Generative Adversarial Networks (GANs) and I am thinking if this reverse engineering of an image to prompt is explored, what it might lead to?

PS: And interestingly enough, if it is extended to reverse engineer from an answer of an LLM to what the actual question/prompt was?