This represents such an entirely wrong idea of what’s going on that it’s almost hilarious. The model doesn’t just have an internal database of all the training data that it, like, collages together. Holy fuck, if a model internally contained all of the training data in a form it could simply be retrieved from it would be far more valuable as data compression than anything else it might be or do.

What’s much closer is that an image generation model has seen millions of pictures, each labeled with metadata about what it is a picture of, each of which causes tiny adjustments to the neural networks weights, because the model is being trained to be able to look at a novel picture and recognize “this picture is of a dog pissing on an apple tree”. Any model that can do that can also be run in reverse - essentially “here is a block of white noise I am saying is supposed to be a really noisy picture of a dog pissing on an apple tree, try to clean up the noise and show me how it looks like a dog pissing on an apple tree”, then run that in a loop until the picture doesn’t change much between iterations. It doesn’t need to have seen something precisely identical to what it’s generating - with enough training data it has an internal model for things like dogs, clowns or what makes something look like anime and the generation just tickles the bits until the picture rates as probably being what’s being asked for.