But that’s not the default behaviour of these models at all. Instead you have to work pretty hard to extract these original images.

When using the models normally, they do generate new content, even if the style and subjects are of course interpolated from training data.

If you deliberately try to make it do so. Even then it usually doesn’t work.

It can do that for some images, and that's a training mistake.

There's no way it can do that for more than a tiny fraction.

"we bias our search towards duplicated training examples because these are orders of magnitude more likely to be memorized than non-duplicated example"