This similarity of model output to training data, by pixel to pixel similarity, is often included in ML papers. The GAN here will very likely be producing uniqu

This similarity of model output to training data, by pixel to pixel similarity, is often included in ML papers. The GAN here will very likely be producing uniques images in each case, some will be closer to the training data. But in general GANs are very capable and can for instance imagine a person with red hair wearing a green jacket even if it has never seen this combination of things before.