That's essentially precisely what's happening here. You can see in the different pictures where different sets of training data were used---buildings, faces, an

That's essentially precisely what's happening here. You can see in the different pictures where different sets of training data were used---buildings, faces, animals. Give the machine millions of reference images to work from and then tell it to find those images in noise, and it will succeed (because it literally can't "imagine" anything else for the noise to be).