Really cool. You could generate all kinds of interesting art with this.
I can't help but think of people who report seeing faces in their toast. Humans are biased towards seeing faces in randomness. A neural network trained on millions of puppy pictures will see dogs in clouds.
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).
I can't help but think of people who report seeing faces in their toast. Humans are biased towards seeing faces in randomness. A neural network trained on millions of puppy pictures will see dogs in clouds.