Machine learning is getting pretty crazy but it's still not quite there sometimes. Especially when you use it in the wrong way.
A Variational Autoencoder is a neural network architecture that can be trained on a batch of similar looking images to generate new, but similar looking images. I trained a Variational Autoencoder on a sequence of images of pixels filling in a square - basically two opposite images and every stage of difference in between.
A total of 8 different filling in sequences were tested. Horizontal left to right, horizontal right to left, vertical left to right, vertical right to left, and each of their color inverses. I had no idea what to expect. My stupid human brain kind of secretly hoped it would make sense. Like a simple average of the images. Guess what machine learning is pretty complicated. I ended up with some weird gradients. I'm pretty amused.