The Sphere Encoder is a new way to make images fast by teaching an autoencoder to place all images evenly on a big imaginary sphere and then decode random spots on that sphere back into pictures.
Normalizing Flows are models that learn how to turn real images into simple noise and then back again.