NanoFlow: Scalable Normalizing Flows with Sublinear Parameter Complexity

Compared with the variational autoencoder [18] and generative adversarial network [10], NFs exhibit the distinct characteristic of an exact probability density estimation from a principled maximum likelihood training.