Auto-regressive Generative Models
PixelRNN — Recurrent NN which predicts pixel categories given the previous pixels.
PixelCNN — Output pixels are represented as 256 categorical variables. Use a masked convolution to include only preceding pixel information. Faster but lower accuracy than
PixelCNN++ — Instead of representing output pixels as categorical variables, represent output pixels as a set of distributions parameterized by parameters produced by the network.