Closing the Dequantization Gap: PixelCNN as a Single-Layer Flow

Neural Information Processing Systems 

Flow models have recently made great progress at modeling ordinal discrete data such as images and audio.