PixelCNN++ improves PixelCNN with discretized logistic mixture likelihood and other modifications

AI Impact Summary

PixelCNN++ introduces a discretized logistic mixture likelihood to model pixel values more accurately than the standard PixelCNN, along with additional architectural tweaks. This can improve density estimates and sample realism on image datasets, but increases computational complexity for training and inference due to the mixture computations. Teams should reassess data preprocessing (pixel quantization), tuning the number of logistic mixtures, and potential changes to hyperparameters and training time to realize the fidelity gains.