Why do you not use pixelCNN for CIFAR experiments?

[sverneka]

[kimiyoung]

Because we did not find it necessary in terms of performance. The reason is as follows: PixelCNN is used in our framework to prevent generating on-manifold samples, but it is hard to generate examples that are "too realistic" on CIFAR-10. Empirically, we did not see substantial gains using PixelCNNs.