Normalizing Flow Learns CelebA Faces

This video shows samples from a normalizing flow using a modified (shrunk) GLOW structure throughout the first 350 training epochs on the CelebA dataset. Temperature of the samples increases linearly from left (0.5) to right (0.9).

The samples clearly show faces, but still display significant noise.

I produced this as a preparatory exercise for a project on neural spline flows.