Insights from Generative Modeling for Neural Video Compression