Where does processing keyframes and non-keyframes come into play in the video compression/encoding pipeline?

There's a lot of recent literature/attempts at compressing using small deep nets which replace handcrafted encoders