Modelling video as a series of frames seems like such a waste; and a great point of focus for optimisation. The vast majority of video content has a lot of red

Modelling video as a series of frames seems like such a waste; and a great point of focus for optimisation. The vast majority of video content has a lot of redundant inter-frame information. De-duping this is a key part of most compression schemes and (as an AI simpleton) seems like on obvious entry point for minimising token usage. Or is this simply a case where token windows are expected to / have already grown to a point where this sort of optimisation is not needed?