Video decoding is resource intensive. We’re used to it, we have hardware acceleration for some of it, but spewing something around 52 million pixels every second from a highly compressed data source is not cheap. I’m not sure how both compare, but small LLM models are not that costly to run if you don’t factor their creation in.

I mean, it would. For example Jellyfin implements it, but it does so by extracting the pictures ahead of time and saving them. It takes days to do this for my library.

It is useful for internet streams though, not really for local or lan video.