Given how video is compressed (usually, key frames + series of diffs) perhaps there's some internal optimization leveraging that (key frame: bunch of tokens, di

Given how video is compressed (usually, key frames + series of diffs) perhaps there's some internal optimization leveraging that (key frame: bunch of tokens, diff frames: much fewer tokens)