In video generation, the computational cost of self-attention is proportional to the square of the number of tokens; after VAE encoding and patching, a 4-second 720p 24fps video can yield tens of thousands to over a hundred thousand tokens in total.