How does STA handle edge cases, such as videos with sudden scene changes or complex motion, where locality might not be as strong?

Sliding tile attention accelerates Hunyuan video generation by 3x with no quality drop and no need for training