Training-Free Efficient Video Generation via Dynamic Token Carving | Read Paper on Bytez