Last released Aug 19, 2026
Training-free sparse-attention acceleration for video diffusion and world models