Preprint
Aug 2026
SparSTAR: Sparse Attention for SpaceTime AutoRegressive Video Synthesis
SarSTAR is introduced, a training-free block-sparse attention method tailored to this setting, showing that recomputing block selection at each target scale is more reliable than reusing a transferred mask as reuse spans increasingly distant scales.
Jongbeom Lee, Hyunwoo Yu, Jincheol Yang et al.
· 0 citations