Sparsity Should Be ALLOTted by Output: Training-Free Allocation of Block-Sparse Attention for Video Diffusion Transformers

One-line takeaway

KlingAI's training-free ALLOT allocates block-sparse attention in video DiTs by estimating each layer-step cell's contribution to final-video error, exploiting that the most sensitive layer-step pairs stay stable across inputs.

可灵团队的 ALLOT 无需训练，为视频扩散 Transformer 分配块稀疏注意力：发现不同层和去噪步对画质影响差异大且跨输入稳定，于是按对最终视频误差的贡献联合优化各层各步的稀疏度。