DFSAttn: Dynamic Fine-grained Sparse Attention for Efficient Video Generation

Researchers introduced DFSAttn, a training-free sparse attention method for diffusion transformers in video generation.

DFSAttn combines token reordering, hierarchical block scoring, and sparse mask caching with adaptive ratios for fine-grained sparsity.

The method cuts the high cost of full spatiotemporal attention and delivers up to 2.1x end-to-end speedup.

It also maintains or improves quality versus prior sparse-attention methods, especially at high sparsity levels.