The authors develop iCATS, a training-free method for accelerating attention in Diffusion Transformer (DiT) based video generation.
Interaction-aware sparse attention and timestep-adaptive sparsity accelerate video generation
iCATS achieves up to 2× speedup on video DiTs with minimal quality loss by clustering tokens based on query-key interactions and adjusting sparsity per denoising step.
Chinese Tech
Chengfeng Han · Baole Ai · Xianlu Bian · Jie Yao · Zilong Huang · Ang Wang · +1 more
Hangzhou Normal University · T-Head Semiconductor Co., Ltd. · Alibaba Group
Research Digest··2 min read
Han et al.
Why this paper
From Alibaba Group and 2 others
In one line
iCATS achieves state-of-the-art efficiency-quality trade-off in video generation through interaction-aware sparse attention and SNR-guided sparsity.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors (2 noted)
- ✓Reports numbers on named benchmarks (2 benchmarks)
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§