Agentic data science with the Data Agent Kit
cloud.google.com·6h ago
AI Research
arXiv CS.AI·Aug 13·5 min read
LoSA: Near-Lossless Sparse Attention for Training-Free Video Diffusion Acceleration
TL;DR
LoSA introduces a training-free sparse attention method for video diffusion transformers, achieving a 1.36x speedup with minimal quality loss. By retaining 99% of attention mass, it optimizes computational efficiency without sacrificing fidelity.
✦ Why It Matters
Implement LoSA in your video diffusion projects to enhance performance without retraining.
Key Takeaways
- Adopt LoSA for efficient video diffusion without retraining.
- Utilize the 99% attention mass threshold for optimal performance.
- Combine LoSA with feature caching for maximum speedup.
- Evaluate performance on datasets like Wan2.1-1.3B and HunyuanVideo.
How It Works
LoSA operates by fixing a threshold for retained attention mass at 99%, rather than focusing on a specific sparsity ratio. It measures block attention masses during an initial dense step and retains only the necessary key/value blocks for each attention head and query block.
This allows for significant reductions in computation while ensuring that the majority of attention mass is preserved across denoising steps.