Token Radius Attention for Efficient Video Generation

Researchers introduced Token Radius Attention, a training-free sparse attention method for video diffusion transformers.

It estimates per-token attention needs from entropy and turns them into query-centered sparse radii instead of dense 3D self-attention.

Across Wan2.1, Wan2.2, and HunyuanVideo text-to-video and image-to-video tests, it kept only a small fraction of attention links.

The method delivered notable speedups while maintaining competitive video quality, reducing the compute cost of generation.