Test-time Sparsity for Extreme Fast Action Diffusion

Researchers propose a test-time sparsity method for action diffusion models to cut inference cost.

The approach dynamically prunes residual computation, parallelizes encoding and pruning, and reuses cached features across steps.

It reduces FLOPs by 92% and makes generation about 5x faster, reaching 47.5 Hz.

The result enables much faster diffusion-based action generation without hurting quality, which matters for real-time robotics and interactive systems.