Conceptual

Dynamics-Aware Token Pruning for Diffusion Model Acceleration

A training-free method (DaTo) that unifies feature caching and token pruning to accelerate diffusion-model inference. Observing that feature caching makes activations across adjacent denoising timesteps too similar and flattens their temporal dynamics, DaTo selectively prunes low-dynamics tokens so that only high-dynamics tokens pass through the self-attention layers, restoring feature dynamics across timesteps. It achieves both temporal and token-wise reuse, delivering roughly 9x speedup on Stable Diffusion while improving FID.