DeCoPrune Optimizes KV-Cache for Autoregressive Video Diffusion
October 4, 2026
DeCoPrune is a training-free method that prunes KV caches by using denoising difficulty as a proxy to identify tokens that provide unique visual evidence.
HOW THIS AFFECTS YOU
●
builderYou can reduce memory overhead in long-form video generation without training new models.
●
researcherThis links token value directly to denoising consistency in autoregressive models.