DeCoPrune Prunes 85% of Video Diffusion KV Cache Training-Free, 4x Faster Continuation

mmlab-ntu · hf · 2026-10-08

NTU MMLab introduces DeCoPrune, a training-free KV-cache pruning method for autoregressive video diffusion models.

Problem: KV cache grows unboundedly with generated history; existing fixed-window or similarity-based token selection can't directly measure whether a new chunk adds information beyond retained context.
Method: Frames cache compression as a denoising-consistency problem — empirically, denoising difficulty proxies a token's long-term retention value. Tokens with larger discrepancy between intermediate clean predictions and final denoised values carry less predictable visual evidence and are kept in long-term cache.
Benchmark: Releases CMBench with 58 1-minute context episodes and 116 Reappear/Revisit continuation tasks testing recall of previously observed objects/scenes.
Results: On LingBot World v2, prunes over 85% of historical KV tokens while preserving near-FullKV long-range recall and accelerating continuation generation by 4x+, substantially beating compression baselines.

More from Research