DeCoPrune: Efficient KV-Cache Pruning for Autoregressive Video Diffusion via Denoising Consistency
DeCoPrune cuts over 85% of video-diffusion KV tokens while preserving long-range recall.
DeCoPrune is a training-free method that compresses the growing KV cache of autoregressive video diffusion by measuring denoising consistency. Tokens whose intermediate clean predictions differ strongly from their final denoised values are retained, while low-discrepancy tokens are pruned. On CMBench, with 58 roughly one-minute episodes and 116 Reappear or Revisit tasks, experiments using LingBot World v2 keep near-full-cache long-range recall while pruning over 85% of historical KV tokens and speeding continuation generation by more than 4x.
- DeCoPrune is a training-free KV-cache pruning method.
- It keeps tokens with high denoising discrepancy in the long-term cache.
- CMBench has 58 one-minute episodes and 116 recall continuation tasks.
- On LingBot World v2 it prunes over 85% of historical tokens and speeds generation over 4x.
From the original article
1 min read at huggingface.co
Autoregressive video diffusion supports streaming generation and interactive control, but its KV cache grows continuously with the generated history. Existing compression strategies either discard history using fixed windows or select tokens through local attention and similarity signals, which do not directly measure whether the current chunk contributes information beyond the retained context.
By Zeqi Xiao, Qingle Liu, Kaiwen Zhang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.