AI

DeCoPrune: Efficient KV-Cache Pruning for Autoregressive Video Diffusion via Denoising Consistency

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 07/10/2026, 19:18:24 UTC.

Briefings

DeCoPrune is a training-free method for autoregressive video diffusion that uses denoising consistency to prune KV cache tokens. By measuring the discrepancy between intermediate clean predictions and final denoised values, the method retains high-discrepancy tokens while pruning low-discrepancy ones. Experiments on LingBot World v2 show DeCoPrune preserves near-FullKV long-range recall while pruning over 85% of historical KV tokens and accelerating continuation generation by over 4 times.

Key points

• DeCoPrune is a training-free method that treats KV cache compression as a denoising-consistency problem.
• The method uses the discrepancy between intermediate clean predictions and final denoised values to identify tokens to retain.

Context

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. DeCoPrune addresses this by using denoising consistency as a proxy for token value.

Limitations

The evaluation relies on LingBot World v2 and uses CMBench with 58 episodes and 116 continuation tasks. Results are based on empirical findings from a single project homepage and GitHub repository; no independent validation or broader dataset comparisons are provided in the evidence.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 01/11/2026, 10:34:08 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeigj33vs7s2plgvbgq7dcfl7lmy3d2zzltqjubr5rrplvhorhh7s3m