SemanticKV compresses KV cache for agentic video generation
Original title: Semantically Aligned KV Cache Compression for Efficient Agentic Autoregressive Video Generation
Authors report a 3.35x KV-memory cut and 1.22x speedup on Causal Forcing at compression ratio 0.7.
Researchers introduce SemanticKV, a training-free KV-cache compression method for agentic autoregressive video generation. It uses planner-extracted semantic constraints and text cross-attention to score cached entries, then keeps a single top-k set without retraining. The authors evaluate four autoregressive video backbones on three benchmarks for 5-, 30-, and 60-second generation. On Causal Forcing at ratio 0.7, they report a VBench Total Score of 84.10 versus 84.05 for the strongest compressed baseline and 83.99 for Full KV.
Unconfirmed: Scores and speedups are author-reported and not independently verified.
- SemanticKV leaves retained keys, values, and positions unchanged.
- It runs independently on each candidate without extra video generations.
- Reported KV-cache memory falls from 6.04 GB to 1.81 GB.
- The paper calls SemanticKV the first constraint-guided method of this kind.
From the original article
1 min read at arxiv.orgAgentic autoregressive video generation combines semantic planning, sequential generation, and visual verification for controllable synthesis. Growing visual histories enlarge the key-value (KV) cache, increasing memory use and attention costs across candidate rollouts.
By Weizhong Huang · arxiv.org
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.