Reproduce “TempoKV: Predictive KV Cache Quantization for Autoregressive Video Generation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jiahui Hu, Weile Mo, Peiliang Cai, Jiacheng Liu, Jinkui Ren, Xiantao Zhang, Xiaobing Tu, Linfeng Zhang
TempoKV reduces KV-cache storage for long-video generation by predicting cache entries from history and storing low-precision corrections. It predicts keys from earlier keys and values from value history plus the aligned current key. On Self-Forcing with MovieGen prompts at 832×480 for 717 frames, k5v3 achieves 3.65× compression and Visual Quality 74.04, below BF16’s 74.23; compressed decoding adds latency.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.