Reproduce “NanoKV: High-Ratio KV Cache Compression 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

Jiacheng Lu, Haoyi Zhu, Sipei Yi, Enze Xie, Cheng Zhuo, Yu Li

NanoKV compresses autoregressive video generators’ retained Key-Value cache by storing shared cluster centers and compact token residuals, with codecs fitted to the frozen generator’s downstream behavior. On HY-WorldPlay at nominal 32× coefficient compression, it achieves 26.09× effective compression versus KVTC’s 13.50× and PSNR of 19.56 versus 18.56 dB, measured against matched Original rollouts, not ground truth.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.