Autoregressive video models generate successive portions of a video using cached keys and values, internal representations that help retrieve relevant information from earlier scenes.
Video generators can nearly halve history storage while improving reconstruction quality
Across three video models, the authors report better quality-storage trade-offs, with nearly twice a competing method’s compression and a 1.0-decibel quality improvement on HY-WorldPlay.
Big Tech
Jiacheng Lu · Haoyi Zhu · Sipei Yi · Enze Xie · Cheng Zhuo · Yu Li
Zhejiang University · NVIDIA
Research Digest··2 min read
Before you read this
Three ideas this paper builds onkv cache compressionReducing the size of the key-value cache in transformers via selective retention, low-rank updates, or adaptive compression.61 papers →autoregressive video generationGenerates video frames one at a time, each conditioned on previously generated frames.5 papers →role decouplingSeparating distinct functions such as planning, reasoning, and execution into separate components improves efficiency, quality, and reliability.46 papers →
Lu and colleagues present NanoKV, a way to compress the stored history that autoregressive video models use to maintain scene consistency.
How it works
before and afterThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.