Autoregressive video generation models often require substantial memory for the key-value (KV) cache, which stores intermediate representations. Token merging offers a way to cut this cost by clustering temporally redundant tokens and keeping only a small set of representatives.
The paper introduces a calibrated cache compression approach that uses offline information to improve online generation. While existing token-merging methods have been used for this purpose, the authors argue that their offline-calibrated method makes online performance better. The abstract does not detail how the calibration is performed or how much memory is saved.
The significance is in reducing memory consumption during video generation, which could help scale models to longer or higher-resolution outputs. However, the abstract stops short of stating these benefits explicitly, so the practical gains remain to be seen.