DySink: Dynamic Frame Sinks for Autoregressive Long Video Generation

Researchers introduced DySink, a retrieval-based method for long autoregressive video generation.

DySink keeps a compact memory bank, retrieves visually relevant past frames as dynamic sinks, and uses an anomaly gate to block collapse-prone context.

On minute-long video tests, it outperformed strong baselines on dynamic degree and temporal quality.

The approach tackles a key long-video bottleneck by making long-range context more adaptive and reducing outdated or collapsed visual outputs.