A retrieval-based memory system called DySink aims to stop AI video generators from regressing toward their own opening frames.
Current autoregressive video generators pair a rolling window of recent frames with a fixed set of early frames — so-called "frame sinks" — that serve as long-range anchors. As a video grows longer and the visual scene changes, those anchors stay in memory regardless of relevance. The result is a failure mode the researchers call "sink collapse": the model's attention converges on stale context, and generated content drifts back toward the start of the clip. DySink replaces static early-frame anchors with a retrieval system that selects the most visually relevant historical frames on the fly, paired with an "anomaly gate" that detects when attention is collapsing onto retrieved context and suppresses it before the damage spreads.
Long-form video is where the AI video sector quietly struggles. Tools that produce clean five-second clips degrade fast past the thirty-second mark, and anchor staleness is one underreported reason why. A memory architecture that selects context by relevance rather than by age is exactly the kind of fix that never appears in a demo reel but matters for anything meant to run longer than a TikTok.
The paper reports consistent improvement in what it calls "dynamic degree" over baselines on minute-long clips, and the code and weights are promised on GitHub — which means anyone building on open video pipelines can find out quickly whether those numbers survive contact with real workloads.