DySink: Dynamic Frame Sinks for Autoregressive Long Video Generation

Researchers introduce DySink, a novel framework for autoregressive long video generation that dynamically selects relevant historical frames instead of using static early-frame anchors. The method addresses the problem of outdated context degrading video quality and introduces a sink anomaly gate to prevent content collapse, demonstrating improvements in temporal consistency for minute-long videos.

DySink represents a meaningful advancement in addressing a fundamental limitation of current autoregressive video generation systems. Traditional approaches use fixed memory allocation strategies that cache early frames as anchors for long-range context, but this rigid approach becomes problematic as generated video content diverges from those initial frames. The framework's dynamic retrieval mechanism fundamentally rethinks how systems maintain long-range dependencies, selecting frames based on actual visual relevance rather than temporal position.

The technical innovation directly tackles two distinct failure modes in current systems. The first involves context bias toward outdated cues that no longer align with current generation state. The second, more severe issue involves RoPE-induced phase re-alignment causing inter-head attention homogenization and sink collapse—where generated content regresses toward the cached early frames rather than progressing naturally. By coupling adaptive retrieval with an anomaly detection gate, DySink suppresses these collapse-prone contexts while maintaining efficient memory usage through a compact bank.

For the AI video generation ecosystem, this work signals an important maturation beyond baseline streaming approaches. Developers and organizations building long-form video systems will benefit from more temporally coherent outputs, particularly for extended sequences where traditional sink strategies degrade. The improvement demonstrates that memory allocation strategies deserve the same optimization attention as core architectures.

The release of code and weights accelerates adoption among researchers and practitioners. Future work likely explores whether these dynamic selection principles apply to other sequence modeling tasks beyond video, potentially influencing how foundation models handle long-context generation more broadly.

• →DySink replaces static early-frame anchors with dynamically selected relevant historical frames for improved video coherence
• →A sink anomaly gate detects and suppresses context that causes output collapse and regression toward cached frames
• →Experiments demonstrate consistent improvements in temporal quality and visual diversity for minute-long video generation
• →The framework efficiently maintains long-range dependencies through adaptive retrieval rather than fixed memory allocation
• →Open-source release enables broader adoption among video AI development teams