analiza aiqbench
DySink addresses a critical bottleneck in autoregressive long video generation by replacing static memory anchors with dynamically retrieved frames to prevent content regression. This retrieval-based framework actively suppresses attention collapse caused by outdated context while improving temporal quality and adaptability over minute-long sequences. The approach represents significant progress in managing bounded-memory constraints for high-fidelity generative models without sacrificing efficiency.
DySink: Dynamic Frame Sinks for Autoregressive Long Video Generation
arXiv:2605.21028v3 Announce Type: replace-cross Abstract: Autoregressive long video generation often adopts bounded-memory streaming for efficiency, typically combining local windows for short-term continuity with static early-frame sinks as long-range anchors. However, this fixed allocation keeps early frames cached even when the current visual state has substantially diverged from them, while discarding potentially more relevant intermediate history. As a result, the retained long-range context may become less adaptive and bias generation toward outdated cues; in severe cases, RoPE-induced phase re-alignment can homogenize inter-head attention and cause sink collapse, where content regresses toward sink frames. We propose DySink, a retrieval-based framework that maintains a compact memory bank and selects visually relevant historical frames as dynamic frame sinks. DySink couples adaptive retrieval with a sink anomaly gate, which detects excessive inter-head consensus over retrieved context and suppresses collapse-prone context. Experiments on minute-long videos show that DySink consistently improves dynamic degree over strong baselines while also achieving higher temporal quality. The code and model weights will be released at https://github.com/yebo0216best/DySink.
Abstract:Autoregressive long video generation often adopts bounded-memory streaming for efficiency, typically combining local windows for short-term continuity with static early-frame sinks as long-range anchors. However, this fixed allocation keeps early frames cached even when the current visual state has substantially diverged from them, while discarding potentially more relevant intermediate history. As a result, the retained long-range context may become less adaptive and bias generation toward outdated cues; in severe cases, RoPE-induced phase re-alignment can homogenize inter-head attention and cause sink collapse, where content regresses toward sink frames. We propose DySink, a retrieval-based framework that maintains a compact memory bank and selects visually relevant historical frames as dynamic frame sinks. DySink couples adaptive retrieval with a sink anomaly gate, which detects excessive inter-head consensus over retrieved context and suppresses collapse-prone context. Experiments on minute-long videos show that DySink consistently improves dynamic degree over strong baselines while also achieving higher temporal quality. The code and model weights will be released at this https URL.
Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
arXiv:2605.21028 [cs.CV]
(or arXiv:2605.21028v3 [cs.CV] for this version)
Text integral sub licența deschisă a publicației · 179 cuvinte
Subiecte conexe
DySink: Dynamic Frame Sinks for Autoregressive Long Video Generation
arXiv · 9 iun. 2026
DySink: Dynamic Frame Sinks for Autoregressive Long Video Generation
arXiv · 3 aug. 2026
A kernel-centric path to real-time video generation on Trainium
Amazon Science · 25 sept. 2026
Build a Physical AI model factory with NVIDIA Cosmos 3 on SageMaker HyperPod
AWS · 4 sept. 2026