2025/08/28 by Shengqu Cai, Ceyuan Yang, Cai, Shengqu +24 · 3 voices · 28 citations
Computer Science · #Artificial Intelligence (cs.AI) #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #Generative Adversarial Networks and Image Synthesis #Graphics (cs.GR) #Video Analysis and Summarization #cs.AI #cs.CV #cs.GR
paper · pdf · doi:10.48550/arxiv.2508.21058
openalex publication_date 2025/08/28 · arxiv published 2025/08/28 · openalex created_date 2025/10/10 · arxiv updated 2025/12/09 · openalex updated_date 2026/07/28
Long video generation is fundamentally a long context memory problem: models must retain and retrieve salient events across a long range without collapsing or drifting. However, scaling diffusion transformers to generate long-context videos is fundamentally limited by the quadratic cost of self-attention, which makes memory and computation intractable and difficult to optimize for long sequences. We recast long-context video generation as an internal information retrieval task and propose a simple, learnable sparse attention routing module, Mixture of Contexts (MoC), as an effective long-term memory retrieval engine. In MoC, each query dynamically selects a few informative chunks plus mandatory anchors (caption, local windows) to attend to, with causal routing that prevents loop closures. As we scale the data and gradually sparsify the routing, the model allocates compute to salient history, preserving identities, actions, and scenes over minutes of content. Efficiency follows as a byproduct of retrieval (near-linear scaling), which enables practical training and synthesis, and the emergence of memory and consistency at the scale of minutes.