Research map / Video / Video generation
Mixture of Contexts for Long Video Generation
Shengqu Cai, Ceyuan Yang, Lvmin Zhang, Yuwei Guo, Junfei Xiao, Ziyan Yang, Yinghao Xu, Zhenheng Yang, Alan Yuille, Leonidas Guibas, Maneesh Agrawala, Lu Jiang and 1 more
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace dense attention with learned top-k context routing (MoC) | LCT 3B MMDiT · multi-shot 8-shot scenes (VBench) | Dynamic Degree: 0.4583 -> 0.5625 | Improved |
| replace dense attention with MoC at 85% sparsity | LCT 3B MMDiT · multi-shot 8-shot scenes (VBench) | attention FLOPs: 1.7 x 10^13 -> 2.3 x 10^12 | Improved |
| replace dense attention with MoC | LCT 3B MMDiT · multi-shot 8-shot scenes (VBench) | Image Quality: 0.5140 -> 0.5003 | Hurt |
| replace dense attention with MoC | 3B MMDiT · single-shot 8s 320x192 (VBench) | Dynamic Degree: 0.6875 -> 0.7500 | Improved |
| replace dense attention with MoC on short single-shot videos | 3B MMDiT · single-shot 8s 320x192 (VBench) | end-to-end speed | Hurt |
| replace dense attention with MoC | Wan-2.1-1.3B · single-shot VBench | Dynamic Degree: 0.4219 -> 0.6250 | Improved |
| add forced intra-shot links (ablation) | MoC · multi-shot forced-link ablation (10k iters) | Dynamic Degree: 0.0208 -> 0.5729 | Improved |
| add forced cross-modal links on top of intra-shot links (ablation) | MoC · multi-shot forced-link ablation (10k iters) | Image Quality: 0.4472 -> 0.5104 | Improved |
| add context drop-in and drop-out (ablation) | MoC · multi-shot forced-link ablation (10k iters) | Background Consistency: 0.9426 -> 0.9579 | Improved |
| reduce chunk size from 256 to 64 or 128 (ablation) | MoC · single-shot ablation (30k iters) | Dynamic Degree: 0.4612 -> 0.3413 | Hurt |
| increase chunk size from 256 to 1024 (ablation) | MoC · single-shot ablation (30k iters) | sparsity: 83% -> 35% | Hurt |
| use top-k 1 instead of 3 (ablation) | MoC · single-shot ablation (30k iters) | Dynamic Degree: 0.4612 -> 0.1313 | Hurt |
Closest papers by meaning
- AdaVideoRAG: Omni-Contextual Adaptive Retrieval-Augmented Efficient Long Video Understanding (NeurIPS 2025 · similarity 0.89)
- OmniMem: Scalable and Adaptive Memory Retrieval for Long Video Generation (NeurIPS 2026 · similarity 0.88)
- SALOVA: Segment-Augmented Long Video Assistant for Targeted Retrieval and Routing in Long-Form Video Analysis (CVPR 2025 · similarity 0.87)
- ReWind: Understanding Long Videos with Instructed Learnable Memory (CVPR 2025 · similarity 0.87)
- Vgent: Graph-based Retrieval-Reasoning-Augmented Generation For Long Video Understanding (NeurIPS 2025 · similarity 0.87)
- Temporal Chain of Thought: Long-Video Understanding by Thinking in Frames (NeurIPS 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
