Research map / Video / Video generation
Dual-Granularity Memory for Efficient Video Generation
Hongjun Wang, Lin Liu, Jianguo Li, Tao Lin
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 |
|---|---|---|---|
| add sink columns (Nsink=3) to chunked GSTPN (ablation) | GSTPN WanVideo-1.3B · VBench quality score | Quality Score: 79.1 -> 83.5 | Improved |
| increase sink columns from 3 to 4 (ablation) | GSTPN WanVideo-1.3B · VBench quality score | Quality Score: +0.1 | No clear effect |
| add boundary buffer (Nbuf=2) (ablation) | GSTPN WanVideo-1.3B · VBench quality score | Quality Score: +0.7 (to 83.5) | Improved |
| increase boundary buffer from 2 to 3 (ablation) | GSTPN WanVideo-1.3B · VBench quality score | Quality Score: +0.1 at 50% more memory | No clear effect |
| combine sink columns and boundary buffer (ablation) | GSTPN WanVideo-1.3B · VBench quality score | Quality Score: 79.1 -> 83.6 | Improved |
| add second and third scanning orientations (WTH, HTW) (ablation) | GSTPN · VBench quality score | Quality Score: 80.2 -> 83.5 | Improved |
| reduce chunk size from 200 to 100 (ablation) | GSTPN · VBench quality score | Quality Score: 83.5 -> 84.2; latency 67s -> 80s | Mixed |
| increase chunk size from 200 to 400 (ablation) | GSTPN · VBench quality score | Quality Score: 83.5 -> 81.9; latency 67s -> 58s | Mixed |
| raise LCaM retrieval top-K from 1 to 3 (ablation) | LCaM · Vision Reward | Vision Reward: 0.042 -> 0.052 | Improved |
| raise LCaM retrieval top-K from 3 to 5 (ablation) | LCaM · Overall quality | Overall Quality: 84.8 -> 84.2 | Hurt |
| use similarity threshold 0.3 instead of 0.1 (ablation) | LCaM · Quality score | Quality Score: 83.9 -> 84.8 | Improved |
| raise similarity threshold from 0.3 to 0.5 (ablation) | LCaM · Quality score | Quality Score: 84.8 -> 83.8 | Hurt |
Closest papers by meaning
- Recurrent Autoregressive Diffusion: Global Memory Meets Local Attention (ECCV 2026 · similarity 0.88)
- Learning-to-Memorize: Dynamic Context Management for Long-Horizon Autoregressive Video Generation (NeurIPS 2026 · similarity 0.88)
- Snap Video: Scaled Spatiotemporal Transformers for Text-to-Video Synthesis (CVPR 2024 · similarity 0.87)
- StreamingT2V: Consistent, Dynamic, and Extendable Long Video Generation from Text (CVPR 2025 · similarity 0.87)
- Attend Locally, Remember Linearly: Linear Attention as Cross-Frame Memory for Autoregressive Video Diffusion (NeurIPS 2026 · similarity 0.86)
- VSTAR: Generative Temporal Nursing for Longer Dynamic Video Synthesis (ICLR 2025 · similarity 0.86)
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.
