Research map / Video / Video generation
AnchorWeave: World-Consistent Video Generation with Retrieved Local Spatial Memories
Zun Wang, Han Lin, Jaehong Yoon, Jaemin Cho, Yue Zhang, Mohit Bansal
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 |
|---|---|---|---|
| condition on local per-frame point clouds instead of a fused global point cloud (ablation) | AnchorWeave (CogVideoX-5B) · RE10K+DL3DV partial-revisit | PSNR: 16.31 -> 20.96 | Improved |
| retrieve more memories per chunk (K=1 to K=4) (ablation) | AnchorWeave (CogVideoX-5B) · RE10K+DL3DV partial-revisit | PSNR: 19.01 -> 20.01 -> 20.96 | Improved |
| use multiple anchors (K=4) instead of single anchor (ablation) | AnchorWeave (Wan2.2-5B) · RE10K+DL3DV partial-revisit | SSIM: 0.6180 -> 0.6739 | Improved |
| add pose-conditioned fusion instead of simple averaging of anchors (ablation) | multi-anchor controller · RE10K+DL3DV partial-revisit | visual artifacts (qualitative) | Improved |
| use joint attention across anchors instead of separate attention (ablation) | multi-anchor controller · RE10K+DL3DV partial-revisit | visual quality (qualitative) | Improved |
| replace CogVideoX-5B backbone with Wan2.2-5B | AnchorWeave · RE10K+DL3DV partial-revisit | VBench quality total: 80.30 -> 80.98 | Improved |
| use AnchorWeave single-anchor variant vs single-anchor baselines | AnchorWeave (K=1, CogVideoX) · RE10K+DL3DV partial-revisit | PSNR: 19.01 vs 18.23 (Gen3C) | Improved |
| use AnchorWeave (CogVideoX) versus SEVA on reconstruction fidelity | AnchorWeave · RE10K+DL3DV partial-revisit | PSNR: 20.96 vs 21.13 | No clear effect |
| use AnchorWeave versus SEVA on temporal quality | AnchorWeave · RE10K+DL3DV partial-revisit | Temporal Flicker: 91.00 -> 93.90 | Improved |
| add target camera pose control alongside anchors (no retrieval, first frame only) | AnchorWeave · Camera control (300 videos) | Translation error: 1.96 -> 1.72 (vs SEVA) | Improved |
| increase K from 1 to 4 (ablation) | AnchorWeave (Wan2.2-5B) · Efficiency | Generation time / GPU memory: 221 -> 275 s; 34 -> 39 GB | Hurt |
| fuse more frames into one global point cloud (ablation) | global point cloud rendering · Rendering diagnostic | PSNR (dB): 44.93 -> 16.15 | Hurt |
Closest papers by meaning
- Video World Models with Long-term Spatial Memory (NeurIPS 2025 · similarity 0.88)
- SynCamMaster: Synchronizing Multi-Camera Video Generation from Diverse Viewpoints (ICLR 2025 · similarity 0.88)
- GEN3C: 3D-Informed World-Consistent Video Generation with Precise Camera Control (CVPR 2025 · similarity 0.87)
- WorldWeaver: Generating Long-Horizon Video Worlds via Rich Perception (NeurIPS 2025 · similarity 0.87)
- Learning World Models for Interactive Video Generation (NeurIPS 2025 · similarity 0.87)
- SimVS: Simulating World Inconsistencies for Robust View Synthesis (CVPR 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.
