Research map / Video / Video generation
Rethinking Position Embedding as a Context Controller for Multi-Reference and Multi-Shot Video Generation
Binyuan Huang, Yuning Lu, Weinan Jia, Hualiang Wang, Mu Liu, Daiqing Yang
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 SideInfo-RoPE to the multi-shot reference-to-video model (ablation) | VACE-14B based PoCo · Multi-shot test set (half schedule) | CrossShot-FaceSim: 77.29 -> 81.55 | Improved |
| add SideInfo-RoPE (Tlow) (ablation) | VACE-14B based PoCo · Multi-shot test set (half schedule) | FaceSim: 45.42 -> 60.35 | Improved |
| add SideInfo-RoPE (Tlow) (ablation) | VACE-14B based PoCo · Multi-shot test set (half schedule) | CrossShot-DINO: 91.25 -> 91.32 | No clear effect |
| add SideInfo-RoPE (Tlow) (ablation) | VACE-14B based PoCo · Multi-shot test set (half schedule) | Prompt following: 52.40 -> 52.96 | Improved |
| assign low-frequency rather than high-frequency temporal channels to SideInfo axis (ablation) | SideInfo-RoPE · Multi-shot test set (half schedule) | FaceSim: 55.54 -> 60.35 | Improved |
| use high-frequency temporal channels (Thigh) for SideInfo instead of low (ablation) | SideInfo-RoPE · Multi-shot test set (half schedule) | Prompt following: 52.96 -> 53.55 | Mixed |
| use PoCo (multi-shot joint model) instead of single-shot VACE-14B | VACE-14B · Multi-shot multi-reference test set | Alignment FaceSim: 67.05 -> 70.12 | Improved |
| use PoCo instead of Phantom-14B | multi-shot generation · Multi-shot multi-reference test set | CrossShot-DINO: 73.24 -> 92.66 | Improved |
| use PoCo instead of Phantom-14B | multi-shot generation · Multi-shot multi-reference test set | Alignment FaceSim: 72.75 -> 70.12 | Hurt |
| use PoCo instead of Phantom-14B | multi-shot generation · Multi-shot multi-reference test set | Prompt alignment: 55.25 -> 51.29 | Hurt |
Closest papers by meaning
- RefAlign: Representation Alignment for Reference-to-Video Generation (ECCV 2026 · similarity 0.86)
- AlignVid: Taming Visual Dominance via Training-Free Attention Modulation in Text-guided Image-to-Video Generation (ICML 2026 · similarity 0.86)
- EntityBench: Towards Entity-Consistent Long-Range Multi-Shot Video Generation (NeurIPS 2026 · similarity 0.84)
- FRESCO: Spatial-Temporal Correspondence for Zero-Shot Video Translation (CVPR 2024 · similarity 0.84)
- RefDrop: Controllable Consistency in Image or Video Generation via Reference Feature Guidance (NeurIPS 2024 · similarity 0.84)
- Mask^2DiT: Dual Mask-based Diffusion Transformer for Multi-Scene Long Video Generation (CVPR 2025 · similarity 0.84)
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.
