Research map / Video / Video generation
RefAlign: Representation Alignment for Reference-to-Video Generation
Lei Wang, YuXin Song, Ge Wu, Haocheng Feng, Hang Zhou, Jingdong Wang, Yaxing Wang, Jian 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 RA loss (ablation) | RefAlign (Wan2.1 1.3B) · OpenS2V-Eval | TotalScore: 49.93% -> 55.73% | Improved |
| add RA loss (ablation) | RefAlign (Wan2.1 1.3B) · OpenS2V-Eval | NaturalScore: 40.46% -> 62.96% | Improved |
| add RA loss (ablation) | RefAlign (Wan2.1 1.3B) · OpenS2V-Eval | FaceSim: 68.45% -> 53.15% | Hurt |
| add negative loss L_neg to RA loss (ablation) | RefAlign · OpenS2V-Eval | TotalScore: 51.75% -> 55.73% | Improved |
| align to DINOv3 via loss instead of feeding DINOv3 as extra input (ablation) | RefAlign · OpenS2V-Eval | TotalScore: 52.15% -> 55.73% | Improved |
| add DINOv3 input features (config D) vs no RA loss (ablation) | Wan2.1 baseline · OpenS2V-Eval | TotalScore: 49.93% -> 52.15% | Improved |
| use DINOv3-H+ instead of DINOv3-L as alignment encoder (ablation) | RefAlign · OpenS2V-Eval | TotalScore: 55.73% -> 55.30% | No clear effect |
| use SigLIP2-So instead of DINOv3-L as alignment encoder (ablation) | RefAlign · OpenS2V-Eval | TotalScore: 55.73% -> 54.25% | Hurt |
| use Qwen2.5-VL-7B instead of DINOv3-L as alignment encoder (ablation) | RefAlign · OpenS2V-Eval | TotalScore: 55.73% -> 53.12% | Hurt |
| increase alignment depth from 9 to 30 (ablation) | RefAlign · OpenS2V-Eval | FaceSim: 54.28% -> 9.59% | Hurt |
| increase alignment depth from 9 to 30 (ablation) | RefAlign · OpenS2V-Eval | NaturalScore: 66.62% -> 80.21% | Improved |
| increase alignment depth from 9 to 30 (ablation) | RefAlign · OpenS2V-Eval | TotalScore: 55.40% -> 50.12% | Hurt |
Closest papers by meaning
- Scaling Zero-Shot Reference-to-Video Generation (CVPR 2026 · similarity 0.89)
- Dynamic Reflections: Probing Video Representations with Text Alignment (ICLR 2026 · similarity 0.88)
- Phantom: Subject-Consistent Video Generation via Cross-Modal Alignment (ICCV 2025 · similarity 0.87)
- RefDrop: Controllable Consistency in Image or Video Generation via Reference Feature Guidance (NeurIPS 2024 · similarity 0.87)
- Refine-by-Align: Reference-Guided Artifacts Refinement through Semantic Alignment (ICLR 2025 · similarity 0.87)
- Can Text-to-Video Generation help Video-Language Alignment? (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.
