Research map / Video / Video generation
DreamRelation: Relation-Centric Video Customization
Yujie Wei, Shiwei Zhang, Hangjie Yuan, Biao Gong, Longxiang Tang, Xiang Wang, Haonan Qiu, Hengjia Li, Shuai Tan, Yingya Zhang, Hongming Shan
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 hybrid mask training strategy (ablation) | DreamRelation · Relational video customization (26 NTU RGB+D relations) | Relation Accuracy: 0.3574 -> 0.4452 | Improved |
| add space-time relational contrastive loss (ablation) | DreamRelation · Relational video customization (26 NTU RGB+D relations) | Relation Accuracy: 0.3416 -> 0.4452 | Improved |
| add relation LoRAs (ablation) | DreamRelation · Relational video customization (26 NTU RGB+D relations) | Relation Accuracy: 0.3626 -> 0.4452 | Improved |
| add subject LoRAs (ablation) | DreamRelation · Relational video customization (26 NTU RGB+D relations) | Relation Accuracy: 0.3769 -> 0.4452 | Improved |
| add FFN LoRAs (ablation) | DreamRelation · Relational video customization (26 NTU RGB+D relations) | Relation Accuracy: 0.4021 -> 0.4452 | Improved |
| place Relation LoRA in Q,K rather than V (Subject LoRA in V) (ablation) | DreamRelation · Relational video customization (26 NTU RGB+D relations) | Relation Accuracy: 0.3444 -> 0.4452 | Improved |
| place Relation LoRA in Q,K rather than K,V (ablation) | DreamRelation · Relational video customization (26 NTU RGB+D relations) | Relation Accuracy: 0.3937 -> 0.4452 | Improved |
| add RCL to MotionInversion (ablation) | MotionInversion · Relational video customization (26 NTU RGB+D relations) | Relation Accuracy: 0.3151 -> 0.3633 | Improved |
| add RCL to MotionInversion (ablation) | MotionInversion · Relational video customization (26 NTU RGB+D relations) | CLIP-T: 0.3217 -> 0.3181 | No clear effect |
| raise contrastive loss weight lambda1 from 0.01 to 1.00 (ablation) | DreamRelation · Relational video customization (26 NTU RGB+D relations) | Relation Accuracy: 0.4452 -> 0.2998 | Hurt |
| raise contrastive loss weight lambda1 from 0.01 to 0.10 (ablation) | DreamRelation · Relational video customization (26 NTU RGB+D relations) | Relation Accuracy: 0.4452 -> 0.3964 | Hurt |
| vary mask weight lambda_m (1, 25, 50, 100) (ablation) | DreamRelation · Relational video customization (26 NTU RGB+D relations) | Relation Accuracy: 0.3469 -> 0.4452 at 50, 0.4018 at 100 | Mixed |
Closest papers by meaning
- DreamRelation: Bridging Customization and Relation Generation (CVPR 2025 · similarity 0.89)
- VideoMage: Multi-Subject and Motion Customization of Text-to-Video Diffusion Models (CVPR 2025 · similarity 0.87)
- Few-Shot Referring Relationships in Videos (CVPR 2023 · similarity 0.86)
- DreamVideo: Composing Your Dream Videos with Customized Subject and Motion (CVPR 2024 · similarity 0.85)
- Multi-subject Open-set Personalization in Video Generation (CVPR 2025 · similarity 0.84)
- PolyVivid: Vivid Multi-Subject Video Generation with Cross-Modal Interaction and Enhancement (NeurIPS 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.
