Research map / Video / Video generation
Motion Attribution for Video Generation
Xindi Wu, Despoina Paschalidou, Jun Gao, Antonio Torralba, Laura Leal-Taixé, Olga Russakovsky, Sanja Fidler, Jonathan Lorraine
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 motion masking to attribution (ablation) | Motive · VBench (Wan2.1-T2V-1.3B) | Dynamic degree (%): 43.8 -> 47.6 | Improved |
| add motion masking to attribution (ablation) | Motive · VBench (Wan2.2-TI2V-5B) | Dynamic degree (%): 43.8 -> 48.3 | Improved |
| add motion masking to attribution (ablation) | Motive · VBench (LTX-2B) | Dynamic degree (%): 41.8 -> 45.1 | Improved |
| select 10% data with Motive instead of random 10% | fine-tuning Wan2.1-T2V-1.3B · VBench | Dynamic degree (%): 41.3 -> 47.6 | Improved |
| fine-tune on Motive 10% data instead of full dataset | fine-tuning Wan2.1-T2V-1.3B · VBench | Dynamic degree (%): 42.0 -> 47.6 | Improved |
| fine-tune on Motive 10% data instead of full dataset | fine-tuning Wan2.2-TI2V-5B · VBench | Dynamic degree (%): 45.3 -> 48.3 | Improved |
| Motive selection vs V-JEPA selection | fine-tuning Wan2.1-T2V-1.3B · VBench | Dynamic degree (%): 41.6 -> 47.6 | Improved |
| Motive selection vs motion magnitude selection | fine-tuning LTX-2B · VBench | Dynamic degree (%): 39.6 -> 45.1 | Improved |
| Motive selection vs full fine-tuning on imaging quality | fine-tuning Wan2.1-T2V-1.3B · VBench | Imaging quality (%): 63.9 -> 64.6 | Hurt |
| add Motive-selected fine-tuning vs base model (human preference) | Wan2.1-T2V-1.3B · Human evaluation | Win rate (%): 74.1 | Improved |
| add motion masking (human preference) (ablation) | Motive · Human evaluation | Win rate (%): 46.9 win, 33.1 loss | Improved |
| increase projection dimension from 128 to 512 (ablation) | Motive gradient projection · Gradient ranking preservation | Spearman correlation (%): 46.9 -> 74.7 | Improved |
Closest papers by meaning
- Direct Motion Models for Assessing Generated Videos (ICML 2025 · similarity 0.88)
- VideoJAM: Joint Appearance-Motion Representations for Enhanced Motion Generation in Video Models (ICML 2025 · similarity 0.88)
- MotionRAG: Motion Retrieval-Augmented Image-to-Video Generation (NeurIPS 2025 · similarity 0.88)
- Motion Prompting: Controlling Video Generation with Motion Trajectories (CVPR 2025 · similarity 0.88)
- Make Pixels Dance: High-Dynamic Video Generation (CVPR 2024 · similarity 0.88)
- Enhancing Motion in Text-to-Video Generation with Decomposed Encoding and Conditioning (NeurIPS 2024 · similarity 0.87)
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.
