Research map / Video / Video generation
Generative Video Motion Editing with 3D Point Tracks
Yao-Chih Lee, Zhoutong Zhang, Jiahui Huang, Jui-Hsien Wang, Joon-Young Lee, Jia-Bin Huang, Eli Shechtman, Zhengqi Li
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 |
|---|---|---|---|
| replace naive Gaussian-kernel sampling with cross-attention sampling/splatting (ablation) | 3D track conditioner · in-the-wild videos | EPE: 16.18 -> 7.03 | Improved |
| use 3D tracks instead of 2D tracks (ablation) | cross-attention conditioner · DyCheck | PSNR: 13.88 -> 14.82 | Improved |
| use 3D tracks instead of 2D tracks (ablation) | cross-attention conditioner · in-the-wild videos | FVD: 305.93 -> 335.85 | Hurt |
| inject track z embeddings (ablation) | 3D track conditioner · in-the-wild videos | EPE: 7.44 -> 6.12 | Improved |
| add real-data fine-tuning after synthetic training (ablation) | Edit-by-Track · in-the-wild videos | FVD: 623.32 -> 306.44 | Improved |
| add synthetic-data stage before real data (ablation) | Edit-by-Track · in-the-wild videos | EPE: 63.98 -> 6.12 | Improved |
| use two-stage training instead of single-stage mixed data (ablation) | Edit-by-Track · in-the-wild videos | FVD: 347.95 -> 306.44 | Improved |
| use Edit-by-Track instead of track-conditioned I2V ATI | Wan base · DyCheck | LPIPS: 0.468 -> 0.406 (.468 -> .406) | Improved |
| use 1.3B Edit-by-Track instead of 14B ATI | Wan · in-the-wild videos | FVD: 268.80 -> 306.44 | Hurt |
| use 1.3B Edit-by-Track instead of 14B ATI | Wan · in-the-wild videos | EPE: 11.44 -> 6.12 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- ReVideo — ReVideo: Remake a Video with Motion and Content Control (NeurIPS 2024) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Wan-Move: Motion-controllable Video Generation via Latent Trajectory Guidance (NeurIPS 2025 · similarity 0.88)
- Dynamic Motion Blending for Versatile Motion Editing (CVPR 2025 · similarity 0.88)
- Motion Prompting: Controlling Video Generation with Motion Trajectories (CVPR 2025 · similarity 0.87)
- Extrapolating and Decoupling Image-to-Video Generation Models: Motion Modeling is Easier Than You Think (CVPR 2025 · similarity 0.87)
- Direct Motion Models for Assessing Generated Videos (ICML 2025 · similarity 0.87)
- MotionClone: Training-Free Motion Cloning for Controllable Video Generation (ICLR 2025 · 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.
