Research map / Video / Video generation
Decouple and Track: Benchmarking and Improving Video Diffusion Transformers For Motion Transfer
Qingyu Shi, Jianzong Wu, Jinbin Bai, Jiangning Zhang, Lu Qi, Yunhai Tong, Xiangtai 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 |
|---|---|---|---|
| use shared temporal Conv1D kernel instead of LoRA (ablation) | DeT (CogVideoX-5B) · MTBench (30-video subset) | Motion Fidelity: 80.4 -> 85.6 | Improved |
| use shared temporal Conv1D kernel instead of Local Attn (ablation) | DeT (CogVideoX-5B) · MTBench (30-video subset) | Motion Fidelity: 73.1 -> 85.6 | Improved |
| use shared temporal Conv1D kernel instead of Conv3D (ablation) | DeT (CogVideoX-5B) · MTBench (30-video subset) | Edit Fidelity: 27.1 -> 31.6 | Improved |
| add dense point tracking loss (ablation) | DeT (CogVideoX-5B) · MTBench (30-video subset) | Motion Fidelity: 82.3 -> 85.6 | Improved |
| add dense point tracking loss (ablation) | DeT (CogVideoX-5B) · MTBench (30-video subset) | Edit Fidelity: 31.2 -> 31.6 | Improved |
| drop temporal kernel in last 65% of layers instead of last 75% (ablation) | DeT (CogVideoX-5B) · MTBench (30-video subset) | Motion Fidelity: 83.1 -> 85.6 | Improved |
| drop temporal kernel in last 65% of layers instead of last 55% (ablation) | DeT (CogVideoX-5B) · MTBench (30-video subset) | Edit Fidelity: 30.4 -> 31.6 | Improved |
| kernel size 3 instead of 7 (ablation) | DeT (CogVideoX-5B) · MTBench (30-video subset) | Motion Fidelity: 81.7 -> 85.6 | Improved |
| mid dim 128 instead of 256 (ablation) | DeT (CogVideoX-5B) · MTBench (30-video subset) | Motion Fidelity: 82.6 -> 85.6 | Improved |
| lower denoising loss weight from 1.0 to 0.1 (ablation) | DeT (CogVideoX-5B) · MTBench (30-video subset) | Edit Fidelity: 31.6 -> 32.1 | Mixed |
| DeT vs MotionInversion on the same DiT | CogVideoX-5B · MTBench (All) | Edit Fidelity: 26.6 -> 31.2 | Improved |
| DeT vs MotionDirector (U-Net) | DeT · MTBench (All) | Motion Fidelity: 67.7 -> 85.8 | 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.
- MotionClone — MotionClone: Training-Free Motion Cloning for Controllable Video Generation (ICLR 2025) · better on 3 of 3 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
- Video Motion Transfer with Diffusion Transformers (CVPR 2025 · similarity 0.93)
- Emergent Temporal Correspondences from Video Diffusion Transformers (NeurIPS 2025 · similarity 0.92)
- Motion Consistency Model: Accelerating Video Diffusion with Disentangled Motion-Appearance Distillation (NeurIPS 2024 · similarity 0.90)
- Video Diffusion Models are Training-free Motion Interpreter and Controller (NeurIPS 2024 · similarity 0.90)
- EDEN: Enhanced Diffusion for High-quality Large-motion Video Frame Interpolation (CVPR 2025 · similarity 0.89)
- Space-Time Diffusion Features for Zero-Shot Text-Driven Motion Transfer (CVPR 2024 · similarity 0.89)
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.
