Research map / Video / Video generation
When Distillation Breaks Motion Control: Restoring Generative Trajectories for Fast Video Generators
Jintao Rong, Xin Xie, Xinyi Yu, Linlin Ou, Xinyu Zhang, Chunhua Shen, Dong Gong
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 guidance to distilled student (ablation) | TurboV2 student · TurboBench (TurboV2) | motion fidelity: 0.562 -> 0.833 | Improved |
| add motion guidance to distilled student (ablation) | TurboV2 student · TurboBench (TurboV2) | text alignment: 0.348 -> 0.324 | Hurt |
| add teacher guidance (ablation) | TurboV2 + MG · TurboBench (TurboV2) | motion fidelity: 0.833 -> 0.912 | Improved |
| add teacher guidance (ablation) | TurboV2 + MG · TurboBench (TurboV2) | inference time: 7s -> 20s | Hurt |
| add adaptive acceleration (ablation) | TurboV2 + MG + TG · TurboBench (TurboV2) | inference time: 20s -> 13s | Improved |
| add adaptive acceleration (ablation) | TurboV2 + MG + TG · TurboBench (TurboV2) | motion fidelity: 0.912 -> 0.933 | Improved |
| raise teacher guidance strength lambda from 0.1 to 0.5 (ablation) | MotionEcho · TurboBench (TurboV2) | FID: 332.23 -> 320.10 (MF 0.924 -> 0.930, best 0.936 at 0.3) | Mixed |
| raise motion guidance strength eta from 0.05 to 0.2 (ablation) | MotionEcho · TurboBench (TurboV2) | motion fidelity: 0.929 -> 0.940 (TC 0.978 -> 0.966) | Mixed |
| raise activation threshold delta1 from 360 to 500 (ablation) | MotionEcho · TurboBench (TurboV2) | inference time: 19.89 -> 13.29 (MF 0.9314 -> 0.9308) | Improved |
| raise truncation threshold delta2 from 380 to 500 (ablation) | MotionEcho · TurboBench (TurboV2) | inference time: 17.83 -> 12.84 | Improved |
| reduce student steps from 16 to 4 (ablation) | MotionEcho · TurboBench (TurboV2) | FID: 322.97 -> 347.91 (ITC 13s -> 6s) | Mixed |
| apply MotionEcho to AnimateDiff-Lightning with MotionClone | AD-L · AnimateBench | motion fidelity: 0.648 -> 0.868 | 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 5 of 5 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 0.89)
- Motion Consistency Model: Accelerating Video Diffusion with Disentangled Motion-Appearance Distillation (NeurIPS 2024 · similarity 0.88)
- MotionPro: A Precise Motion Controller for Image-to-Video Generation (CVPR 2025 · similarity 0.88)
- Extrapolating and Decoupling Image-to-Video Generation Models: Motion Modeling is Easier Than You Think (CVPR 2025 · similarity 0.88)
- Motion Prompting: Controlling Video Generation with Motion Trajectories (CVPR 2025 · similarity 0.88)
- Motion-Zero: A Zero-Shot Trajectory Control Framework of Moving Object for Diffusion-Based Video Generation (AAAI 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.
