Research map / Video / Video generation
Beyond the Golden Data: Resolving the Motion-Vision Quality Dilemma via Timestep Selective Training
Xiangyang Luo, Qingyu Li, Yuming Li, Guanbo Huang, Yongjie Zhu, Wenyu Qin, Meng Wang, Pengfei Wan, Shao-Lun Huang
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 TQD timestep-selective training | Wan 1.3B baseline · Set-A (all four quality groups) | MQ: 2.1388 -> 2.2557 | Improved |
| add TQD | Wan 1.3B baseline · Set-B (only imbalanced data) | MQ: 2.0905 -> 2.1477 | Improved |
| add TQD | Wan 1.3B baseline · Set-C (golden data only) | VQ: 3.3378 -> 3.3743 | Improved |
| add TQD | CogVideoX-5B baseline · Set-A | VQ: 3.2593 -> 3.2570 | No clear effect |
| add TQD | CogVideoX-5B baseline · Set-A | MQ: 2.1855 -> 2.2345 | Improved |
| add adaptive timestep sampling to baseline (ablation) | Wan Set-A · Set-A ablation | Dynamic: 0.5312 -> 0.6340 | Improved |
| add quality dropout to baseline (ablation) | Wan Set-A · Set-A ablation | MQ: 2.1388 -> 2.1909 | Improved |
| combine adaptive timestep and quality dropout (ablation) | Wan Set-A · Set-A ablation | MQ: 2.2193 -> 2.2557 | Improved |
| add adaptive timestep on top of data filtering (ablation) | Wan Set-A · VideoPhy2 hard cases | Prompt Consistency: 3.756 -> 3.843 | Improved |
| raise kappa_max from 10 to 30 (ablation) | Wan TQD Set-A · Set-A kappa_max sweep | MQ: 2.2282 -> 2.2557 | Improved |
| raise kappa_max from 20 to 40 (ablation) | CogVideoX TQD · Set-A kappa_max sweep | MQ: 2.2345 -> 2.2192 | Hurt |
| inject 20% Gaussian noise into MQ/VQ scores (ablation) | Wan TQD Set-A · Set-A noise robustness | MQ: 2.2557 -> 2.2114 | Hurt |
Closest papers by meaning
- Motion Consistency Model: Accelerating Video Diffusion with Disentangled Motion-Appearance Distillation (NeurIPS 2024 · similarity 0.87)
- FreeLong: Training-Free Long Video Generation with SpectralBlend Temporal Attention (NeurIPS 2024 · similarity 0.87)
- Identifying and Solving Conditional Image Leakage in Image-to-Video Diffusion Model (NeurIPS 2024 · similarity 0.86)
- Timestep Embedding Tells: It's Time to Cache for Video Diffusion Model (CVPR 2025 · similarity 0.86)
- LongDiff: Training-Free Long Video Generation in One Go (CVPR 2025 · similarity 0.86)
- VideoCrafter2: Overcoming Data Limitations for High-Quality Video Diffusion Models (CVPR 2024 · similarity 0.86)
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.
