Research map / Video / Video generation
Synthetic Video Enhances Physical Fidelity in Video Synthesis
Qi Zhao, Xingyu Ni, Ziyu Wang, Feng Cheng, Ziyan Yang, Lu Jiang, Bohan Wang
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 forward and following camera shots in synthetic data instead of random cameras (ablation) | synthetic data for video DiT · Human motion collapse | collapse rate: 87% -> 23% | Improved |
| use high-quality 3D assets in synthetic videos (ablation) | synthetic data for video DiT · Spin shot | success rate: 22.5% -> 85% | Improved |
| use high-quality 3D assets in synthetic videos (ablation) | synthetic data for video DiT · Layer decomposition | success rate: 92.5% -> 95% | No clear effect |
| use high-quality rendering instead of low-cost rendering (ablation) | synthetic data for video DiT · Gym | success rate: 41.7% -> 83.3% | Improved |
| use high-quality rendering instead of low-cost rendering (ablation) | synthetic data for video DiT · Layer decomposition | success rate: 17.5% -> 95% | Improved |
| add fine-grained element-wise captions instead of generic VLM captions (ablation) | synthetic data captions · Uprock dance move | success rate: 2% -> 98% | Improved |
| add fine-grained element-wise captions instead of generic VLM captions (ablation) | synthetic data captions · Spin dance move | success rate: 16% -> 84% | Improved |
| add fine-grained element-wise captions instead of generic VLM captions (ablation) | synthetic data captions · Freeze dance move | success rate: 0% -> 66% | Improved |
| add special tags (e.g. animated, rendered) to synthetic captions (ablation) | synthetic data captions · Dance move | success rate: 12.5% -> 90% | Improved |
| add special tags to negative prompts at inference (ablation) | special-tag captions · Dance move | success rate: 90% -> 92.5% | No clear effect |
| raise synthetic mix from 10% to 50% at 10000 steps (ablation) | mixed training · Mix rate study | success rate without synthetic artifacts: 40% -> 85% | Improved |
| train longer (10000 to 15000 steps) with 50% synthetic data (ablation) | mixed training · Mix rate study | success rate without synthetic artifacts: 85% -> 80% | Hurt |
Closest papers by meaning
- PhysChoreo: Physics-Controllable Video Generation with Part-Aware Semantic Grounding (ECCV 2026 · similarity 0.89)
- VideoPhy: Evaluating Physical Commonsense for Video Generation (ICLR 2025 · similarity 0.88)
- PhysCtrl: Generative Physics for Controllable and Physics-Grounded Video Generation (NeurIPS 2025 · similarity 0.87)
- Hierarchical Fine-grained Preference Optimization for Physically Plausible Video Generation (NeurIPS 2025 · similarity 0.86)
- 4Real: Towards Photorealistic 4D Scene Generation via Video Diffusion Models (NeurIPS 2024 · similarity 0.86)
- SimVS: Simulating World Inconsistencies for Robust View Synthesis (CVPR 2025 · similarity 0.85)
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.
