Research map / Video / Video generation
$PhyWorldBench$: A Comprehensive Evaluation of Physical Realism in Text-to-Video Models
Jing Gu, Xian Liu, Yu Zeng, Ashwin Nagarajan, Fangrui Zhu, Daniel Hong, Yue Fan, Qianqi Yan, Kaiwen Zhou, Ming-Yu Liu, Xin Eric 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 |
|---|---|---|---|
| tell the MLLM the video is AI-generated (context-aware prompt) (ablation) | CAP evaluator (GPT-o1) · PhyWorldBench evaluation | ROC-AUC (PC): 65.6 -> 75.1 | Improved |
| add two-step describe-then-judge chain-of-thought (ablation) | CAP evaluator (GPT-o1) · PhyWorldBench evaluation | ROC-AUC (PC): 73.6 -> 75.1 | Improved |
| add CAP to GPT-o1 judge | GPT-o1 · PhyWorldBench evaluation | ROC-AUC (PC): 61.6 -> 75.1 | Improved |
| add CAP to Qwen-VL-2.0 judge | Qwen-VL-2.0 · PhyWorldBench evaluation | ROC-AUC (PC): 59.8 -> 74.1 | Improved |
| add CAP to GPT-4o judge | GPT-4o · PhyWorldBench evaluation | ROC-AUC (SA): 72.1 -> 78.5 | Improved |
| fine-tune GPT-4o evaluator on 80% of annotated videos | CAP with GPT-4o · PhyWorldBench evaluation | PC: 73.9 -> 80.7 | Improved |
| use physics-enhanced prompt instead of event prompt | Hunyuan 720p · PhyWorldBench | physics-following fraction: 0.159 -> 0.198 | Improved |
| use physics-enhanced prompt instead of event prompt | LTX-Video · PhyWorldBench | physics-following fraction: 0.056 -> 0.065 | No clear effect |
| use detailed narrative prompt instead of event prompt | Hunyuan 720p · PhyWorldBench | physics-following fraction: 0.159 -> 0.155 | No clear effect |
| use detailed narrative prompt instead of event prompt | CogVideoX-1.5 · PhyWorldBench | physics-following fraction: 0.123 -> 0.168 | Improved |
| prompt erratic visual changes vs softened prompts | 12 T2V models · paired prompts, human eval | success rate: 42.1% -> 22.1% | Hurt |
| prompt higher-complexity multi-force interactions vs simpler cases | 12 T2V models · paired prompts, human eval | success rate: 18% -> 12% | Hurt |
Closest papers by meaning
- PhysCtrl: Generative Physics for Controllable and Physics-Grounded Video Generation (NeurIPS 2025 · similarity 0.91)
- Towards World Simulator: Crafting Physical Commonsense-Based Benchmark for Video Generation (ICML 2025 · similarity 0.91)
- VideoPhy: Evaluating Physical Commonsense for Video Generation (ICLR 2025 · similarity 0.90)
- How Far Is Video Generation from World Model: A Physical Law Perspective (ICML 2025 · similarity 0.90)
- PhysGen3D: Crafting a Miniature Interactive World from a Single Image (CVPR 2025 · similarity 0.90)
- Hierarchical Fine-grained Preference Optimization for Physically Plausible Video Generation (NeurIPS 2025 · 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.
