Research map / Video / Video generation
Evaluating Intuitive Physics Understanding in Video Diffusion Models via Likelihood Preference
Jianhao Yuan, Fabio Pizzati, Francesco Pinto, Lars Kunze, Ivan Laptev, Paul Newman, Philip Torr, Daniele De Martini
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 |
|---|---|---|---|
| scale up video diffusion model size (ablation) | video diffusion models · LikePhys benchmark | average PPE | Improved |
| train on larger data corpora (ablation) | video diffusion models · LikePhys benchmark | average PPE | Improved |
| increase number of input frames (context window) (ablation) | video diffusion models · LikePhys benchmark (5 scenarios, 5 models) | PPE | Improved |
| vary classifier-free guidance scale from 1.0 to 8.0 (ablation) | video diffusion models · LikePhys benchmark (5 scenarios, 5 models) | PPE | No clear effect |
| use DiT-based architecture instead of UNet | video diffusion models · LikePhys benchmark | average PPE: UNet models often above 50% -> top three DiT models 43.6-43.8% | Improved |
| vary text prompt among eight variants (ablation) | LikePhys · four scenarios | PPE | No clear effect |
| uniformly sample ten timesteps rather than single region (ablation) | LikePhys · four scenarios | denoising loss difference | Improved |
| use PPE likelihood preference instead of VLM judges | physics evaluators · human preference alignment | Kendall tau: VideoPhy 38.9, Qwen2.5 VL 33.3 -> 44.4 | Improved |
| compare PPE with VBench aesthetic quality | LikePhys · VBench correlation | correlation with -PPE: r = -0.05 | No clear effect |
Closest papers by meaning
- Latent Intuitive Physics: Learning to Transfer Hidden Physics from A 3D Video (ICLR 2024 · similarity 0.89)
- Lost in Latent Space: An Empirical Study of Latent Diffusion Models for Physics Emulation (NeurIPS 2025 · similarity 0.88)
- PISA Experiments: Exploring Physics Post-Training for Video Diffusion Models by Watching Stuff Drop (ICML 2025 · similarity 0.87)
- VideoREPA: Learning Physics for Video Generation through Relational Alignment with Foundation Models (NeurIPS 2025 · similarity 0.87)
- How Far Is Video Generation from World Model: A Physical Law Perspective (ICML 2025 · similarity 0.86)
- GEOPHYS: The Geometry of Physical Plausibility (NeurIPS 2026 · 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.
