Research map / Video / Video generation
Inference-time Physics Alignment of Video Generative Models with Latent World Models
Jianhao Yuan, Xiaofeng Zhang, Felix Friedrich, Nicolas Beltran-Velez, Melissa Hall, Reyhane Askari-Hemmat, Xiaochuang Han, Nicolas Ballas, Michal Drozdzal, Adriana Romero-Soriano
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 WMReward best-of-N search (16 particles) | vLDM · PhysicsIQ I2V | PhysicsIQ score: 27.76 -> 32.90 | Improved |
| add guidance on top of WMReward BoN | vLDM · PhysicsIQ I2V | PhysicsIQ score: 32.90 -> 33.44 | Improved |
| add guidance on top of BoN | MAGI-1 V2V · PhysicsIQ V2V | PhysicsIQ score: 60.34 -> 62.00 | Improved |
| WMReward guidance alone without search | MAGI-1 V2V · PhysicsIQ V2V | PhysicsIQ score: 55.22 -> 55.02 | No clear effect |
| WMReward guidance alone without search | vLDM I2V · PhysicsIQ I2V | PhysicsIQ score: 27.76 -> 27.88 | No clear effect |
| use WMReward BoN instead of Qwen2.5-VL reward BoN (ablation) | MAGI-1 I2V · PhysicsIQ I2V | PhysicsIQ score: 24.99 -> 36.56 | Improved |
| use VideoMAE reward BoN on Wan2.2 (ablation) | Wan2.2 I2V · PhysicsIQ I2V | PhysicsIQ score: 38.26 -> 37.91 | Hurt |
| add WMReward BoN | Wan2.2 · PhysicsIQ I2V | PhysicsIQ score: 38.26 -> 44.39 | Improved |
| add WMReward BoN to API model | Sora2 · PhysicsIQ I2V | PhysicsIQ score: 42.30 -> 46.43 | Improved |
| add WMReward BoN | MAGI-1 · VideoPhy overall PC | physics consistency (PC): 25.0 -> 32.8 | Improved |
| add WMReward BoN | MAGI-1 · VideoPhy overall SA | semantic adherence (SA): 54.4 -> 45.3 | Hurt |
| use BoN instead of SMC sampling with the WMReward (ablation) | vLDM I2V (16 particles) · PhysicsIQ I2V | PhysicsIQ score: 29.24 -> 32.90 | Improved |
Closest papers by meaning
- Physics-Informed Video Generation via Mixture-of-Experts Latent Alignment (NeurIPS 2026 · similarity 0.91)
- Hierarchical Fine-grained Preference Optimization for Physically Plausible Video Generation (NeurIPS 2025 · similarity 0.89)
- VideoREPA: Learning Physics for Video Generation through Relational Alignment with Foundation Models (NeurIPS 2025 · similarity 0.89)
- PhysCtrl: Generative Physics for Controllable and Physics-Grounded Video Generation (NeurIPS 2025 · similarity 0.89)
- Inference-Time Text-to-Video Alignment with Diffusion Latent Beam Search (NeurIPS 2025 · similarity 0.88)
- GEOPHYS: The Geometry of Physical Plausibility (NeurIPS 2026 · similarity 0.88)
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.
