Research map / Video / Video generation
GEOPHYS: The Geometry of Physical Plausibility
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Shows that trajectory geometry (speed, curvature, acceleration, linear-prediction residual) of per-frame features from frozen image encoders gives a training-free physical-plausibility score that detects physics violations (SoTA on LikePhys and IntPhys2) and improves best-of-N video generation on PhysicsIQ.
Problem
Detecting physically implausible videos and selecting physically plausible generations is slow and expensive, relying on video world models, MLLM judges, or physics-specific training.
Approach
Physical plausibility is reflected in the kinematic geometry of frame-level feature trajectories from frozen image encoders (speed variation, turning-angle consistency, acceleration, linear AR residual), combined into a training-free scalar score usable for detection and best-of-N verification.
Key results
- Ensemble (OR) of GeoPhys backbones reaches 98.3% pairwise accuracy on LikePhys and 93.3% on IntPhys2, versus best published baselines around 57-58% on IntPhys2 (V-JEPA 2 57.5%).
- As a best-of-16 verifier, GeoPhys lifts MAGI-1 24B V2V PhysicsIQ from 50.01% to 64.50%, above WMReward at 62.29%.
- GeoPhys verifier runs at 0.25 s/video and 1.2 GB VRAM for a single backbone versus 1.5 s/video and 9.3 GB for WMReward.
Surprising findings
- Frozen image encoders with no video or physics training outperform billion-parameter video world models and multimodal LLM judges on physics-violation detection.
- All 12 video diffusion models score near chance (39-56%) on LikePhys when ranked by likelihood.
Limitations
- Signals are time-symmetric, so reversed videos get the same score.
- Best-of-N gains are bounded by candidate diversity; PhysicsIQ oracle ceiling is below 100%.
- Signal is a statistical correlate of plausibility, not a physics model.
Experimental setup
Datasets: LikePhys IntPhys2 PhysicsIQ VOE EEG stimuli (ADEPT-rendered)
Benchmarks: LikePhys IntPhys2 PhysicsIQ
Metrics: pairwise accuracy ROC AUC PhysicsIQ score FVD LPIPS FVMD VBench-Q wall-clock VRAM
Baselines: V-JEPA 2 GPT-4o Gemini 2.5 Qwen-VL Hunyuan T2V WMReward Qwen3-VL VideoMAE Oracle
Builds on: perceptual straightening linearisation objective (Goroshin et al.) V-JEPA 2 WMReward DINOv2
Closest papers by meaning
- VideoHallu: Evaluating and Mitigating Multi-modal Hallucinations on Synthetic Video Understanding (NeurIPS 2025 · similarity 0.88)
- QUANTIPHY: A Quantitative Benchmark Evaluating Physical Reasoning Abilities of Vision-Language Models (CVPR 2026 · similarity 0.87)
- VideoPhy: Evaluating Physical Commonsense for Video Generation (ICLR 2025 · similarity 0.87)
- RoboScape: Physics-informed Embodied World Model (NeurIPS 2025 · similarity 0.86)
- SciVideoBench: Benchmarking Scientific Video Reasoning in Large Multimodal Models (ICML 2026 · similarity 0.86)
- Physics-Driven Spatiotemporal Modeling for AI-Generated Video Detection (NeurIPS 2025 · 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.