GEOPHYS: The Geometry of Physical Plausibility

2026-06-15

AI Summary

This work addresses the computational expense, reliance on external models, or need for training modifications that plague existing methods for evaluating physical plausibility in videos. The authors propose GEOPHYS, which reveals—for the first time—that frame-wise embeddings from a frozen image encoder inherently encode five geometric properties that serve as effective signals of physical reasonableness. Leveraging this insight, GEOPHYS efficiently discriminates physically implausible events without any additional training and functions as a best-of-N verifier for physics alignment in video generation. Experiments demonstrate its superior performance, achieving 98.3% and 93.3% accuracy on LikePhys and IntPhys2 benchmarks, respectively. When applied to MAGI-1 24B video generation, it improves physical plausibility to 64.50% while reducing computational time by 1.5× and memory usage by 4.65× compared to prior approaches.

Abstract

While humans can identify physically implausible events within milliseconds, machine learning approaches addressing the same problem are extremely slow and expensive. They either rely on external multimodal-LLM judges or require ad-hoc modifications to the training procedure. In this work, we argue that indicators of physical plausibility are implicitly captured by five geometric properties of the per-frame embeddings produced by frozen image encoders. In aggregate, we call them GEOPHYS. First, we show that these signals correlate with human EEG responses to two forms of object-permanence violations. Second, GEOPHYS robustly discriminates physically implausible videos from realistic ones, achieving state-of-the-art physics-violation detection: 98.3% on LikePhys and 93.3% on IntPhys2, whereas V-JEPA 2, GPT-4o, Gemini, and twelve modern video diffusion models perform near chance. Third, used as a best-of-N verifier for physical alignment during video generation, GEOPHYS lifts MAGI-1 24B from 50.01% to 64.50% on PhysicsIQ at 1.5x lower wall-clock and 4.65x lower memory than the V-JEPA 2 world-model verifier. Ultimately, GEOPHYS demonstrates that physical plausibility in videos can be assessed by leveraging the emergent geometric properties of temporal features extracted from image encoders.

Problem

• physical plausibility
• video understanding
• physics violation detection
• machine learning
• geometric properties

Innovation

• GEOPHYS
• physical plausibility
• geometric properties
• frozen image encoders
• video generation