🤖 AI

GEOPHYS: The Geometry of Physical Plausibility

This paper introduces GEOPHYS, a method that leverages five geometric properties of per-frame embeddings from frozen image encoders to efficiently and accurately detect physical implausibilities in videos, outperforming state-of-the-art multimodal models and significantly improving physical alignment in video generation with lower computational costs.

Original authors: Christian Internò, Alexander Pondaven, Habon Issa, Fabio Pizzati, Francesco Pinto, Markus Olhofer, Ivan Laptev, Philip Torr, Eero P. Simoncelli, Barbara Hammer, David Klindt

Published 2026-06-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Christian Internò, Alexander Pondaven, Habon Issa, Fabio Pizzati, Francesco Pinto, Markus Olhofer, Ivan Laptev, Philip Torr, Eero P. Simoncelli, Barbara Hammer, David Klindt

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are watching a magic show. A magician throws a ball into the air, and it suddenly turns into a butterfly and flies away. You don't need a degree in physics to know something is wrong; your brain instantly screams, "That's impossible!"

For a long time, computers struggled to do this. To teach an AI to spot these "magic tricks" (physical violations), researchers had to build massive, expensive super-computers or train them on millions of hours of video. It was like teaching a child physics by making them read every textbook in the library before they could understand that a ball shouldn't float.

GEOPHYS is a new, surprisingly simple trick that changes the game. The researchers discovered that you don't need a super-computer or a physics degree. You just need to look at the "shape" of how an AI sees the world.

The Core Idea: The "Smooth Path" vs. The "Jagged Path"

Think of a frozen image encoder (a standard AI that looks at still photos) as a camera that takes a snapshot of a video frame by frame.

Real Life (The Smooth Path): When you watch a real video of a ball bouncing, the AI's internal "view" of that ball moves in a smooth, predictable line. It's like a car driving down a straight, well-paved highway. The AI expects the next frame to be just a little bit different from the last one.
Fake Life (The Jagged Path): When the ball suddenly defies gravity or disappears, the AI's internal view gets confused. The "car" suddenly swerves off the road, hits a wall, or teleports. The path becomes jagged, erratic, and unpredictable.

The paper calls this GEOPHYS (Geometry of Physical Plausibility). It's a math score that measures how "bumpy" or "smooth" the AI's internal path is.

Smooth path? The video is likely real.
Bumpy path? The video is likely fake or physically impossible.

Why This is a Big Deal

The researchers tested this idea in three ways, and the results were shocking:

1. It Matches Human Brains
They showed the AI and human volunteers videos of objects appearing and disappearing (like a ball vanishing behind a box).

The Humans: When the object vanished, their brains showed a specific electrical signal (EEG) indicating surprise.
The AI: At the exact same moment, the GEOPHYS score spiked.
The Takeaway: Even though the AI was never taught about "surprise" or "physics," its geometric confusion matched human brain reactions perfectly. It's as if the AI's "gut feeling" about a broken path is the same as a human's.

2. It Beats the Giants
They pitted GEOPHYS against the biggest, most expensive AI models in the world (like GPT-4o, Gemini, and massive video-diffusion models).

The Giants: These models, trained on billions of parameters, often guessed randomly (about 50% accuracy) when asked to spot fake physics.
GEOPHYS: Using simple, frozen image cameras (models that have never seen a video before), GEOPHYS achieved 98.3% accuracy on one test and 93.3% on another.
The Metaphor: It's like a child with a magnifying glass spotting a fake painting better than a team of art historians with a million-dollar lab.

3. It Makes AI Video Generators Better
When AI tries to create new videos, it often makes physics mistakes (like water flowing uphill). Usually, to fix this, you need a massive "judge" AI to check the work, which is slow and uses a lot of electricity.

The Old Way: Using a massive "World Model" verifier (like V-JEPA 2) to check videos. It's accurate but heavy and slow.
The GEOPHYS Way: Using this simple geometric score as a judge.
The Result: GEOPHYS improved the quality of AI-generated videos significantly (from 50% to 64.5% on a physics test) while using 4.65 times less memory and running 1.5 times faster than the heavy-duty judges.

The "Magic" Behind the Curtain

The most surprising part of the paper is that the AI models used for GEOPHYS were frozen.

They were trained only on still images.
They were never shown a video.
They were never taught physics.

The researchers argue that because these models learned to recognize objects in photos, they accidentally learned the "rules of the road" for how objects move. When an object breaks those rules, the model's internal geometry gets messy. They didn't need to teach the AI physics; they just needed to listen to the "noise" the AI made when physics broke.

Summary

GEOPHYS proves that you don't need a giant, expensive brain to spot fake physics. You just need to look at the geometry of how a simple image-viewer processes a video. If the path is smooth, it's real. If the path is jagged, it's a lie. And the best part? It's fast, cheap, and surprisingly accurate.

Technical Summary: GEOPHYS: The Geometry of Physical Plausibility

Problem Statement

While humans can identify physically implausible events in videos within milliseconds, current machine learning approaches for the same task are computationally expensive, slow, and often require ad-hoc modifications. Existing solutions typically rely on:

External multimodal Large Language Models (LLMs) acting as judges.
Billion-parameter video-pretrained world models (e.g., V-JEPA 2).
Physics-targeted model training or neuro-symbolic pipelines.
Hand-crafted forensic checks.

These methods share a core assumption: that signals for physical plausibility must be explicitly induced into a model or reasoned about through complex training procedures. This work challenges that assumption, asking whether a vision model trained only on still images (with no video or physics supervision) inherently contains a geometric signature of physical plausibility.

Methodology: GEOPHYS

The authors propose GEOPHYS, a training-free metric that assesses physical plausibility by analyzing the geometric properties of temporal feature trajectories extracted from frozen image encoders.

Core Hypothesis

The paper posits that vision models learn to "linearize" natural video dynamics. Physically plausible videos map to smooth, locally predictable trajectories in feature space, whereas physical violations disrupt this linearity, creating irregular signals.

The Pipeline

Input: A video $V = \{x_1, \dots, x_T\}$ is fed into a frozen visual backbone $f_\theta$ (no fine-tuning).
Feature Extraction: Each frame $x_t$ is mapped to a spatial embedding $Z_t$ . A spatial average-pooling operation extracts a single feature vector $\bar{z}_t$ per frame, forming a discrete trajectory $\Gamma(V) = (\bar{z}_1, \dots, \bar{z}_T)$ in representation space.
Geometric Signals: The system computes five kinematic statistics on this trajectory:
- Speed ( $\phi_{speed}$ ): Temporal standard deviation of frame-to-frame displacement magnitudes. Flags erratic jumps or disappearances.
- Curvature ( $\phi_{curv}$ ): Mean turning angle between consecutive displacement vectors. Measures deviation from straight trajectories.
- Angle Consistency ( $\phi_{ang}$ ): Standard deviation of turning angles. Separates consistent trajectories from erratic ones.
- Acceleration ( $\phi_{accel}$ ): Magnitude of the second-order difference (change in velocity). Captures abrupt changes in speed or direction.
- Prediction Residual ( $\phi_{perr}$ ): The error of a linear auto-regressive predictor. Measures how well the next feature lies in the affine subspace spanned by previous features.
Scoring: These signals are combined into a single scalar plausibility score. Larger values indicate less regular (more implausible) dynamics.
Backbones: The method is validated across four distinct frozen backbones, none trained on video or physics:
- Self-supervised Transformers: DINOv2 and DINOv3.
- Neuro-inspired CNNs: CORnet-S (recurrent ventral stream) and VOneNet (ResNet-50 with fixed Gabor V1 front-end).

Key Contributions

1. Biological Grounding (Neural Alignment)

The authors validate that GEOPHYS signals are not mere artifacts of pretraining statistics but align with human neurobiology.

Experiment: The per-frame GEOPHYS signals were compared to human EEG recordings (specifically Contralateral Delay Activity, CDA) during object-permanence violation tasks (Create and Vanish scenarios).
Result: GEOPHYS signals (specifically speed and acceleration in CORnet-S IT layers) track the time course of human surprise responses, rising rapidly after a violation and scaling with object count, mirroring the CDA response.

2. State-of-the-Art Detection

GEOPHYS achieves superior performance in discriminating physically plausible videos from violated ones on standard benchmarks, outperforming models with explicit video training or physics supervision.

LikePhys Benchmark: Achieved 98.3% accuracy (OR ensemble), surpassing 12 video diffusion models, GPT-4o, Gemini, and V-JEPA 2 (which performed near chance).
IntPhys2 Benchmark: Achieved 93.3% accuracy, significantly outperforming V-JEPA 2 (57.5%) and other baselines.
Complementarity: Different backbones specialize in different types of violations (e.g., CORnet-S excels at "Flag" scenarios, VOneNet at "Fluid"), allowing ensemble methods to reach near-ceiling performance.

3. Test-Time Scaling for Video Generation

GEOPHYS is deployed as a "best-of-N" verifier for physically aligned video generation, offering a computationally efficient alternative to world-model verifiers.

Setup: Used to rerank candidates generated by MAGI-1 (24B) and other generators on the PhysicsIQ benchmark.
Result: Lifted MAGI-1 24B performance from 50.01% to 64.50%.
Efficiency: Achieved this improvement with 1.5× lower wall-clock time and 4.65× lower memory usage compared to the previous state-of-the-art verifier (V-JEPA 2 based WMReward).
Quality: The selection improved physics-fidelity metrics (FVD, LPIPS) without degrading generic video quality (VBench-Q).

Significance and Claims

The paper argues that physical plausibility in videos can be assessed by leveraging the emergent geometric properties of temporal features extracted from standard image encoders.

Paradigm Shift: It challenges the necessity of explicit physics training or massive video-pretrained world models for detecting physical violations.
Cost-Efficiency: It demonstrates that a "weak" verifier (frozen image encoders, sub-billion parameters) can outperform "strong" verifiers (billion-parameter world models) in both accuracy and efficiency.
Mechanism: The authors do not claim the models understand or simulate physics. Instead, they argue that the geometry of feature-space trajectories is statistically disrupted by physical violations in a way that is reliable enough to be useful.
Implications: This suggests that for many applications, the bottleneck in generating physically plausible video is not the sophistication of the verifier, but the presence of any verifier with the correct inductive bias (geometric linearity).

The work concludes that while frozen backbones do not represent physics explicitly, their geometric signatures serve as a powerful, training-free correlate for physical plausibility, bridging computational vision, neuroscience, and generative AI.