Principia

Relational Physics Tests for Video Models

Varun Varma Thozhiyoor1*, Shivam Tripathi1*, Venkatesh Babu Radhakrishnan1, Anand Bhattad2

1Indian Institute of Science    2Johns Hopkins University

*Equal contribution

Paper Benchmark

TL;DR

Modern video generators produce visually realistic motion, but fail surprisingly simple Newtonian physics tests. Principia evaluates physics using calibration-independent relational constraints between paired objects and finds that no evaluated generator exceeds 0.42 physical consistency despite all scoring above 0.7 on VBench.

Abstract

Evaluating physical reasoning in video models is difficult because absolute motion measurements depend on frame rate, object scale, and camera calibration, all of which are often ambiguous or unavailable in generated video. We propose a different approach. When two objects in the same scene obey the same physical law, their motions must satisfy predictable relationships, and these relationships hold independent of calibration. We introduce Principia, a benchmark that evaluates Newtonian physics through relational consistency between paired objects.

Principia spans eight phenomena—gravity, restitution, friction, rotational inertia, projectile motion, momentum, pendulum, and mass–spring oscillation—across translational, rotational, collisional, and oscillatory dynamics, using real-world scenes recorded under controlled protocols. We also introduce a calibration-independent consistency score that quantifies physical violation directly in image space. Across thousands of generations from five state-of-the-art video generators, no model exceeds 0.42 on Principia despite all scoring around 0.8 on VBench. Vision-language models are evaluated on their ability to detect relational physics violations, with the best model achieving only 67% accuracy and most performing near chance level.

Why Evaluate Relatively?

Single object sliding down an incline, motion looks plausible

Single Object Motion

A generated video of a single block sliding down an incline, shown at three timesteps and as a stroboscopic composite (right): motion appears plausible.

Two blocks of different mass released on identical inclines, arrival times compared

Paired Object Comparison

A generated video of two blocks with different masses released simultaneously on identical inclines, where the blocks arrive at different times, violating the mass-independence invariant of gravitational acceleration. This violation only becomes detectable through comparison.

Physics Scenarios

Speed
Hover to preview · Click to inspect

Each scenario defines a relational invariant that must hold whenever two objects obey the same physical law. Hover a card to preview the invariant; click to inspect the recorded scene.

Leaderboard

Principia consistency score for video generators (out of 1.0).

Principia VBench
Wan2.2-14B
Principia 0.419
VBench 0.780
Omni
Principia 0.409
VBench 0.810
Veo-3.1
Principia 0.379
VBench 0.800
Wan2.2-5B
Principia 0.205
VBench 0.800
Cosmos-2.5-14B
Principia 0.199
VBench 0.780
Cosmos-2.5-2B
Principia 0.148
VBench 0.780

Physical Consistency Analysis

We evaluate both the ability of video generators to produce physically consistent motion and the ability of vision-language models to detect violations of relational physics.

Relational Physics Consistency

For each phenomenon φ, the invariant defines two scalar quantities F☉(o₁) and F☉(o₂) that should be equal under correct physics. Depending on the phenomenon, these quantities may correspond either to directly measured values (friction) or to ratios derived from the measured values (gravity, spring, restitution, rotational inertia, pendulum).

S☉ equals 1 when the invariant holds exactly and decreases toward 0 as the violation grows. A score of S☉ = 0.9 corresponds to roughly 10% relational asymmetry. The normalization makes S☉ unit-free and bounded in [0, 1].

Principia vs. VBench

Scatter plot comparing VBench visual quality score against Principia physical consistency score

Visual quality and physical fidelity are largely decoupled. All five video generators score above 0.7 on VBench but below 0.5 on Principia, clustering in the high-visual-quality, low-physics-fidelity region.

Per-phenomenon consistency radar chart for video generators

Video Generators

Each polygon exhibits a distinct performance profile rather than uniform weakness. Veo-3.1 peaks at inertia and gravity; Wan2.2-14B performs particularly well on friction and pendulum; Cosmos-2.5-2B exhibits a smaller polygon, with its strongest performance concentrated on gravity and restitution. No polygon fills the chart.

Per-phenomenon agreement radar chart for vision-language models

Vision-Language Models

VLMs are evaluated on whether they can detect relational physics violations rather than generate physically correct motion. Their polygons are smoother and more uniform than those of the generators, but performance remains limited: three of four models perform near chance level, while the best-performing model reaches only 0.67 average detection accuracy, just above chance.

Video Gallery

Select a physics scenario and an example scene to compare generations across all six models.

Result Gallery

Generation Prompt

Browse All Generations

Synthetic Videos

Use the left and right buttons to browse paired clips

Each scenario includes a synthetic video that demonstrates the underlying physics invariant and an anti-physics counterpart that violates the invariant.

Citation

If you find Principia useful in your research, please consider citing our work. 

@InProceedings{Thozhiyoor_2026_CVPR,
    author    = {Thozhiyoor, Varun Varma and Tripathi, Shivam and Radhakrishnan, Venkatesh Babu and Bhattad, Anand},
    title     = {Objects in Generated Videos Are Slower Than They Appear: Models Suffer Sub-Earth Gravity and Don't Know Galileo's Principle...for now},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Findings},
    month     = {June},
    year      = {2026},
    pages     = {3830--3839}
}