MORPHEUS: Benchmarking Physical Reasoning of Video Generative Models with Real Physical Experiments

Antonios Tragoudaras^2,*, Chenyu Zhang^2,1,*, Daniil Cherniavskii^2,*, Antonios Vozikis^2,†, Thijmen Nijdam^2,†, Derck W. E. Prinzhorn², Mark Bodracska², Nicu Sebe¹, Andrii Zadaianchuk^2,††, Stratis Gavves^2,††

¹University of Trento, Italy
²University of Amsterdam, the Netherlands
^*Equal contribution of first authors
^†Equal contribution of second authors
^††Equal advising

arXiv Code Dataset

Comparison of evaluation methods for video generative models. Human or VLM-based judgments provide only qualitative and subjective assessments of physical plausibility. Trajectory matching compares generated and ground-truth paths but may misclassify physically valid trajectories. For example, many projectile parabolas remain plausible when a VGM is conditioned on a single image with unknown velocity. Morpheus instead evaluates generations with physics-informed scores that test conservation of invariants and consistency with governing equations of motion.

Summary

Recent advances in image and video generation raise hopes that these models possess world modeling capabilities—the ability to generate realistic, physically plausible videos. This could revolutionize applications in robotics, autonomous driving, and scientific simulation. However, before treating these models as world models, we must ask: Do they adhere to physical laws? Current evaluation methods rely on subjective judgments or trajectory matching, limiting their usage for physical reasoning estimation, where many generations could be physically plausible.

We introduce Morpheus, a benchmark for evaluating video generation models on physical reasoning. It features 130 real-world videos capturing physical phenomena, guided by conservation laws. Using those as conditioning for video generation, we assess physical plausibility using physics-informed metrics evaluated with respect to infallible conservation laws known per physical setting, leveraging advances in physics-informed neural networks and vision-language foundation models.

MORPHEUS

Video Examples

Experiments

No videos available for this configuration yet.

Aggregated scores across all physical experiments (higher is better). We select the best scores between using the enhanced and plain versions of the textual prompt to represent the best model's ability. For comparison, real-world videos obtain Dynamical scores in the range of 0.96-0.99, and Physical Invariance score of 0.90 or above.

Loading aggregated score breakdown...

Discarded videos

An example of the disappearance of the object (the projectile ball). Model: PyramidalFlow, multi-frame conditioning, enhanced text prompt.

An example of the duplication of the object (non-holonomic pendulum). Model: COSMOS, multi-frame conditioning, plain text prompt.

An example of the stillness of the object (double pendulum). Model: LTX, single frame conditioning, plain text prompt.

The analysis of the major reasons behind high discard rates reveals the absence of motion (i.e. stillness) and the presence of duplicate objects, as well as, to a lesser extent, the disappearance of the object from the video.

Loading discard breakdown...

Augmentations

Experiments

No videos available yet for this experiment.

The relative difference between the score on the original, lab videos and augmented realistic settings. The realistic augmented settings appear to be much harder to predict.

Loading augmentation improvements...

Why Alternative Evaluation Methods Fall Short

Two common alternatives to Morpheus — comparing generated trajectories directly against a reference (trajectory matching) and asking a vision-language model to judge physical plausibility — both struggle to certify physically clean, real-world footage as physical. The following findings, drawn from the paper's appendix, motivate Morpheus's physics-informed scoring instead.

Trajectory matching penalizes physically valid, divergent realizations

To isolate the failure mode of trajectory matching from any generative-model noise, we ran it real-vs-real: for each experiment, every pair of distinct real-world recordings of the same phenomenon was cross-compared using Physics-IQ (Motamed et al. 2026) overlap metrics. Since both videos in every pair are clean physical realizations, a metric that measured physical correctness should score them near-perfectly — instead, overlap is low across every category, showing that small, unobservable differences in initial conditions, mass distribution, or friction make trajectory matching reject physically valid futures. Morpheus, which scores functional adherence to governing equations and conserved invariants instead of matching paths, scores near-perfectly on these same real-world references.

Experiment	Spatiotemporal IoU	Spatial IoU	Weighted Spatial IoU
Bouncing Ball	0.100	0.274	0.169
Collision	0.021	0.184	0.112
Double Pendulum	0.087	0.561	0.408
Falling	0.164	0.396	0.342
Holonomic Pendulum	0.072	0.532	0.397
Projectile	0.010	0.040	0.034
Rolling	0.306	0.773	0.725
Sliding Book	0.189	0.240	0.202
Spring	0.189	0.298	0.233

Real-vs-real Physics-IQ overlap scores. Even perfect-vs-perfect physics yields low overlap — trajectory matching penalizes valid but divergent realizations, not physical incorrectness.

VLM-based auto-eval (VideoPhy2-AutoEval) can't reliably tell real physics from generated

We also benchmarked against VideoPhy2-AutoEval (Bansal et al. 2026), a VLM fine-tuned for physical commonsense judgments, on both real-world recordings and generations from 9 models. Aggregated across all experiments, the VLM assigns real-world footage a Joint Physical Consistency (JPC) pass rate of only 23.6% (95% CI [16.1, 31.1]), barely above generated videos at 13.9% ([12.1, 15.8]) — the VLM fails to assign ceiling scores to real physics and barely separates real from generated, with no significant gap in rule adherence.

The per-category breakdown is the sharpest evidence: the VLM scores several physically clean real-world videos (bouncing ball, all three collision variants, three of four falling variants, both rolling-can variants, sliding book) at a 0% JPC pass rate, and for double pendulum it rates the generated videos as more physically plausible (41.4%) than the real ones (30.0%). An evaluator that cannot confidently certify that a real video obeys physical laws is fundamentally unsuited to benchmarking generative models — precisely what motivates Morpheus's physics-informed scoring.

Experiment	Real-world JPC pass	Generated JPC pass
Falling ball	100.0%	34.5%
Rolling orange	100.0%	39.1%
Spring	85.7%	72.4%
Double pendulum	30.0%	41.4%
Holonomic pendulum	27.8%	13.8%
Bouncing ball	0.0%	5.7%
Collision (big hits small)	0.0%	4.6%
Collision (equal)	0.0%	0.0%
Collision (small hits big)	0.0%	3.4%
Falling apple	0.0%	0.0%
Falling marker	0.0%	0.0%
Falling tape	0.0%	0.0%
Projectile	0.0%	2.3%
Rolling empty can	0.0%	0.0%
Rolling full can	0.0%	0.0%
Sliding book	0.0%	5.7%

Joint Physical Consistency (PC≥4 and rule adherence) pass rate assigned by VideoPhy2-AutoEval, real-world vs. generated, per experiment.