cs.CVSep 5, 2026

PhysWeep: Does a Video Generator Realize the Physics You Ask For?

Authors: Rasul Khanbayov, Hasan Kurban

Organizations: College of Science and Engineering, Hamad Bin Khalifa University, Doha, Qatar

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Image-to-video generators are often credited with absorbing physical dynamics as implicit world models, a claim the community currently checks with plausibility scores that ask whether a clip looks consistent with real-world motion. Plausibility is the wrong test on its own, because a clip can look natural while encoding the wrong value of the governing physical parameter, and no existing benchmark measures this gap directly. PhysWeep closes it with a fixed, label-free audit, treating a frozen generator as a black box, recovering the realized parameter from generated pixels, and reporting how often generation is trackable at all, how far the realized value sits from the requested one, and which, if either, of the literature's two proposed failure mechanisms the data support. A deterministic-simulator positive control confirms every score is exactly checkable. Applied to three open generators across six sweep axes, PhysWeep finds a specific, reproducible, previously undocumented failure. Conditional on producing trackable motion, two of the three generate confident, well-fit dynamics that converge to one of a small number of fixed, wrong values selected by the sampling seed rather than by the request, reproducing across two independent model families, two physical systems, and an independent tracker. It matches neither the prior reversion nor the case-based clamping the literature anticipates, because the reversion target is seed-conditional rather than a single global default, and a leave-one-out selection rule rejects both; the in-range faithfulness slope is statistically indistinguishable from zero wherever a response is estimable at all. A benchmark averaging over seeds would never see this: each sample is confidently locked to a wrong constant, exactly the failure a plausibility score is structurally blind to. We release the protocol, suite, and analysis code.

Figures & tables

Figure 1: The seed picks the physics, and it is not tracking noise. (a) CogVideoX-5B-I2V, projectile, in-range grid, text conditioning: recovered ∣g^∣|\hat{g}| is flat in the request and separated by sampling seed , with within-seed spread narrower than the plotted markers, while the same tracker and fitter recover the swept parameter from simulator video at slope 1.00011.0001 . Each point is a mean over that seed’s gated clips at that gg ( 80%80\% trackable in-range), not a single clip. (b) The same signature in a different architecture on a different law: DynamiCrafter returns near-perfect elasticity whatever restitution is asked for, at a fit quality that rules out a tracking artifact ( 25.0%25.0\% trackable in-range; one fitter-degenerate outlier, e^>1\hat{e}>1 , physically impossible for a restitution coefficient, is excluded from the plotted line). (c) Schematic, drawn to obey Eq. 3 : case clamping tracks the identity until θedge\theta_{\text{edge}} then goes flat, prior reversion is a straight line of slope α\alpha crossing the identity at θ0\theta_{0} , both defined in Sec. 4 as functions of θ\theta alone, so neither can express a response flat in θ\theta and offset by seed, which is why the selection rule rejects both ( Proposition 3 ). Sign convention in (a) is downward-negative; magnitudes are plotted.
Systemθ\thetaRecovered statisticSweep grid: in-range / out-of-range
Projectileggparabola curvature{5,7,9.8,12,15}\{5,7,9.8,12,15\} / {1.6,3,20,25}\{1.6,3,20,25\}
Damped pendulumω\omegazero-crossing period{2,3,4,5}\{2,3,4,5\} / {0.5,1,8,12}\{0.5,1,8,12\}
 ζ\zetaenvelope decay{.02,.05,.1}\{.02,.05,.1\} / {.3,.6,.9}\{.3,.6,.9\}
Bouncing balleebounce-height ratio{.6,.7,.8,.9}\{.6,.7,.8,.9\} / {.2,.35,.97}\{.2,.35,.97\}
Spring–masskkoscillation period{20,40,60,80}\{20,40,60,80\} / {5,10,200,400}\{5,10,200,400\}
Inclined slideμ\mualong-slope acceleration{.1,.2,.3,.4}\{.1,.2,.3,.4\} / {.01,.7,1.0}\{.01,.7,1.0\}
Table 1: The PhysWeep suite. Five deterministic renderers give six sweep axes, each with a closed-form map from a tracked statistic to the parameter, so recovery is exact and label-free. Units: gg m/s 2 , ω\omega rad/s, kk N/m; e,ζ,μe,\zeta,\mu dimensionless. Rendering settings and per-pair seed counts are in Appendix D .
Figure 2: The measurement pipeline. A known θ\theta drives a deterministic renderer, whose conditioning frames are handed to the frozen generator under audit; the resulting clip is tracked and fit in closed form to recover θ^\hat{\theta} , which is compared back against the known θ\theta to compute every metric in Sec. 4 . The fit-quality gate ( R2 ⁣≥ ⁣0.8R^{2}\!\geq\!0.8 ) is the only point where a generation can be excluded, and it is excluded as non-physical , not silently dropped. Only the generator (dark box) is untrusted; every other stage is exact and label-free by construction.
Trackability (%)proj.slideball
LTX-Video20\mathbf{20}35\mathbf{35}55
CogVideoX80\mathbf{80}0055
DynamiCrafter64\mathbf{64}80\mathbf{80}2525
Table 2: Coverage on the three axes short clips can identify at all : fraction of in-range clips whose tracked trajectory clears the R2 ⁣≥ ⁣0.8R^{2}\!\geq\!0.8 fit gate, under text conditioning. Shading is proportional to trackability, gray marks a measured zero (CogVideoX, inclined slide), and bold marks the five cells yielding enough gated clips for a slope. The remaining three axes are frame-budget-bound rather than a conditioning result ( Sec. 7 ) and are omitted here.
   ExtrapolationMechanism
ModelAxisnoutn_{\text{out}}EG [95% CI]PRIΔ\Delta AICSDI
LTX-Videoprojectile770.9850.985[ −0.33-0.33 , 2.732.73 ]1.0001.000−4.0-4.00.9950.995
LTX-Videobouncing ball44—1.0001.000−9.8-9.80.9850.985
LTX-Videoinclined slide440.2640.264[ −0.51-0.51 , 1.001.00 ]0.5140.514−11.3-11.30.5110.511
CogVideoXprojectile12121.0681.068[ 0.270.27 , 1.891.89 ] ∗0.9740.974−47.9-47.90.9190.919
CogVideoXbouncing ball99—0.9990.999−100.4-100.40.9970.997
Table 3: Mechanism adjudication , every model-axis pair with ≥4\geq 4 gated out-of-range points; all eight select as neither prior reversion nor case clamping, as Proposition 3 predicts. EG=PRE(Sout)−PRE(Sin)\mathrm{EG}=\mathrm{PRE}(S_{\text{out}})-\mathrm{PRE}(S_{\text{in}}) , undefined (—) on the bouncing-ball rows since θ\theta enters PRE\mathrm{PRE} ’s denominator and the in-range grid never clears the gate at a stable point. PRI ⁣= ⁣1\mathrm{PRI}\!=\!1 means the output ignores the request; ΔAIC<0\Delta\mathrm{AIC}<0 favours prior reversion over clamping; both use out-of-range data only. SDI\mathrm{SDI} is the column that decides the mechanism (read in text above).
Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Systemθ\thetaIn-range gridOut-of-range grid
Projectilegg{5,7,9.8,12,15}\{5,7,9.8,12,15\}{1.6, 3, 20, 25}\{1.6,\,3,\,20,\,25\}
Damped pendulumω\omega{2,3,4,5}\{2,3,4,5\} rad/s{0.5, 1, 8, 12}\{0.5,\,1,\,8,\,12\}
 ζ\zeta{0.02,0.05,0.1}\{0.02,0.05,0.1\}{0.3, 0.6, 0.9}\{0.3,\,0.6,\,0.9\}
Bouncing ballee{0.6,0.7,0.8,0.9}\{0.6,0.7,0.8,0.9\}{0.2, 0.35, 0.97}\{0.2,\,0.35,\,0.97\}
Spring–masskk{20,40,60,80}\{20,40,60,80\} N/m{5, 10, 200, 400}\{5,\,10,\,200,\,400\}
Inclined slideμ\mu{0.1,0.2,0.3,0.4}\{0.1,0.2,0.3,0.4\}{0.01, 0.7, 1.0}\{0.01,\,0.7,\,1.0\}
Appendix
Table A1 : Sweep grids per system. Each row has one swept axis; m≥5m\geq 5 seeded initial conditions per grid point.
ModelAxisseeds/pt (in/out)clips gen.clips gated
LTX-Videoprojectile5/105/1055551212
LTX-Videopendulum ω\omega5/555/5513013011
LTX-Videopendulum ζ\zeta5/155/15606000
LTX-Videobouncing ball5/155/15656555
LTX-Videospring-mass5/155/15808000
LTX-Videoinclined slide5/55/535351111
Appendix
Table A2 : Seed counts and clip yields per model-axis pair. “seeds/pt (in/out)” is the number of seeds run at each in-range and out-of-range grid value respectively; most pairs use the base 55 seeds on both splits, but pairs close to select_mechanism ’s four-gated-point leave-one-out threshold received a larger targeted follow-up batch on the out-of-range split only (LTX-Video/pendulum- ω\omega : extended to 5555 seeds/point, i.e. 110110 out-of-range attempts across its two out-of-range grid values, still yielding only 11 gated clip; CogVideoX/bouncing-ball: extended to 3333 seeds/point). No in-range seed count was extended beyond the base 55 , since the out-of-range split is what select_mechanism ’s adjudication gate requires.
ModelRepositoryLicenceLoaded via
LTX-VideoLightricks/LTX-VideoLTX-Video Open Weights (custom)diffusers==0.39.0
CogVideoX-5B-I2Vzai-org/CogVideoX-5b-I2VCogVideoX custom licence adiffusers==0.39.0
DynamiCrafter 512Doubiiu/DynamiCrafter_512weights: research/non-commercial bstandalone repo, subprocess
Qwen2.5-VL-7B-InstructQwen/Qwen2.5-VL-7B-InstructApache 2.0transformers==4.57.6
CoTracker3facebookresearch/co-trackerCC-BY-NC 4.0 ctorch.hub.load
Appendix
Table A3 : Every model used, all frozen and inference-only, loaded with no explicit commit/revision pin — each resolved to that repository’s default-branch head at run time, a gap against bit-for-bit reproducibility we disclose rather than paper over.
More in this topic

Physical Consistency in Video Generation

Physical consistency in video generation studies how to make and evaluate videos whose geometry, motion, interactions, and state changes obey real-world laws.

See all 191 papers

Explore similar work

CardsList
Mar 20, 2026

Physion-Eval: Evaluating Physical Realism in Generated Video via Human Reasoning

Video generation models are increasingly used as world simulators for storytelling, simulation, and embodied AI. As these models advance, a key question arises: do generated videos obey the physical laws of the real world? Existing evaluations largely rely on automated metrics or coarse human judgments such as preferences or rubric-based checks. While useful for assessing perceptual quality, these methods provide limited insight into when and why generated dynamics violate real-world physical constraints. We introduce Physion-Eval, a large-scale benchmark of expert human reasoning for diagnosing physical realism failures in videos generated by five state-of-the-art models across egocentric and exocentric views, containing 10,990 expert reasoning traces spanning 22 fine-grained physical categories. Each generated video is derived from a corresponding real-world reference video depicting a clear physical process, and annotated with temporally localized glitches, structured failure categories, and natural-language explanations of the violated physical behavior. Using this dataset, we reveal a striking limitation of current video generation models: in physics-critical scenarios, 83.3% of exocentric and 93.5% of egocentric generated videos exhibit at least one human-identifiable physical glitch. We hope Physion-Eval will set a new standard for physical realism evaluation and guide the development of physics-grounded video generation. The benchmark is publicly available at https://huggingface.co/datasets/PhysionLabs/Physion-Eval.

Qin Zhang, Peiyu Jing, Hong-Xing Yu +7

Physion Labs · Stanford University · MIT +2

Physical Consistency in Video GenerationVideo Generation EvaluationPhysical Reasoning in Video
AskPDF
Oct 19, 2026

PVBench: Quantifying Physical Plausibility in Video Generation with Object-Level Physical Measurements

Recent video generation models produce visually compelling outputs, yet their physical plausibility remains poorly measured. Physical plausibility asks whether generated objects obey geometry, kinematics, dynamics, and conservation laws. Existing benchmarks rely on coarse visual-language model judgments or target embodied world models rather than video generation itself. We introduce PVBench, a physical video benchmark that quantitatively measures the plausibility of generated videos across 46 dimensions organized into seven domains, spanning: i) instruction and semantic fidelity, ii) behavioral and causal consistency, iii) perceptual and temporal quality, iv) geometric and spatial consistency, v) material properties and field states, vi) kinematic and spatiotemporal persistence, vii) and mechanical plausibility with conservation laws. PVBench comprises 600 carefully crafted prompts, of which 30% involve physical embodied agents such as robot arms, humanoids, quadrupeds, and multi-agent systems. For challenging physically tractable dimensions, it pairs each with specialized vision models that derive verifiable, physically grounded measurements, such as object-level 6D pose, metric scale, surface normals, kinematics, and numerical conservation-law residuals. It thus avoids relying solely on subjective scoring, while supporting both text-to-video and image-to-video generation. Finally, we validate the agreement between automatic metrics and human judgments through human evaluation. Meanwhile, we evaluate mainstream open-source and closed-source models and present several insights. By releasing the benchmark, prompts, and calibration protocol, we aim to foster physically grounded video generation.

Kaiwei Chen, Moran Li, Zhenye Gan +5

APRIL Lab, Zhejiang University · Huazhong University of Science and Technology · Tencent Youtu Lab

Physical Consistency in Video GenerationVideo Generation EvaluationGeometric Consistency in Video GenerationTemporal Consistency in Video Generation
AskPDF
May 27, 2026

Proprio: Latent Self-Scoring and Inference-Time Refinement for Physically Plausible Video Generation

Modern video generative models produce visually impressive results, yet frequently violate basic physical principles. We propose Proprio, a training-free framework that enables a frozen video generator to assess and improve the physical plausibility of its own outputs. Inspired by proprioception, the biological sense of one's own movement, Proprio treats the model's flow residual under controlled latent perturbations as a self-scoring signal. Samples that are better explained by the generator's learned dynamics induce smaller and more stable residuals. We aggregate this signal across timesteps and perturbations, focus it on motion-relevant regions with a dynamic spatiotemporal mask, and use it for best-of-N search, gradient-based self-refinement, or both. Across text-to-video and image-to-video benchmarks, Proprio consistently improves physical plausibility, outperforming VLM-based scoring, and external world-model baselines in several settings. With TurboWan2.2, Proprio improves Physics-IQ from 32.2 to 37.5 (+16.5%) and VideoPhy2-hard physical commonsense from 45.6 to 55.0 (+20.6%). Human evaluation further shows that raters prefer Proprio-selected or refined videos for physical plausibility in roughly two-thirds of comparisons. These results suggest that frozen video generators contain actionable internal signals for evaluating and improving the physical plausibility of their own outputs.

Mariam Hassan, Kaouther Messaoud, Wuyang Li +1

1 École Polytechnique Fédérale de Lausanne (EPFL) · Télécom Paris, IP Paris

Physical Consistency in Video GenerationImage-to-Video GenerationVideo Diffusion ModelsT2V Generation
AskPDF