600
paired prompt groups
600 carefully designed scenes, each with matched T2V and I2V variants.
PVBench: Quantifying Physical Plausibility in Video Generation with Object-Level Physical Measurements
Scroll to reveal the paper↓
CONDITIONAL & SEMANTIC FIDELITY✦BEHAVIORAL, CAUSAL & COMMONSENSE CONSISTENCY✦PERCEPTUAL & TEMPORAL QUALITY✦GEOMETRIC & SPATIAL CONSISTENCY✦MATERIAL PROPERTIES & FIELD STATES✦KINEMATIC & SPATIOTEMPORAL PERSISTENCE✦MECHANICAL PLAUSIBILITY & CONSERVATION LAWS✦CONDITIONAL & SEMANTIC FIDELITY✦BEHAVIORAL, CAUSAL & COMMONSENSE CONSISTENCY✦PERCEPTUAL & TEMPORAL QUALITY✦GEOMETRIC & SPATIAL CONSISTENCY✦MATERIAL PROPERTIES & FIELD STATES✦KINEMATIC & SPATIOTEMPORAL PERSISTENCE✦MECHANICAL PLAUSIBILITY & CONSERVATION LAWS✦
01 / WHY PVBENCHLooking real is not
Looking real is not
the same as being physical.
Modern video generators can produce visually compelling results while still violating geometry, kinematics, dynamics, and conservation laws. PVBench turns these failures into measurable object-level quantities.
46
evaluation dimensions
Three assessment layers organized across seven domains.
25
physical dimensions
Object-intrinsic properties, object–space relations, and dynamics across space and time.
30%
embodied-agent prompts
180 of 600 prompts (30%) feature robot arms, humanoids, quadrupeds, or multi-agent systems.
02 / THE BENCHMARKFrom pixels to
From pixels to
physical quantities.
PVBench evaluates 46 dimensions across seven domains and three layers. Physical quantities are measured numerically when possible, checked against physical laws, and supplemented by Qwen3.8-27B when direct measurement is difficult.
0114 dims
Semantic layer
Condition and semantic fidelity (CSF), plus behavioral, causal, and commonsense consistency (BC).
027 dims
Perceptual layer
Perceptual and temporal quality (PTQ): subject consistency, aesthetics, motion smoothness, flicker, dynamics, background, and imaging quality.
0325 dims
Physical layer
Four domains: geometry and spatial consistency (GS), material properties and field states (MFS), kinematics and spatiotemporal persistence (KP), and mechanical plausibility and conservation laws (MC).
EVALUATION PARADIGMS
17 specialized-model quantitative evaluations · 4 hybrid measurement + MLLM judgments · 4 pure MLLM semantic scores
03 / DATA CURATIONDesigned for
Designed for
measurable physics.
Each of the 600 prompt groups binds target dimensions to a physically observable scene; T2V and I2V variants share the same initial scene.
DATA PIPELINETarget dimension → scene specification → human verification → paired T2V / I2V generation
PROMPT DISTRIBUTIONDomain coverage, physical dimensions, prompt lengths, and key vocabulary
04 / EVALUATION & LEADERBOARDSee what models
See what models
actually preserve.
PVBench reports percentile scores for six models in both T2V and I2V. Radar plots and layer-level score maps make the gap between visual quality and physical plausibility directly comparable.
PAPER RESULTST2V / I2V
The figure combines seven-domain radar plots with semantic, perceptual, and physical-layer scores; no separate score-map panel is repeated below.
PAPER SCORES · PERCENTILE SCALE
6 models · T2V + I2VPVBench leaderboard
| Config. | Model | Semantic layer | Perceptual layer | Physical layer | ||||
|---|---|---|---|---|---|---|---|---|
| CSF | BC | PTQ | GS | MFS | KP | MC | ||
| T2V | SANA-Video 2.0 | 55.1 | 39.0 | 58.2 | 47.0 | 27.4 | 52.2 | 18.7 |
| Wan2.1 | 64.8 | 43.7 | 63.4 | 50.6 | 16.4 | 55.3* | 24.8 | |
| Wan2.2 | 73.1 | 53.6 | 68.1* | 49.2 | 18.1 | 52.9 | 20.5 | |
| LingBot-Video | 72.8 | 59.8 | 58.0 | 42.8 | 28.9 | 51.1 | 22.5 | |
| MiniMax H3 | 88.0* | 79.3* | 55.4 | 52.6 | 30.8* | 51.5 | 25.3* | |
| Seedance 2.5 | 92.2 | 83.3 | 68.4 | 52.2* | 47.2 | 55.6 | 30.0 | |
| I2V | SANA-Video 2.0 | 56.0 | 22.9 | 48.4 | 35.8 | 21.3 | 48.0 | 19.5 |
| Wan2.1 | 84.6 | 41.1 | 59.2 | 40.9 | 28.0 | 53.9 | 24.0 | |
| Wan2.2 | 88.6 | 55.1 | 66.6* | 45.8 | 31.1 | 52.6 | 23.6 | |
| LingBot-Video | 80.2 | 61.8 | 58.1 | 50.8 | 33.4 | 60.7 | 26.8* | |
| MiniMax H3 | 92.8* | 79.9 | 62.5 | 48.9 | 34.2* | 50.7 | 20.1 | |
| Seedance 2.5 | 96.8 | 79.7* | 67.7 | 50.3* | 43.4 | 54.4* | 29.0 | |
CSF: Conditional & Semantic Fidelity · BC: Behavioral, Causal & Commonsense · PTQ: Perceptual & Temporal Quality · GS: Geometric & Spatial · MFS: Material Properties & Field States · KP: Kinematic & Spatiotemporal Persistence · MC: Mechanical Plausibility & Conservation Laws. Bold marks the highest score within each configuration; * marks the second-highest.
05 / VISUAL COMPARISONCompare the physical
Compare the physical
cases side by side.
Browse one physical case at a time and compare five video-generation models. Switch between T2V and I2V while the key prompt, physical dimension, and evaluation method stay visible beside the videos.
Case 01 · matched benchmark sceneT2V · five-model view
PHYSICAL DIMENSIONR01 · DepthGeometry & spatial consistency
EVALUATION METHODVGGT-Ω · geometryObject-level measurement pipeline
KEY PROMPT
Selected prompt preview appears here.
25 physical cases · 5 models · T2V + I2VChoose a case to jump directly
OPEN RESEARCHMake video generation
Make video generation
physically usable.
PVBench helps researchers move from “looks good” to “obeys the world.” Explore the evaluation code on GitHub and read the paper on arXiv.
06 / CITATION
Cite PVBench
PVBench is available on arXiv: https://arxiv.org/abs/2610.26711.
@article{chen2026pvbench,
title = {PVBench: Quantifying Physical Plausibility in Video Generation with Object-Level Physical Measurements},
author = {Chen, Kaiwei and Li, Moran and Gan, Zhenye and Wang, Yabiao and Wang, Chengjie and Zhang, Kexin and Liu, Yong and Zhang, Jiangning},
year = {2026},
eprint = {2610.26711},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.26711}
}