Text-to-audio-video (T2AV) generation is central to applications such as filmmaking and world modeling. However, current models often fail to produce physically plausible sounds. Previous benchmarks primarily focus on audio-video temporal synchronization, while largely overlooking explicit evaluation of audio-physics grounding, thereby limiting the study of physically plausible audio-visual generation. To address this issue, we present PhyAVBench, the first benchmark that systematically evaluates the audio-physics grounding capabilities of T2AV, image-to-audio-video (I2AV), and video-to-audio (V2A) models. PhyAVBench offers PhyAV-Sound-11K, a new dataset of 25.5 hours of 11,605 audible videos collected from 184 participants to ensure diversity and avoid data leakage. It contains 337 paired-prompt groups with controlled physical variations that drive sound differences, each grounded with an average of 17 videos and spanning 6 audio-physics dimensions and 41 fine-grained test points. Each prompt pair is annotated with the physical factors underlying their acoustic differences. Importantly, PhyAVBench leverages paired text prompts to evaluate this capability. We term this evaluation paradigm the Audio-Physics Sensitivity Test (APST) and introduce a novel metric, the Contrastive Physical Response Score (CPRS), which quantifies the acoustic consistency between generated videos and their real-world counterparts. We conduct a comprehensive evaluation of 17 state-of-the-art models. Our results reveal that even leading commercial models struggle with fundamental audio-physical phenomena, exposing a critical gap beyond audio-visual synchronization and pointing to future research directions. We hope PhyAVBench will serve as a foundation for advancing physically grounded audio-visual generation. Prompts, ground-truth, and generated video samples are available at https://github.com/imxtx/PhyAVBench.
Figures & tables
Figure 1. The proposed audio-physics sensitivity test.
Acoustic Physical Dimensions
Controlled Setting
Newly Collected
#GT Videos per Prompt
Evaluation Metric
Sound Source Mechanism
Fluid & Aerodynamics
Sound Propagation
Observer Physics
Time & Causality
Complex Coupling
TAVGBench ( Mao et al., 2024 )
✗
✗
✗
✗
✗
✗
✗
✗
1
AV-Align
SAVGBench ( Shimada et al., 2024 )
✗
✗
✗
✗
✓
✗
✗
✗
-
AV & Spatial-Align
Verse-Bench ( Wang et al., 2025a )
✗
✗
✗
✗
✗
✗
✗
✓
1
AV-Align
JavisBench ( Liu et al., 2026a )
✗
✗
✗
✗
✗
✗
✗
✓ (partial)
1
AV-Align
VABench ( Hua et al., 2025 )
✓
✓
✗
✗
✓
✗
✗
✗
0
AV & Stereo Align
Table 1. Comparison with the existing audio-video generation benchmarks.
Figure 2. The data curation pipeline of PhyAVBench.
Model
AV Quality
Semantic Consistency
AV Synchronization
Physical
FVD ↓
KVD ↓
FAD ↓
TV-IB ↑
TA-IB ↑
CLIP ↑
CLAP ↑
AV-IB ↑
DeSync ↓
AVH ↑
Javis ↑
CPRS ↑
Text-to-Audio-Video (T2AV)
Commercial (Closed-source)
Sora 2 ( OpenAI, 2025 )
578.70
72.04
4.25
0.3415
0.2028
0.2713
0.3592
0.2137
0.0370
0.2118
0.1803
0.4512
Wan 2.6 ( Wanxiang, 2025 )
656.62
108.25
1.48
0.3434
0.1820
0.2792
0.3858
0.2143
0.2645
0.2141
0.1846
0.4120
Seedance 1.5 Pro ( Seedance et al., 2025 )
968.03
315.38
3.09
0.3274
0.2258
0.2663
0.3600
0.2794
0.0240
0.2754
0.2412
0.4044
Table 2. Quantitative evaluation on PhyAVBench. We categorize models by task (T2AV, I2AV, V2A) and source type. Bold and underlined values denote the best and second-best results within each task category, respectively.
Figure 3. Qualitative comparison of generated audio-visual results on PhyAVBench.
Figure 4. The audio-physics grounding performance of SOTA T2AV, I2AV, and V2A models measured by CPRS.
Figure 5. Human evaluation results.
Metric
−FAD
TA-IB
CLAP
AV-IB
−DeSync
AVH
Javis
CPRS
P.C.
0.51
0.77
0.86
0.76
0.60
0.72
0.71
0.92
Table 3. Pearson correlation (P.C.) across T2AV models, using per-model PVR-MOS means. For lower-is-better metrics, the score is defined as −DeSync and −FAD .
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
Major Dimension
m Code
Sub Category
c Code
Test Point
t Code
Example
Sound Source Mechanics
m01
Material Properties
c01
Hardness / Damping
t01
Wrench dropping on concrete vs. carpet
Density
t02
Solid vs. hollow ball impact
Surface Texture
t03
Scrapping wood vs. metal
Object Geometry
c02
Size
t04
Dropping large vs. small bottoles
Shape
t05
Tapping long tubes vs. flat plates
Thickness
t06
Thin vs. thick steel plates
Appendix
Table 4. The fine-grained test points and illustrative examples of PhyAVBench.
Figure 6. Prompt groups and newly recorded video examples in PhyAV-Sound-11K.
Figure 7. The duration distribution of PhyAV-Sound-11K.
Figure 8. The difficulty distribution of PhyAV-Sound-11K.
Rank
Task
Model
CPRS ± std
Cos ± std
p± std
f(p)± std
01
I2AV
MOVA
0.3613 ± 0.1891
0.0853 ± 0.2335
0.172 ± 0.526
0.1799 ± 0.2841
02
I2AV
LTX
0.3285 ± 0.1544
0.0652 ± 0.1712
0.148 ± 0.579
0.1244 ± 0.2452
03
I2AV
Ovi
0.3198 ± 0.1429
0.0580 ± 0.1840
0.104 ± 0.386
0.1106 ± 0.2177
04
I2AV
UniVerse-1
0.2729 ± 0.0812
0.0084 ± 0.1172
0.018 ± 0.283
0.0416 ± 0.1236
01
T2AV
Sora 2
0.4512 ± 0.2149
0.1694 ± 0.2213
0.389 ± 0.580
0.3177 ± 0.3432
02
T2AV
Wan 2.6
0.4120 ± 0.2173
0.1209 ± 0.2261
0.275 ± 0.643
0.2635 ± 0.3455
Appendix
Table 5. CPRS ranking based on the LAION-CLAP embedding model.
Rank
Task
Model
CPRS ± std
Cos ± std
p± std
f(p)± std
01
I2AV
MOVA
0.3688 ± 0.2001
0.0501 ± 0.1925
0.128 ± 0.712
0.2125 ± 0.3274
02
I2AV
LTX
0.3125 ± 0.1422
0.0203 ± 0.1297
0.062 ± 0.485
0.1149 ± 0.2377
03
I2AV
Ovi
0.3005 ± 0.1269
0.0238 ± 0.1322
0.050 ± 0.394
0.0892 ± 0.2070
04
I2AV
UniVerse-1
0.2844 ± 0.1038
0.0061 ± 0.1066
0.010 ± 0.378
0.0657 ± 0.1711
01
T2AV
Sora 2
0.4288 ± 0.2214
0.1148 ± 0.1883
0.350 ± 0.756
0.3003 ± 0.3716
02
T2AV
Wan 2.6
0.3850 ± 0.1961
0.0766 ± 0.2145
0.286 ± 0.947
0.2317 ± 0.3206
Appendix
Table 6. CPRS ranking based on the ImageBind embedding model.
Rank
Task
Model
CPRS ± std
Cos ± std
p± std
f(p)± std
01
I2AV
MOVA
0.3677 ± 0.1972
0.0879 ± 0.2245
0.197 ± 0.666
0.1914 ± 0.3075
02
I2AV
LTX
0.3098 ± 0.1307
0.0602 ± 0.1834
0.096 ± 0.395
0.0895 ± 0.1961
03
I2AV
Ovi
0.2983 ± 0.0935
0.0760 ± 0.1701
0.085 ± 0.251
0.0586 ± 0.1318
04
I2AV
UniVerse-1
0.2720 ± 0.0794
0.0226 ± 0.1375
0.014 ± 0.218
0.0326 ± 0.1173
01
T2AV
Wan 2.6
0.3595 ± 0.1690
0.1179 ± 0.2207
0.162 ± 0.409
0.1600 ± 0.2519
02
T2AV
Kling-v2-6
0.3482 ± 0.1356
0.1347 ± 0.2054
0.185 ± 0.315
0.1291 ± 0.1962
Appendix
Table 7. CPRS ranking based on the CAV-MAE-Sync embedding model.
Metric
LAION-CLAP
ImageBind
CAV-MAE-Sync
P.C.
0.9230
0.9457
0.8927
Appendix
Table 8. Pearson correlation (P.C.) between CPRS metrics based on three embedding models and the PVR-MOS.
Figure 9. The example questionnaire for human evaluation.
Figure 10. The prompt for cross-model validation.
Figure 11. The prompt used to generate prompt groups based on the taxonomy of PhyAVBench.
Generative video-to-audio (V2A) models produce highly plausible soundtracks, but it remains unclear whether they capture the underlying physical processes. Existing evaluations emphasize perceptual realism and overlook physical correctness under controlled interventions. In this paper, we introduce FlatSounds, a benchmark that audits the physical reasoning of V2A models through: 1) controlled counterfactual pairs in which a single physical factor is varied, and 2) single-video pattern tests that probe internal consistency and directional trends. These settings test whether the generated audio correctly reflects specific physical properties and timings. Our evaluation of state-of-the-art models reveals a consistent trade-off: models rely more on text captions than the visual stream to infer physics and semantics. Captions generally improve physical and semantic accuracy, but paradoxically degrade temporal alignment. Our results highlight the need to move beyond audio quality toward learning physical processes directly from pixels. Finally, we find that our physics-based metrics correlate strongly with human preference tests on our own data. Project webpage: https://research.nvidia.com/labs/cosmos-lab/flatsounds/
Text-to-Audio-Video (T2AV) generation is rapidly becoming a core interface for media creation, yet its evaluation remains fragmented. Existing benchmarks largely assess audio and video in isolation or rely on coarse embedding similarity, failing to capture the fine-grained joint correctness required by realistic prompts. We introduce AVGen-Bench, a task-driven benchmark for T2AV generation featuring high-quality prompts across 11 real-world categories. To support comprehensive assessment, we propose a multi-granular evaluation framework that combines lightweight specialist models with Multimodal Large Language Models (MLLMs), enabling evaluation from perceptual quality to fine-grained semantic controllability. Our evaluation reveals a pronounced gap between strong audio-visual aesthetics and weak semantic reliability, including persistent failures in text rendering, speech coherence, physical reasoning, and a universal breakdown in musical pitch control. Code and benchmark resources are available at http://aka.ms/avgenbench.
Ziwei Zhou, Zeyuan Lai, Rui Wang +6
Fudan University · University of Science and Technology of China · Microsoft Research Asia
Joint audio-video generation models are rapidly approaching professional production quality, raising a central question: do they understand audio-visual physics, or merely generate plausible sounds and frames that violate real-world consistency? We introduce AV-Phys Bench, a benchmark for evaluating physical commonsense in joint audio-video generation. AV-Phys Bench tests models across three scene categories: Steady State, Event Transition, and Environment Transition. It covers physics-grounded subcategories drawn from real-world scenes, plus Anti-AV-Physics prompts that deliberately request physically inconsistent audio-video behavior. Each generation is evaluated along five dimensions: visual semantic adherence, audio semantic adherence, visual physical commonsense, audio physical commonsense, and cross-modal physical commonsense. Across three proprietary and four open-source models, we find that Seedance 2.0 performs best overall, but all models remain far from robust physical understanding. Performance drops sharply on event-driven and environment-driven transitions, and even strong proprietary systems collapse on Anti-AV-Physics prompts. We further introduce AV-Phys Agent, a ReAct-style evaluator that combines a multimodal language model with deterministic acoustic measurement tools, producing rankings that closely align with human ratings. Our results identify cross-modal physical consistency and transition-driven scene dynamics as key open challenges for joint audio-video generation.
Zijun Cui, Xiulong Liu, Hao Fang +8
University of Texas at Dallas · University of Washington · University of California, Los Angeles