acceptodds
Under review as a conference paper at ICLR 2027

DamEQ-Bench: A Physics-Aware Benchmark for Gravity-Dam Seismic Video Generation

Abstract

Generating physically plausible seismic-response videos of gravity dams remains challenging for modern video generation models. Unlike active motion in natural videos, gravity-dam vibration is a passive forced response governed by the prescribed seismic excitation, reservoir condition, and structural state; consequently, visual realism does not necessarily imply physically correct dynamics. To address this gap, we introduce DamEQ-Bench, a comprehensive benchmark for state-conditioned gravity-dam seismic video generation. DamEQ-Bench is constructed from controlled large-scale shaking-table experiments and comprises 198 seismic conditions spanning empty-reservoir and impounded-reservoir states, together with approximately 1,300 multi-view UAV recordings and synchronized sensor measurements as experimental references. We formulate gravity-dam seismic video generation as visually conditioned I2V/V2V tasks driven by prescribed seismic conditions and an initial image or video. To jointly assess visual quality and physical-response fidelity, we establish a comprehensive evaluation framework combining six general visual-quality metrics with eight response-oriented metrics that characterize structural dynamic response and modal characteristics. Under a unified protocol, we further evaluate sixteen representative open- and closed-source video generation models, together with two non-generative static control baselines, and investigate whether lightweight LoRA-based domain adaptation can improve seismic-response fidelity. DamEQ-Bench provides a unified, interpretable, and reproducible testbed for distinguishing visually plausible motion from structural responses that are consistent with prescribed seismic conditions.

Primary area:
datasets and benchmarks
Keywords:
seismic video generation, physics-aware benchmark, seismic response, structural dynamics, gravity dams

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_RYD9OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces DamEQ-Bench, a physics-aware benchmark for gravity-dam seismic video generation built from large-scale shaking-table experiments. The benchmark contains 198 controlled seismic conditions (empty/impounded reservoir states) with ~1,300 multi-view UAV recordings and synchronized sensor measurements. The authors formulate the task as I2V/V2V generation conditioned on seismic prompts and evaluate 16 video generation models plus 2 static baselines using 6 visual-quality and 8 response-oriented metrics. They also investigate LoRA-based domain adaptation.

Critical strengths

Genuine physical experimental data at scale: The 198 shaking-table conditions with synchronized multi-view video and sensor data represent a substantial, costly, and reusable resource. This is far more valuable than synthetic or internet-scraped benchmarks and enables physically grounded evaluation that cannot be replicated without such infrastructure. Rigorous separation of visual and physical metrics: The explicit demonstration that visually plausible videos can be physically incorrect—and the provision of interpretable diagnostic metrics—advances the community's understanding of what video generation models actually learn. This framing is consequential for how future physics-aware generation work should be evaluated.

Critical weaknesses

Limited generalization scope: All data comes from a single 1:50 scaled specimen at one facility. The benchmark evaluates "within-system condition generalization" only, as the authors acknowledge. This limits claims about whether models learn generalizable seismic dynamics versus memorizing this specific structure's response. While the authors are transparent about this, it constrains the benchmark's broader impact. Reference model circularity risk: The ODE-Response and related metrics depend on a learned Neural ODE as "ground truth" for conditions where direct sensor comparison may be unavailable. If this model is imperfect, the benchmark may penalize physically correct generations or reward ones that match the ODE's biases. The paper does not report ODE validation error against held-out sensor data.

Questions for authors

ODE validation: What is the prediction error of the Neural ODE against held-out sensor measurements? How sensitive are the reported rankings to ODE accuracy? Evaluation subset size: How many samples were actually evaluated per model in Tables 2/H3/H4? Are the reported differences statistically significant (e.g., confidence intervals or standard errors)? LoRA training protocol: What was the training set size, and were evaluation conditions strictly held out? Could the observed improvements simply reflect overfitting to the experimental scene?

1 comment
AuthorsOpenReview ↗

Response to Reviewer RYD9

Response to Reviewer RYD9

We sincerely thank the reviewer for recognizing the exceptional value of our large-scale experimental dataset and the importance of physically grounded video-generation evaluation. We have strengthened DamEQ-Bench through additional experimental validation, statistical significance analysis, ranking robustness tests, and a strictly partitioned LoRA evaluation protocol. Our key findings are summarized below.

  1. Large-scale experimental foundation and generalization.

DamEQ-Bench is built upon experiments conducted at the National Facility for Earthquake Engineering Simulation, featuring the world's largest shaking table in terms of platform dimensions and payload capacity (20 × 16 m; 1,350 t). We employed a 1:50 large-scale gravity-dam model designed according to physical similitude principles, enabling controlled investigation of seismic responses and complex dam–reservoir interactions.

The resulting dataset comprises 198 experimental seismic conditions, approximately 1,300 multi-view UAV recordings, and synchronized sensor measurements. This experimental infrastructure provides physically traceable observations that are difficult to obtain through synthetic data or conventional video collections.

By systematically varying seismic excitation and reservoir conditions, DamEQ-Bench establishes a rigorous testbed for condition-dependent physical generalization. Its experimental observations provide valuable foundations for evaluating learned structural dynamics and developing future physics-aware generative models.

  1. Physical validation and ranking robustness.

We have incorporated additional quantitative validation and metric-ablation experiments to strengthen the reliability of our physical evaluation.

An independent single-degree-of-freedom response-validation experiment under El Centro excitations of 0.10g and 0.20g achieves Pearson correlations of 0.837 and 0.877, NRMSE values of 0.074 and 0.070, and peak-response errors of 0.1% and 5.3%, respectively. Predicted and measured dominant frequencies also show close agreement. These results provide quantitative evidence of response-reconstruction accuracy in the independent validation system.

We further conducted ranking robustness analyses to assess the influence of learned physical references. After removing ODE-Response, the resulting rankings retain strong agreement with the full evaluation (Spearman ρ = 0.988; Kendall τ = 0.933). Under the independently reported noODE evaluation of base generators, the leading models remain Wan2.1-I2V, LTX-Video-2B-distilled, and ExtDM.

Together, these analyses demonstrate that the benchmark's comparative findings are supported by multiple complementary physical diagnostics rather than a single learned-response metric.

  1. Statistical significance and evaluation transparency.

Our benchmark encompasses 198 experimentally controlled seismic conditions, with multiple generated videos evaluated under each condition across different generative models. We further strengthen statistical reliability through 2,000 condition-level paired bootstrap resamples, accounting for dependencies among videos generated under the same seismic excitation.

The statistical analysis reports model-specific sample sizes and confidence intervals. For example, Wan2.1-I2V outperforms ExtDM by 0.026 in physical score, with a 95% confidence interval of [0.008, 0.045]. Additional metric-ablation analyses further demonstrate the robustness of the model rankings.

  1. LoRA adaptation and held-out evaluation protocol.

We established a condition-level training, validation, and testing protocol for the complete collection of 198 seismic conditions, with a proportional allocation of 82/45/71 conditions, respectively. The protocol is designed to separate seismic excitation records across training and evaluation subsets, enabling controlled assessment of generalization to unseen seismic conditions.

For Wan2.1-I2V, we selected the LoRA checkpoint at training step 2000, achieving a validation loss of 0.006674. The condition-level partitioning, checkpoint selection, and evaluation procedures provide a reproducible foundation for assessing domain adaptation and physical-response fidelity.

This protocol supports a systematic investigation of whether LoRA adaptation improves the generalization of video generation models across different seismic loading and reservoir conditions.

Reviewer_7ioCOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces DamEQ-Bench for generating gravity-dam seismic-response videos from an initial image or video prefix and a seismic-condition prompt. The dataset contains 198 shaking-table conditions, approximately 1,300 multi-view recordings, and synchronized sensor measurements. The benchmark evaluates 16 video generation models using visual and physical-response metrics, includes static controls, and examines LoRA adaptation.

Critical strengths
  1. The experimental dataset is potentially valuable. Controlled seismic loading, two reservoir states, multiple viewpoints, and synchronized measurements provide a useful basis for evaluating structural response beyond visual appearance (Sec. 3.3).

  2. The benchmark includes a broad range of models and useful static controls. Reporting visual and engineering-response scores separately helps expose differences that conventional visual evaluation alone would miss (Table 2).

Critical weaknesses
  1. The generation inputs may not sufficiently specify the response being evaluated. The physical model depends on the time-varying excitation (a_c(t)) (Eq. 1), but the documented generator inputs are an image or short prefix and a condition prompt (Eqs. 3–5). The example prompt gives summary attributes without explaining how the excitation waveform and its time alignment are represented. Yet evaluation measures trajectories and peak timing against a specific experimental response (Secs. 4–5.2). Unless the necessary forcing information is available, these errors may partly reflect missing inputs rather than inadequate physical modeling.

  2. The physical scores need stronger validation and clearer aggregation. StaticRepeat-I2V receives a MAC-Modal score of 0.527, although modal evaluation is described as retaining only reliably excited modes (Sec. 5.2; Table 2). The treatment of motionless outputs therefore needs explanation. Some reported results also require reconciliation: e.g. VideoCrafter1-I2V . These issues matter because reliable physical evaluation is the benchmark’s central contribution.

Questions for authors
  1. Task conditioning: What exact excitation-waveform and temporal-alignment information is included in the prompt? If only summary conditions are supplied, how should reference-specific trajectory and peak-timing errors be interpreted?

  2. Metric validation: How are zero-motion outputs and failed modal identification scored? What scores do real held-out videos obtain against the synchronized sensor references, and how accurately does the Neural ODE reproduce those held-out responses?

  3. Result consistency: Do Tables 2 and 4 use the same samples, normalization, and aggregation? Please reconcile the VideoCrafter1-I2V AF-Profile values

1 comment
AuthorsOpenReview ↗

Response to Reviewer 7ioC

Response to Reviewer 7ioC

We sincerely thank the reviewer for recognizing the value of our large-scale shaking-table experiments, synchronized measurements, and comprehensive visual and physical evaluations. We have further strengthened DamEQ-Bench through improved seismic-conditioning descriptions, validated modal-identification procedures, additional experimental verification, and consistent statistical reporting.

  1. Seismic conditioning and temporal alignment.

DamEQ-Bench is based on large-scale experiments conducted at the National Facility for Earthquake Engineering Simulation, using a 1:50 gravity-dam model designed according to physical similitude principles.

Complete ground-motion records obtained from the PEER database were processed and scaled according to the similitude requirements before being applied to the shaking table. The corresponding acceleration time histories and experimental metadata are preserved for all *98 seismic conditions.

We further extract physically meaningful information, including ground-motion identity, PGA, excitation direction, component-scaling factors, reservoir and structural states, viewpoint, and duration, to construct standardized I2V/V2V conditioning prompts (Appendix D.3).

The complete seismic records provide the experimental excitation and synchronized physical references, while the extracted attributes provide structured conditions for video generation. Video–sensor synchronization (Appendix C.8) and the unified physical-time evaluation grid (Appendices E.1–E.2) enable consistent comparisons of generated responses with experimental observations. The revised manuscript explicitly defines trajectory-phase and peak-timing metrics as reference-specific response-agreement measures.

  1. Static baselines and MAC-Modal validity.

Following the reviewer's observation, we conducted a comprehensive audit of the CPSD–FDD modal-identification procedure and introduced motion-energy and modal-validity checks.

The corrected evaluation achieves 100% rejection of invalid static outputs. Both StaticRepeat-I2V and StaticRepeat-V2V yield zero valid modal identifications. In contrast, real-video self-comparisons and low-amplitude experimental vibrations achieve 100% acceptance, while all temporally shuffled sequences are successfully rejected by the modal-validity criterion.

These positive and negative controls demonstrate the effectiveness of the revised procedure in distinguishing genuine structural vibration from static and temporally inconsistent motion. We have incorporated the validated eligibility criteria and valid-sample coverage into the updated evaluation protocol.

  1. Experimental validation and physical-reference reliability.

We have strengthened physical-reference verification through independent dynamic-response experiments, sensor-supported reconstruction, and ranking robustness analysis.

An independent single-degree-of-freedom response-validation experiment under El Centro excitations of 0.10g and 0.20g achieves Pearson correlations of 0.837 and 0.877, NRMSE values of 0.074 and 0.070, and peak-response errors of 0.1% and 5.3%, respectively. Predicted dominant frequencies also closely match experimental measurements.

We additionally provide experimental sensor-supported response reconstruction results. Representative Dry and Wet cases achieve Pearson correlations of 0.981 and 0.900, respectively, using a conditional transfer-function approach.

The revised evaluation clearly distinguishes direct sensor measurements, conditional transfer-function reconstructions, and Neural ODE-based references. Furthermore, removing ODE-Response preserves strong agreement with the full model ranking (Spearman ρ = 0.988; Kendall τ = 0.933 in the reported ablation analysis).

These complementary experiments and sensitivity analyses strengthen the physical basis and robustness of our evaluation framework.

  1. Consistency between Tables 2 and 4.

We have completed a sample-level verification and unified the evaluation subsets, normalization, and aggregation procedures.

For VideoCrafter1-I2V, the revised AF-Profile results are Dry = 0.410, Wet = 0.313, and Overall = 0.348, consistently computed under the updated Dry/Wet evaluation protocol.

The corresponding results in Tables 2 and 4 and the supplementary materials have been synchronized, with explicit sample-eligibility and aggregation rules to ensure reproducibility and cross-table consistency.

In summary, these revisions provide a stronger experimental and methodological foundation for DamEQ-Bench through physically grounded seismic conditioning, validated modal identification, quantitative dynamic-response verification, ranking robustness analysis, and consistent statistical evaluation. We sincerely appreciate the reviewer's constructive feedback, which has helped further strengthen the scientific quality of our benchmark.

Reviewer_ST5GOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this paper, the authors introduce a benchmark to evaluate video generators for seismic-response videos of gravity dams. This includes 1300 multi-view UAV recordings from 198 seismic conditions. The benchmark includes both visual quality and physical-response fidelity dimensions.

Critical strengths

I think the benchmark is valuable and will be useful to a specific group of the community.

This work provides a comprehensive result of SOTA I2V and V2V models on this benchmark.

Critical weaknesses

The technical contribution of this work is limited. Even though the DamEQ-Bench is niche and maybe valuable in some cases, this work is purely a benchmark with an evaluation protocol and doesn't introduce technical novelty.

I'm not sure if this work fits the deep representation learning conference, or it's better to submit to a more suitable conference/journal where it can reach out to more relevant audiences. I believe it will be more valuable in this way.

In this domain, will people use a simulation + sim2real pipeline or purely video generators? As people may care more about physics correctness, the right approach should be code + simulation + sim2real if needed.

Questions for authors

See weaknesses.

1 comment
AuthorsOpenReview ↗

Response to Reviewer ST5G

Response to Reviewer ST5G

We sincerely thank the reviewer for recognizing the value of DamEQ-Bench and its comprehensive evaluation of I2V and V2V models. We appreciate the concerns regarding technical novelty, relevance to ICLR, and numerical simulation, which we address below.

  1. Technical novelty beyond benchmark construction.

We agree that DamEQ-Bench does not introduce a new video generation architecture. Its methodological contribution lies in extending conventional video evaluation to experimentally measurable structural dynamics.

Our experiments were conducted using the world's largest shaking table in terms of platform dimensions and payload capacity (20 × 16 m; 1,350 t), with a large-scale 1:50 gravity-dam model designed according to physical similitude principles. A larger geometric scale can reduce certain scaling distortions and improve experimental representativeness, although not all multiphysics effects can be reproduced simultaneously. Beyond methodological development, this research supports the seismic safety assessment of critical water infrastructure, disaster risk reduction, and the protection of downstream communities.

As detailed in Section 4 and Appendix G, our framework combines four metrics adapted from Earthquake-Bench with four engineering-oriented response metrics: MAC-Modal, ODE-Response, Peak-Response, and AF-Profile. These characterize modal behavior, excitation-conditioned responses, peak characteristics, and height-dependent acceleration amplification. Supported by 198 seismic conditions and approximately 1,300 multi-view recordings, these metrics provide complementary physical diagnostics beyond conventional visual-quality assessment.

Thus, our contribution is an experimentally grounded evaluation methodology connecting generative learning with structural dynamics, rather than a new generative architecture.

  1. Relevance to ICLR and representation learning.

Although gravity-dam dynamics constitute a specialized application, our central question is relevant to generative learning: Can video generators reproduce measurable, condition-dependent physical dynamics rather than merely visually plausible motion?

Gravity-dam vibration is a passive response governed by seismic excitation and dam–reservoir interactions (Section 3.1, Eq. (1)). Our I2V/V2V tasks (Eqs. (3)–(5)) provide a controlled setting for examining this challenge. Evaluating 16 generative models (Sections 5.1–5.5) reveals their strengths and limitations in physical-response representation. While the present benchmark concerns one experimental structural system, its evaluation question extends to physically grounded generative learning and video-based world models.

  1. Why video generation instead of FEM and Sim2Real?

We agree that FEM and simulation-based approaches are essential for reliable structural engineering analysis. However, our objective is not to replace FEM, but to investigate how video generation can complement physical experiments and numerical simulation.

Physical experiments provide authentic structural responses, including complex dam–reservoir interactions. FEM provides interpretable physical mechanisms and predictions under specified material properties, excitation, and boundary conditions. Video generation offers a complementary data-driven approach to learning observable spatiotemporal dynamics, with the potential to incorporate simulation-derived physical constraints.

DamEQ-Bench provides an experimental foundation for evaluating this complementary approach. By comparing generated videos with measured responses and engineering-oriented metrics, we investigate whether generative models preserve meaningful physical dynamics rather than visual realism alone. We do not claim that current video generators outperform validated FEM or that a fully integrated experiment–simulation–generation framework has already been established.

In summary, DamEQ-Bench provides an experimentally grounded evaluation framework for physically meaningful video generation, establishing a foundation for future integration of physical experiments, numerical simulation, and generative learning.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.