Earthquake-Bench: Video Generation Benchmark for Earthquake Simulation

Generating earthquake videos poses a fundamental challenge for modern video generation models. Unlike active motion patterns where objects move independently, earthquakes involve passive motion dynamics where ground displacement drives structural responses through complex force transmission mechanisms. Moreover, existing models cannot accurately bridge the gap between earthquake records and video generation, making it impossible to generate videos that faithfully represent specific seismic waveforms. To address these limitations, we introduce Earthquake-Bench, the first comprehensive benchmark for physics-based earthquake video generation. We formally define the earthquake video generation task by taking seismic waveforms and structural images/videos as input to generate realistic earthquake videos. We construct the benchmark by selecting common rigid-body building models, reproducing real-world earthquakes by scaling waveforms to drive a 3-axis shaking table, and recording the structural motion with synchronized cameras, creating physically-aligned video-waveform pairs. For evaluation, we propose four physics-aware metrics that compare generated videos against ground-truth videos and sensor data to assess displacement difference, edge difference, segmentation overlap, and temporal consistency. We evaluate seven open-source and nine closed-source state-of-the-art video generation models using this multimodal evaluation suite.