Abstract: Analytical methods based on simple linear relationships often lead to significant deviations between predicted results and measured data when describing the influencing factors of ground motions. To overcome this limitation, this study leverages the advantage of generative neural networks—which can automatically extract features and generate complete seismic motion time histories without relying on prior seismological knowledge—to systematically evaluate the performance of three typical generative neural network models (Variational Autoencoder VAE, Generative Adversarial Network GAN, and Denoising Diffusion Probabilistic Model DDPM) in ground motion simulation. The research employs 1, 451 horizontal ground motion records from 23 independent seismic events in the PEER database as the training dataset for uniform training and comparative analysis of the three models. Comprehensive evaluation of the simulation results in both time and frequency domains demonstrates that among the three models, DDPM exhibits the best simulation performance, followed by GAN, while VAE shows relatively inferior results. Notably, GAN simulation results display a significant enhancement of long-period components, whereas VAE exhibits a pronounced prolongation of duration. Comparative studies with four classical Ground Motion Prediction Equations (GMPEs) reveal that DDPM simulation results show good agreement with GMPE predictions, albeit with a slight systematic underestimation trend.