A Benchmarking Study of GAN Models for High-Quality Data Generation

This study evaluates the performance of two popular Generative Adversarial Network (GAN) architectures, Pix2Pix and BicycleGAN, on paired datasets, including Facade, Cityscapes and Maps. Its performance has been assessed using traditional metrics : FID and IS, as well as novel measures: infinite FID and infinite IS. Our results show that Pix2Pix excels in image fidelity, mainly for facades and cityscapes, with lower FID and higher SSIM scores. BicycleGAN surpasses Pix2Pix in diversity, particularly for Maps, with higher IS values. The trade-off between realism and diversity was highlighted, providing insight to select the appropriate GAN models for specific applications.