The paper introduces the concept of a "relativistic discriminator" to address limitations in standard Generative Adversarial Networks (GANs). The relativistic discriminator estimates the probability that real data is more realistic than fake data, or vice versa, compared to randomly sampled data. This approach is argued to account for the *a priori* knowledge that half of the data in a mini-batch is fake and aligns with divergence minimization. The authors propose two variants: Relativistic Standard GAN (RSGAN) and Relativistic average GAN (RaGAN), which are shown to be more stable and generate higher-quality data samples compared to standard GANs. Empirical results on the CIFAR-10 and CAT datasets demonstrate that RSGAN and RaGAN outperform standard GANs, LSGAN, and WGAN-GP in terms of stability and data quality, especially in challenging settings. The paper also discusses the mathematical implications and future research directions.The paper introduces the concept of a "relativistic discriminator" to address limitations in standard Generative Adversarial Networks (GANs). The relativistic discriminator estimates the probability that real data is more realistic than fake data, or vice versa, compared to randomly sampled data. This approach is argued to account for the *a priori* knowledge that half of the data in a mini-batch is fake and aligns with divergence minimization. The authors propose two variants: Relativistic Standard GAN (RSGAN) and Relativistic average GAN (RaGAN), which are shown to be more stable and generate higher-quality data samples compared to standard GANs. Empirical results on the CIFAR-10 and CAT datasets demonstrate that RSGAN and RaGAN outperform standard GANs, LSGAN, and WGAN-GP in terms of stability and data quality, especially in challenging settings. The paper also discusses the mathematical implications and future research directions.