Enhancing Training Stability in Generative Adversarial Networks via Penalty Gradient Normalization

In the evolution of the generative modeling do-main, Generative Adversarial Networks (GANs) have emerged as focal points of scholarly attention. We propose a new approach to normalize the discriminator gradients in GANs, termed Penalty Gradient Normalization (PGN), aimed at enhancing the stability of GAN training. By doing so, we enhance the discriminator's capacity to discern subtle variations in generated samples. Empirical studies conducted across a variety of datasets demonstrate that the application of penalty gradient normalization in the training of GANs results in state-of-the-art performance. This is evident in both the Inception Score and the Frechet Inception Distance metrics.