LSN -GAN: A Novel Least Square Gradient Normalization for Generative Adversarial Networks

In this study, we introduce LSN-GAN, a new algorithm designed to improve Generative Adversarial Networks (GANs) by implementing Least Square Gradient Normalization. This technique adjusts gradients in the GAN's discriminator, making training more stable by avoiding common issues like gradient blow-up or disappearance. Our goal is to make GANs more reliable and enhance the quality of generated images. We tested LSN-GAN against existing models using the CIFAR-10 and LSUN Church Outdoor datasets. Our method was evaluated based on the Inception Score, which measures image quality and diversity, and the Fréchet Inception Distance (FID), which assesses the similarity between generated and real images. LSN-GAN showed improved results, achieving higher Inception Scores and lower FID values compared to models like SN-GAN and GN-GAN, indicating better image generation capabilities. Our experiments demonstrate that LSN-GAN not only proposes a novel approach to handling gradients in GANs but also significantly enhances the output image quality. This makes LSN-GAN a promising tool for advancing the field of generative models and their applications in image generation.