KGN-Optimized Generative Adversarial Networks

Achieving stability during the training of Generative Adversarial Networks (GANs) continues to be a significant challenge. In this study, we introduce a novel algorithm, k-th Gradient Normalization (KGN), designed to tackle this issue by incorporating a flexible Lipschitz constant parameter. Our experiments demonstrate the efficacy of KGN in enhancing training stability and improving the quality of the generated samples. The results indicate that GANs trained using KGN consistently surpass existing methods on two datasets, as evidenced by improved Frechet Inception Distance(FID) and Inception Score(IS) metrics.