Generative Adversarial Networks (GANs) are powerful models for generating data, but they often suffer from training instability. The Wasserstein GAN (WGAN) approach, which uses the Wasserstein distance to enforce a Lipschitz constraint on the critic, has shown promise in stabilizing training. However, weight clipping, a common method to enforce this constraint, can lead to issues such as poor sample quality and convergence problems. The authors propose an alternative method called Gradient Penalty (WGAN-GP), which penalizes the norm of the gradient of the critic with respect to its input. This method performs better than standard WGAN and enables stable training of various GAN architectures, including deep residual networks and language models with continuous generators. The paper demonstrates high-quality image generation on datasets like CIFAR-10 and LSUN bedrooms, and also achieves state-of-the-art performance on unsupervised CIFAR-10. The authors conclude that their method opens the path for stronger modeling performance on large-scale image datasets and language tasks.Generative Adversarial Networks (GANs) are powerful models for generating data, but they often suffer from training instability. The Wasserstein GAN (WGAN) approach, which uses the Wasserstein distance to enforce a Lipschitz constraint on the critic, has shown promise in stabilizing training. However, weight clipping, a common method to enforce this constraint, can lead to issues such as poor sample quality and convergence problems. The authors propose an alternative method called Gradient Penalty (WGAN-GP), which penalizes the norm of the gradient of the critic with respect to its input. This method performs better than standard WGAN and enables stable training of various GAN architectures, including deep residual networks and language models with continuous generators. The paper demonstrates high-quality image generation on datasets like CIFAR-10 and LSUN bedrooms, and also achieves state-of-the-art performance on unsupervised CIFAR-10. The authors conclude that their method opens the path for stronger modeling performance on large-scale image datasets and language tasks.