WGAN-GP

WGAN-GP stabilizes GAN training by replacing weight clipping with a gradient penalty to enforce 1-Lipschitz continuity.

Gulrajani et al. introduced WGAN-GP in 2017 to solve the vanishing and exploding gradient issues inherent in standard Wasserstein GANs. By penalizing the norm of the critic's gradient with respect to its input (targeting a unit gradient norm), the model ensures stable convergence across architectures like ResNet and deep CNNs. This approach eliminates the need for careful hyperparameter tuning of weight clipping (usually set at 0.01) and prevents the critic from collapsing into simple functions. In practice, WGAN-GP enables the reliable generation of high-quality images (such as 128x128 LSUN bedrooms) without the volatile loss oscillations common in original GAN formulations.

https://arxiv.org/abs/1704.00028