Improved Training of Wasserstein GANs

TLDR

Generative Adversarial Networks are powerful but notoriously unstable to train, and although Wasserstein GANs improve stability, they can still produce low‑quality samples or fail to converge. The authors propose replacing weight clipping in WGANs with a penalty on the critic’s input‑gradient norm. This penalty enforces the Lipschitz constraint by penalizing the critic’s gradient norm, thereby stabilizing training. The gradient‑norm penalty eliminates the instability caused by weight clipping, outperforms standard WGANs, and enables stable training of diverse architectures—including 101‑layer ResNets and discrete‑data language models—while producing high‑quality samples on CIFAR‑10 and LSUN bedrooms.

Abstract

Generative Adversarial Networks (GANs) are powerful generative models, but suffer from training instability. The recently proposed Wasserstein GAN (WGAN) makes progress toward stable training of GANs, but sometimes can still generate only low-quality samples or fail to converge. We find that these problems are often due to the use of weight clipping in WGAN to enforce a Lipschitz constraint on the critic, which can lead to undesired behavior. We propose an alternative to clipping weights: penalize the norm of gradient of the critic with respect to its input. Our proposed method performs better than standard WGAN and enables stable training of a wide variety of GAN architectures with almost no hyperparameter tuning, including 101-layer ResNets and language models over discrete data. We also achieve high quality generations on CIFAR-10 and LSUN bedrooms.