ABC-GAN : Adaptive Blur and Control for improved training stability of Generative Adversarial Networks

Generative Adversarial Networks (GANs) are well known for synthesizing images with impressive quality. However, the images are still often far away from being photo-realistic and the resolutions are typically limited to around 64 by 64 pixels due to training instability. In this work, we propose two simple techniques for improving the stability, training speed and image quality of GANs. First, we show that filtering the inputs of the discriminator with a blur kernel allows for increased image resolution and a significant quality improvement. Second, we propose a control strategy that adjusts the (expected) ratio between generator and discriminator iterations, for improved convergence speed. Our experiments show that the techniques are complementary and can speed up the convergence of DCGAN by a 5x factor and double the resolution, improving the quality significantly.