- [pdf] [supp]
ADEL: Adaptive Distribution Effective-matching Method for Guiding Generators of GANs
Research on creating high-quality, realistic fake images has engendered immense improvement in GANs. However, GAN training is still subject to mode collapse or vanishing gradient problems. To address these issues, we propose an adaptive distribution effective-matching method (ADEL) that sustains the stability of training and enables high performance by ensuring that the training abilities of the generator and discriminator are maintained in balance without bias in either direction. ADEL can help the generator's training by matching the difference between the distribution of real and fake images. As training is ideal when the discriminator and generator are in a balanced state, ADEL works when it is out of a certain optimal range based on the loss value. Through this, ADEL plays an important role in guiding the generator to create images similar to real images in the early stage when training is difficult. As training progresses, it naturally decays and gives model more freedom to generate a variety of images. ADEL can be applied to a variety of loss functions such as Kullback-Liebler divergence loss, Wasserstein loss, and Least-squares loss. Through extensive experiments, we show that ADEL improves the performance of diverse models such as DCGAN, WGAN, WGAN-GP, LSGAN, and StyleGANv2 upon five datasets, including low-resolution (CIFAR-10 and STL-10) as well as high-resolution (LSUN-Bedroom, Church, and ImageNet) datasets. Our proposed method is very simple and has a low computational burden, so it is expandable and can be used for diverse models.