eduzhai > Applied Sciences > Engineering >

Image Augmentations for GAN Training

  • Save

... pages left unread,continue reading

Document pages: 20 pages

Abstract: Data augmentations have been widely studied to improve the accuracy androbustness of classifiers. However, the potential of image augmentation inimproving GAN models for image synthesis has not been thoroughly investigatedin previous studies. In this work, we systematically study the effectiveness ofvarious existing augmentation techniques for GAN training in a variety ofsettings. We provide insights and guidelines on how to augment images for bothvanilla GANs and GANs with regularizations, improving the fidelity of thegenerated images substantially. Surprisingly, we find that vanilla GANs attaingeneration quality on par with recent state-of-the-art results if we useaugmentations on both real and generated images. When this GAN training iscombined with other augmentation-based regularization techniques, such ascontrastive loss and consistency regularization, the augmentations furtherimprove the quality of generated images. We provide new state-of-the-artresults for conditional generation on CIFAR-10 with both consistency loss andcontrastive loss as additional regularizations.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×