Divide and conquer policy for efficient GAN training

Senmao Li, Yaxing Wang, Joost van de Weijer, Fahad Khan, Jian Yang

Keywords: GANs, image generation

Abstract: Recent advances in Generative Adversarial Networks (GANs) have achieved impressive results for the purpose of generating high quality synthetic imagery. While capable of synthesizing high-fidelity images, these models often generate unsatisfactory images which fall outside of the data manifold. A considerable research effort has investigated the data manifold, either by simply discarding images having a lower probability according to the discriminator output, or by filtering real images which are within the sparse regions of the data manifold. While effective, these methods fail to get access to either the fake distribution or the real distribution. In this paper, we propose a divide and conquer policy for GAN training. We first introduce a new local data-manifold detector (LDMD), which estimates whether the generated images are inside or outside of the data manifold. With the proposed LDMD, we further introduce a noise replay mode if it is outside the manifold, and a fake sample reuse mode if it is inside the manifold. Extensive experimental results on a number of GANs variants (e.g., SAGAN,SNGAN,BigGAN and StyleGAN) demonstrate qualitatively and quantitatively that our method improves the GAN’s performance, resulting in more realistic images than previous methods as confirmed by a significant drop in the FID.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Generative models