DIFFUSED INSTANCE CONDITIONED GAN

Junho Park, Jae-Pil Heo

Keywords: generative adversarial networks, GAN, conditional GAN, image generation

TL;DR: Improving image quality and mode coverage of GAN using diffusion based Gaussian mixture in feature space as partition guidance.

Abstract:

Recently, numerous data partitioning methods for generative adversarial networks has been developed for better distribution coverage on complex distribution. Most of these approaches aims to build fine-grained overlapping clusters in data manifold and condition both generator and discriminator with compressed representation about cluster. Although giving larger size of condition can be more informative, existing algorithms only utilize low dimension vector as condition due to dependency on clustering algorithm and unsupervised / self-supervised learning methods. In this work, we take a step towards using richer representation for cluster by utilizing diffusion based Gaussian mixture. Our analysis shows that we can derive continuous representation of cluster with Gaussian mixture when
noise scale is given. Moreover, unlike other counterparts, we do not need excessive computation for acquiring clustered representation. Experiments on multiple datasets show that our model produces better results compared to recent GAN models.

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