Large Scale Image Completion via Co-Modulated Generative Adversarial Networks

Abstract

Co-modulated GANs link image-conditional GANs and unconditional modulated models to address large-scale image completion tasks. Co-modulation brings stochastic and conditional style representations together. To improve existing metrics for image completion, the proposed Paired/Unpaired Inception Discriminative Score (P-IDS/U-IDS) is robust to sampling size, captures subtle differences well, and correlates with human preferences. Experiments using co-modulated GANs lead to high quality and diverse results in free-form image completion and image-to-image translation tasks. We extend the findings by Zhao et al. by performing new image completion experiments to examine the biases of co-modulated GANs.

Tasks

Reproductions