Self-conditioned Generative Adversarial Networks for Image Editing
Event Type
Technical Paper
Interest Areas
Research & Education
Presentation Types
Registration Categories
Full Conference Supporter
Full Conference
Virtual Conference Supporter
Virtual Conference
Exhibitor Additional Full Conference
Exhibitor Full Conference
DescriptionGANs are susceptible to generative bias. They focus on the core of the distribution, leaving the edges behind. We argue that this bias is responsible for the collapse of latent-editing methods away from the distribution's core and propose a method for self-conditioning the GAN — improving treatment of rare modalities.