Self-distilled StyleGAN: Towards Generation From Internet Photos
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
DescriptionWe adapt StyleGAN to work on raw, uncurated internet photos, resulting in high-quality image synthesis. Our self-distillation approach consists of two components: a generative-based, self-filtering of the dataset to eliminate outlier images and perceptual clustering of the generated images to detect inherent data modalities used to improve StyleGAN's "truncation trick".