CVPR2023

Domain Expansion of Image Generators

Yotam Nitzan, Michaël Gharbi, Richard Zhang, Taesung Park, Jun-Yan Zhu, Daniel Cohen-Or, Eli Shechtman

摘要

Zombie direction Domain Adaptation Dormant direction Figure 1. (center) Traversing the latent space of generative models along some directions changes the image significantly while traversing others has no perceptible effect. We call directions of the latter type dormant. (left) Domain adaptation methods, transform the entire generator from a source domain to a target domain, indicated by the color blue. (right) We introduce an approach for a new task -domain expansion. Instead of fully transforming the generator, we expand it to include new data domains. Our method learns to represent the new domain in a disentangled manner by repurposing a single dormant direction.