Fig. 3:

Our framework for domain-scalable UNIT. (a) Adding a new domain $C$ to the current models is easy that one only needs to train an ${\text{Encoder}}_C$ and a ${\text{Regressor}}_C$ on the new domain. (b) The training is performed on the single-domain data. We use a lightweight ${\text{Encoder}}_C$ to map the visual domain to GAN’s latent space (Sec. 3.2), then use a lightweight ${\text{Regressor}}_C$ to reconstruct input images from GAN’s latent space. We show that it is equivalent to aligning the perceptual structures of images of different domains. (c) During inference, we can use the trained encoder and regressor to perform translation between any two domains, as well as generate well-aligned multi-domain images by sampling from the GAN’s