Closed-Form Factorization of Latent Semantics in GANs

Pose on CelebA-HQ Faces (PGGAN) Orientation on LSUN Cars (StyleGAN) Expression on Anime Faces (StyleGAN) Body Pose on LSUN Cats (StyleGAN) Pose on ImageNet Magpies (BigGAN) Layout on LSUN Bedrooms (StyleGAN2) Figure 1. Versatile interpretable directions of the latent space unsupervisedly discovered in different GAN models including PGGAN [16], StyleGAN [17], BigGAN [4], and StyleGAN2 [18]. For each set of images, the middle one is the original output, while the left and the right are the output images by moving the latent code toward and backward the interpretable direction found by SeFa.