PGGAN (Progressive GAN)

NVIDIA

Structures found in this family (2)

Linear Direction, Curvature profile of the representation manifold

By model (3)

PGGAN (trained on CelebA-HQ, 1024x1024)

Linear Direction

Progressive GAN (trained on LSUN scene categories: bedroom, church, kitchen, etc.)

Linear Direction

PGGAN (trained on CelebA, 256px)

Curvature profile of the representation manifold

Observations (5)

GAN units align with object masks and causally control objects

Learned linear latent walks in GANs realize shift, zoom and color edits, with extent tracking training-data variability

InterFaceGAN SVM boundary normals are linear, editable, and partially entangled

Top eigenvectors of A^T A of a pretrained GAN generator's first affine layer are linear latent directions that edit semantic attributes

Image variation in several real pretrained GANs concentrates along surprisingly few, position-homogeneous major axes of the pullback metric

Papers

GAN Dissection: Visualizing and Understanding Generative Adversarial Networks (2019), On the "steerability" of generative adversarial networks (2020), InterFaceGAN: Interpreting the Disentangled Face Representation Learned by GANs (2020), Closed-Form Factorization of Latent Semantics in GANs (2021), The Geometry of Deep Generative Image Models and its Applications (2021)