Self-supervised latent walk optimization

methods / Direction Extraction / Self-supervised latent walk optimization

Techniqueintermediate

Learns a latent-space direction (a linear walk) of a frozen generator by gradient descent so that moving the latent code by alpha times the direction reproduces a programmatic image edit (shift, zoom, color) of the generated image, with no attribute labels or paired source/target images.

Used in (1 observation)

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

structure: Linear Direction · models: BigGAN (class-conditional, 1000 ImageNet categories; resolution not stated in Jahanian et al.), StyleGAN (trained on FFHQ, 1024x1024), StyleGAN (trained on LSUN Car), StyleGAN (trained on LSUN Cat), PGGAN (trained on CelebA-HQ, 1024x1024) · paper: On the "steerability" of generative adversarial networks