• Each image pair corresponds to a channel of an extended InceptionV1 network's 5c_Branch_3_b_1x1 layer.
  • The network was trained on the CelebA classification task.
  • Left images are maximizing the given channel's mean activation in pixel space.
  • Right images are maximizing the mean activation among StyleGAN2 outputs.
  • Click on the StyleGAN2 images for optimization progress videos.
  • The slider is a crude version of infinite scrolling, to avoid downloading all images at once.

    • neuron