What are these animations?

They show sequence of images that are obtained by starting with the unperturbed generated image, and gradually increasing the perturbation to the latent activation values until the resulting generated image is misclassified in the way that is desired. For ImageNet, this is that the classifier outputs the target label. For CelebA-HQ, this is that the model predicts the following attributes positively: ‘Bald’, ‘Blond hair’, ‘Eyeglasses’, ‘Goatee’, ‘Grey hair’, ‘Moustache’, ‘No beard’, ‘Wearing hat’, ‘Wearing necklace’, and ‘Wearing necktie’.

Why are they interesting?

They make it much easier to understand the downstream effect of a perturbation to the latent activations of a generator. By watching the transformation, you get a much clearer sense of the features that are being manipulated than just by looking at the change to still images.

What are the details?

The ImageNet examples are perturbations to the activations at all layers of the BigGAN, as described in the paper. The classifier is the 'robust' ResNet50 model adversarially trained by Engstrom et al. The initial images and target labels used are those used in our experiments.
For CelebA-HQ, fewer examples are provided, but perturbations to subsets of layers are included as indicated in the directory names: 'front' means first 4 layers, 'mid' means next 3 layers, and 'back' means final 3 layers in the Progressive GAN.

Why aren't there more examples?

A more comprehensive dataset of such animations can be made, but an investment of time and compute is required. We will produce such a dataset for the camera-ready publication of the paper. The examples that are present now are enough to clearly understand what this will look like, and to get a good grip on the kinds of features being manipulated.

