Taming Normalizing Flows - Supplementary Videos

Below are different videos demonstrating the process of our method, when it is used to tame a model in order to change some attribute (Sec 4.2).
Each video is created by passing the same latent vectors through the tamed network, along the process.

Forget the blond hair attribute

The video below shows the process of "forgetting" the attribute "Blond Hair".

Change different attributes of the same identity

Each row in the video below corresponds to the same identity, and the different columns show how different attribute changes can be applied. For example, the 1st column shows the process of forgetting the "No Beard" attribute, i.e. increasing the beard.

Increase the attributes "Blond Hair" and "Smiling"

The video below shows the process of emphasizing the attributes "Blond Hair" and "Smiling" (see middle columns in Fig C.2).

Increase the attribute "Eyeglasses"

The video below shows the process of emphasizing the attribute "Eyeglasses".
As most of the images that contains eyeglasses in the training data (CelebA) contain males (~80%), these attributes are coupled, and we can see that when we increase the "Eyeglasses" attribute, we also increase the percentage of males (see identity in 2nd row, 4th column). This can be coped by uncoupling these attributes, selecting only images of females with glasses for the attribute increase, instead of the all images with eyeglasses.