Neural Puppet

Our proposed work shows NIPS attendees executing various high level manipulations of their own face images using generative neural networks. Portrait photos are encoded into the latent space of a variational autoencoder where attribute vectors can be applied. These include opening and closing the mouth, or adding or removing a smile. Images are then decoded from the latent space and videos are created showing these effects. Additionally, participants can define their own attribute vector by having two photos taken and using the difference between them. This new attribute vector can then be applied to provided reference images as a one-shot generalization.