The Taub Faculty of Computer Science Events and Talks

GANs have recently gained attention due to their success in generating realistic new samples of natural images, yet the extent to which such models capture the statistics of full images is poorly understood. In this work we present a simple method to evaluate generative models based on relative proportions of samples that fall into predetermined bins. Applying our method to GANs shows that they typically fail to capture very basic properties of the distribution. As an alternative to the opaque and hard to train GAN, we learn a Gaussian Mixture Model and demonstrate on several datasets that it manages to model the distribution of full images and generate realistic samples. Finally, we discuss how our model can be paired with a pix2pix network to add high-resolution details while maintaining the basic diversity.