Stochastic generation of realistic image content

Our goal is to generate novel photo-realistic images of a given object class (e.g. faces, trees) using a model trained from example images. To achieve this, we treat training images as samples from a texture with spatially varying statistics and synthesize using a modification of the patch-based method of Efros and Freeman. Unfortunately this generates images that are locally consistent, but globally unrealistic. To resolve this we also learn a weak global model of all the image pixels. This creates images with correct global structure but unrealistic local texture. We demonstrate for the case of faces that combining global and local models allows generation of realistic image content.