SimilarityGAN: Using Similarity to Loosen Structural Constraints in Generative Adversarial Models

Abstract

Recently, generative adversarial networks have performed extremely well in image translation. When translating images current models adhere to a strict structural symmetry between the input and output images. This paper, presents a technique for image translation involving a pair of image domains that allows the output image to go beyond the structural symmetry constraints imposed by the input. By using a siamese model as the discriminator, we condition the generator to produce images that are only similar, rather than identical to the input. We show experimentally that using this modified loss a generator can generate realistic images for complex problems that only loosely adhere to the structure of the input.