Image Generation and Style Transfer Using Conditional Generative Adversarial Networks

Generative Adversarial Networks have been utilized as a general-purpose solution to image domain transfer problems because of their ability to learn almost any data distribution. However, in most cases, the network architecture is tailored towards a specific application, and there has not been much exploration on using GANs for applications involving mixed datasets. In this work, we have studied the generalizability of Conditional Generative Adversarial Networks by testing their performance on a combination of image datasets, using the same type of input-output mapping. We show that a significant consequence of using multiple datasets is that the model can transfer characteristics such as style and design from one dataset to another. In order to facilitate this, we have modified the conditional GAN model to handle multiple conditional inputs, taking both the input image map and a class label. The results of our experiments show that a conditional GAN can successfully generalize to different types of input and transfer features of one class to others.