Controllable Generative Adversarial Network

Although it is recently introduced, in last few years, generative adversarial network (GAN) has been shown many promising results to generate realistic samples. However, it is hardly able to control generated samples since input variables for a generator are from a random distribution. Some attempts have been made to control generated samples from a generator, but they have not shown good performances with difficult problems. Conditional GAN is one of the most popular GAN structure to control generated samples; however, conditional GAN frequently fails to generate samples with minor labels which are hardly distinguishable in a training set. Here, we propose controllable GAN (ControlGAN) in this paper. ControlGAN shows powerful performance to control generated samples by making the generator concentrate more on input labels. In this paper, ControlGAN is evaluated with a face image dataset and a room image dataset. Furthermore, we demonstrate that the training of ControlGAN is an actual learning, by feeding a label that had never been trained during the training process.