Conditional image generation using feature-matching GAN

Generative Adversarial Net is a frontier method of generative models for images, audios and videos. In this paper, we focus on conditional image generation and introduce conditional Feature-Matching Generative Adversarial Net to generate images from category labels. By visualizing state-of-art discriminative conditional generative models, we find these networks do not gain clear semantic concepts. Thus we design the loss function in the light of metric learning to measure semantic distance. The proposed model is evaluated on several well-known datasets. It is shown to be of higher perceptual quality and better diversity then existing generative models.