Consistent Embedded GAN for Image-to-Image Translation

Generative Adversarial Networks (GANs) have achieved remarkable progress in image-to-image translation tasks. However, these methods have the common problem that lacking the ability to generate both perceptually realistic and diverse images in the target domain. To tackle the problem, in this paper, we propose a novel model named Consistent Embedded Generative Adversarial Networks (CEGAN) for the image-to-image translation task. It aims to learn conditional generation models for generating perceptually realistic outputs and capture the full distribution of potential multiple modes of results by enforcing tight connections in both the real image space and latent space. To achieve realism, unlike existing GANs models that their discriminators attempt to differentiate between real images from the dataset and fake samples produced by the generator, the discriminator in our model distinguishes the real images and fake images in the latent space to alleviate the impact of the redundancy and noise in generated images. On the other hand, we learn a low-dimensional latent code that is distilled from the possible multiple distribution in the latent space to achieve diversity. By this way, our model avoids the problem of mode collapse and produces more diverse and realistic results. Extensive experimental results demonstrate the superiority of the proposed method.