Abstract:
Generative Adversarial Networks (GANs) have achieved state-of-the-art performance for several image generation and manipulation tasks. Different works have improved the limited understanding of the latent space of GANs by embedding images into specific GAN architectures to reconstruct the original images. In this paper, we investigate the capabilities of the stochastic noise inputs of StyleGAN. We show that the stochastic noise inputs of a StyleGAN model can be used to transfer content and encode color information bypresenting an encoder architecture that, together with a pretrained and fixed StyleGAN model, is able to faithfully reconstruct images from virtually any domain. Thus, we demonstrate a previously unknown grade of generalizablility by training the encoder anddecoder independently and on different datasets. Our proposed architecture processes up to 45 images per second on a single GPU, which is approximately 32× faster than previous approaches. Finally, as one example application, our approach also shows promising results compared to the state of the art on image denoising tasks.
