EST-GAN: Enhancing Style Transfer GANs with Intermediate Game Render Passes

In this paper we introduce EST-GAN, an approach to improve the realism of frames from unsophisticated game scenes. For this purpose, several Generative Adversarial Networks (GANs) structures are applied, which are extended to handle intermediate game render passes generated by conventional rendering pipelines. We present an image-to-image translation method that transforms simple low-poly game scenes into the style of an elaborately produced video game. Through our experiments, we show that the involvement of G-buffer information has a significant impact on the results of these translations and the usage of these leads to a reduction in artifacts.