Year of Graduation
Photorealistic post-processing of rendered 3D scenes
These days it’s conceivable to create near-photorealistic 3D graphics relatively cheaply. However, accomplishing full photorealism requires one order more resources. At the same time, there have been a few fundamental trends inside graphics industry: the strive for photorealism and becoming faster, even real-time. We can try to assist the problem using deep learning models. In this research we are applying different techniques to post-processing with growing difficulty level at the same time. First, we need to explore what kind of data sets we have to use to make the following research useful and meaningful. For a testing purposes we collected totally new dataset. Then find and modificate the model providing meaningful results. In this work we propose new CycleGAN framework with Wasserstein loss and showing translation results of our test images. With new model we propose the training process to converge as well as identity loss addition.