Смирнов Александр Олегович
Conditional Unpaired Image-to-Image Translation for Postprocessing Rendered 3D Scene
Прикладная математика и информатика
Improving the quality of graphics in video games is a rather expensive and time-consuming task, which often requires the work of several specialists in three-dimensional graphics at once. In this paper, we study the possibility of graphics improvement for old games by image translation for the original game, using modern variations of generative adversarial networks. Two new solutions were proposed for solving this problem, one of them uses semantic mapping of frames, the other uses raw source images. In the experimental part, we evaluate the quality of semantic mappings of individual game frames tackling the problem of applicability for proposed task. Also, the second proposed method is compared with one of the existing networks, showing that in some aspects proposed model is better.