The thesis addresses the integration and evaluation of image reconstruction techniques in the Godot game engine. It focuses on modern resolution upscaling methods and an approach for locally enhancing selected parts of the image, with the aim of assessing their impact on image quality and rendering performance in real-time computer graphics.
The work presents the fundamental concepts of image reconstruction, anti-aliasing, temporal accumulation, and super-resolution. The practical part includes the integration of NVIDIA DLSS Super Resolution into Godot via the Streamline framework and the Vulkan graphics backend, a review of the existing AMD FSR2.2 implementation, and the development of a prototype for selective supersampling, in which a selected part of the scene is rendered at a higher resolution and combined with the base image.
The comparison of the methods is based on visual analysis, error maps, image quality measurements, and performance measurements. The PSNR, SSIM, and LPIPS metrics are used to evaluate deviations from the reference render. The results show differences between the methods, particularly in fine details, transparency, volumetric effects, and selected regions of interest.
The thesis presents the practical challenges of integrating temporal image reconstruction methods into a game engine, as well as the advantages and limitations of individual approaches to improving image quality.
|
