Global Illumination (GI) is a strategy in computer graphics to add a certain degree of realism. Several approaches exist to achieve such a visual effect for computer-generated imagery. The most physically accurate approach is through conventional raytracing. It produces similar realistic results by trading-off time and computational-resource intensive, making them unsuitable for real-time usage. For more real-time usage scenarios, a set of faster algorithms exists that utilize post-processing on top of rasterization rather than performing ray-tracing. Despite being faster, the algorithms are resource intensive due to multiple post-processing stages and produce incorrect lighting results due to insufficient information on screen-space features. Hence, we propose a Generative Adversarial Network (GAN) based approach to bring real-time GI effects by following the path of conventional screen-space GI techniques. We acquire surrounding graphical information into account by going beyond screen-space and producing consistent GI effects that are comparatively closer to their physically correct ray-tracing counterpart. Moreover, our model provides a better quality of generated output compared to the other recent model which utilized a similar approach by scoring 0.90811 in SSIM, 0.00093 in MSE, and 30.30576 dB in PSNR on our developed dataset.
By extracting the following the dataset, you will find images for camera orientation, features of cycles and features of eevee. It is recommended to use 7zip for extracting the dataset.
7zip download link: Download (7-zip.org)