On the modifications of one method for rendering global illumination using a generative adversarial neural network

Abstract

Three variants for modifying one method for rendering 3D scenes using a generative adversarial network are proposed to generate realistic lighting in screen space. Experiments using software implementations of neural networks have shown that adding 25% resolution ray traced rendering to the generator input improves image quality when assessed by both the SSIM structural similarity index and the peak signalto- noise ratio PSNR. Adding noise to the discriminator layers improves the quality of network performance in both respects, but not significantly. Passing the ray traced rendering to the hidden convolutional layers of the generator instead of the input leads to a decrease in quality in both metrics compared to other modifications, but avoids unwanted effects on the edges of objects.