Foveated Monte-Carlo Denoising

In this work, we propose a temporally-stable denoising system that is capable of reconstructing MC renderings in a foveated manner. We develop a multi-scale convolutional neural network that starts at a base (downsampled) resolution and denoises progressively higher resolutions. Our network learns to use the lower resolutions and the previous frames to denoise each foveal layer. We demonstrate how this architecture produces accurate denoised results at a much lower computational cost.