Input Ours GT GT Ours Input

Figure 1: We propose a machine learning approach to filter Monte Carlo rendering noise as a post-process. In our method, we use a set of scenes with a variety of distributed effects to train a neural network to output correct filter parameters. We then use the trained network to drive a filter to denoise a new MC rendered image. We show the result of our approach with a cross-bilateral filter for the KITCHEN scene (1200 × 800) on the left and with a non-local means filter for the SAN MIGUEL HALLWAY scene (800 × 1200) on the right. Both of these scenes are path-traced and contain severe noise at 4 samples per pixel (spp). However, our trained network is able to estimate the appropriate filter parameters and effectively reduce the noise in only a few seconds. Note, the tonemapping of the insets has been adjusted for best visibility. Scene credits: KITCHEN – Jo Ann Elliott; SAN MIGUEL HALLWAY – Guillermo M. Leal Llaguno. The most successful approaches for filtering Monte Carlo noise use feature-based filters (e.g., cross-bilateral and cross non-local means filters) that exploit additional scene features such as world positions and shading normals. However, their main challenge is finding the optimal weights for each feature in the filter to reduce noise but preserve scene detail. In this paper, we observe there is a complex relationship between the noisy scene data and the ideal filter pa