A SIGGRAPH 2015 White paper from Nima Khademi Kalantari Steve Bako Pradeep Sen at the University of California, Santa Barbara has a novel approach to reducing noise in Monte Carlo renders, Using neural networks that can be “taught” how.

a machine learning approach for filtering monte carlo noise

As the presentation describes, the best current approaches for filtering Monte Carlo noise is a feature based filtered approach. The current systems will use cross-bilateral and cross non-local means filters that exploit additional scene features in the rendering. This can include features such as world positions and shading normals in the image.

The issue with this method has been to find the optimal weights for each feature in the filter that will reduce the noise, but still leave you with the details that you need in your rendered scene.

The paper, A Machine Learning Approach for Filtering Monte Carlo Noise realizes the complex relationship between the noise in the scene data and the ideal filter parameters. This relationship can be learned using a nonlinear regression model.

The proposed system uses a multilayer perceptron neural network and combines that with a matching filter, during both training and testing. To use the new framework, you first have to train it on a set of noisy images of scenes with a variety of distributed effects.

Then at run-time, the trained network can be used to drive the filter parameters for new scenes to produce filtered images that approximate the ground truth of the Monte Carlo render.