Parametric Integration with Neural Integral Operators

TL;DR

This paper introduces a real-time, neural network-based method for parametric integration in light transport, enabling noise reduction before shading and improving image quality in rendering.

Contribution

It proposes a neural integral operator for denoising that operates in real-time, is material-agnostic, and easily integrates with existing rendering pipelines.

Findings

Operates in real-time with a single frame of data

Compatible with existing denoisers and anti-aliasing techniques

Efficient training and easy integration with physically based rendering

Abstract

Real-time rendering imposes strict limitations on the sampling budget for light transport simulation, often resulting in noisy images. However, denoisers have demonstrated that it is possible to produce noise-free images through filtering. We enhance image quality by removing noise before material shading, rather than filtering already shaded noisy images. This approach allows for material-agnostic denoising (MAD) and leverages machine learning by approximating the light transport integral operator with a neural network, effectively performing parametric integration with neural operators. Our method operates in real-time, requires data from only a single frame, seamlessly integrates with existing denoisers and temporal anti-aliasing techniques, and is efficient to train. Additionally, it is straightforward to incorporate with physically based rendering algorithms.