Efficient Light-Transport Simulation Using Machine Learning
The goal in this dissertation is the efficient synthesis of photorealistic images on a computer. Currently, by far the most popular approach for photorealistic image synthesis is path tracing, a Monte Carlo simulation of the integral equations that describe light transport. We investigate several data-driven approaches for improving the convergence of path tracing, leveraging increasingly sophisticated machine-learning models. Our first approach focuses on the specific setting of "multiple scattering in translucent materials" whereas the following approaches operate in the more general "path-guiding" framework. The appearance of bright translucent materials is dominated by light that scatters beneath the material surface hundreds to thousands of times. We sidestep an expensive, repeated simulation of such long light paths by precomputing the large-scale characteristics of material-internal light transport, which we use to accelerate rendering. Our method employs "white Monte Carlo", imported from biomedical optics, to precompute in a single step the exitant radiance on the surface of large spherical shells that can be filled with a wide variety of translucent materials. Constructing light paths by utilizing these shells is similarly efficient as popular diffusion-based approaches while introducing significantly less error. We combine this technique with prior work on rendering granular materials such that heterogeneous arrangements of polydisperse grains can be rendered efficiently. The computational cost of path construction is not the only factor in rendering efficiency. Equally important is the distribution of constructed paths, because it determines the stochastic error of the simulation. We present two path-guiding techniques that aim to improve this distribution by systematically guiding paths towards scene regions with large energy contribution. To this end, we introduce a framework that learns a path construction scheme on line during rendering while optimally balancing the computational rendering and learning cost. In this framework, we use two novel path-generation models: a performance-optimized spatio-directional tree ("SD-tree") and a neural-network-based generative model that utilizes normalizing flows. Our SD-tree is designed to learn the 5-D light field in a robust manner, making it suitable for production environments. Our neural networks, on the other hand, are able to learn the full 7-D integrand of the rendering equation, leading to higher-quality path guiding, albeit at increased computational cost. Our neural architecture generalizes beyond light-transport simulation and permits importance sampling of other high-dimensional integration problems.