Date of Award


Document Type

Thesis (Ph.D.)

Department or Program

Department of Computer Science

First Advisor

Wojciech Jarosz


Light transport is the study of the transfer of light between emitters, surfaces, media and sensors. Fast simulations of light transport play a pivotal role in photo-realistic image synthesis, and find many applications today, including predictive manufacturing, machine learning, scientific visualization and the movie industry. In order to accurately reproduce the appearance of real scenes, light transport must closely approximate the physical laws governing the flow of light. Physically based rendering is a set of principles for codifying these laws into a mathematical model, and is the predominant rendering methodology today.

The representational power of this model is limited to the effects it chooses to capture. Simultaneously, simulating the model is an additional source of approximation error: The predominant solution framework in use today—Monte Carlo integration—produces the exact image predicted by the model typically only in the limit of infinite computation; at any finite time, an image contaminated with noise is obtained.

In this dissertation, we are concerned with improving the accuracy of physically based light transport. We achieve this both by improving the representational power of the model, and by making the rendering algorithms more efficient, leading to lower error at any given computational budget. In particular, we will investigate correlations and reuse: On the one hand, prevalent models in rendering assume natural processes to arise from random, independent events, and simulate them as such. We will show that for participating media—such as clouds, fog, or smoke—this assumption does not hold, and we introduce an augmented model that can faithfully represent such correlations. On the other hand, the types of solutions that satisfy the rendering problem show a great deal of correlation. Because all pixels in an image view the same scene, the mathematical problems to be solved are greatly interrelated. Where naive rendering algorithms treat each pixel in isolation, we will focus on reusing the same computation over many pixels, exploiting the natural correlations present and thus amortizing computational effort. We improve over prior work by leveraging additional insights about the structure of the rendering problem to allow a greater amount of reuse, and thus efficiency