Please note: This PhD defence will take place online.
Wenyou Wang, PhD candidate
David R. Cheriton School of Computer Science
Supervisor: Professor Toshiya Hachisuka
Light transport simulation is classically formulated through the rendering equation and the path integral over connected paths of surface vertices between light sources and sensors. Directly simulating transport over such complete, connected paths is computationally demanding, motivating alternative representations on which light transport can be formulated. Rethinking light transport beyond connected surface paths, this thesis explores two such directions: decomposing transport paths into reusable path segments for Monte Carlo integration, and shifting transport simulation from surface geometry into a learned latent probe space.
We first introduce a segment-based light transport framework that treats path segments as the basic units of sampling and integration. This formulation naturally accommodates the disconnected subpaths encountered in photon density estimation and path filtering, opening new possibilities for constructing transport paths from reusable segments. Built on marginal multiple importance sampling, this general-purpose framework provides practical strategies for sampling and combining segments. It enables a robust bidirectional path filtering method—challenging to realize under a vertex-based formulation—that outperforms state-of-the-art approaches in efficiency and convergence.
Because the segment-based path integral generally differs from the classical path integral, directly estimating it introduces bias. We next develop a method that removes this bias while preserving efficient segment reuse. Our key idea is to resample before debiasing: a recursive resampling procedure selects candidate paths in segment space without enumerating all possible path configurations. Applying a debiasing kernel only to the selected segment paths and combining the resulting estimator with canonical path sampling preserves unbiasedness with respect to the original path integral while retaining efficient segment reuse. Experiments show that our method matches or outperforms both biased segment-based rendering and unbiased path sampling.
Finally, we introduce TransportFormer, a learned model for diffuse global illumination. The model computes light transport on a spatial grid of probes, learning both what the probes represent and how transport occurs between them. Static geometry and materials are encoded independently of lighting, providing reusable scene context that guides nonlocal transport between probes as lighting changes. The resulting probe representation is then decoded to predict illumination across surfaces, amortizing the cost of transport computation. Trained across scenes, TransportFormer generalizes to unseen scenes without per-scene optimization and responds to lighting changes without temporal accumulation. Experiments demonstrate real-time relighting that closely matches path-traced references, achieving higher fidelity than the evaluated classical probe-based methods.
Together, these contributions demonstrate the benefits of reformulating light transport beyond connected, vertex-based paths. Reusable path segments and learned probe representations offer two distinct avenues for amortizing the cost of light transport simulation, advancing both offline Monte Carlo rendering and real-time global illumination.