Towards the efficient adaptation of offline physically based methods for real-time rendering

Abstract

Rendering physically accurate caustics in real-time remains a persistent challenge due to their complex light interactions and high-frequency features. This thesis presents a comprehensive exploration into adapting oìine physically based rendering techniques for real-time caustic synthesis on modern GPU architectures. Central to this work is CandelaDXR, a GPU-accelerated light tracer that employs novel importance sampling strategies to improve convergence speed and visual adelity for caustics. By generating dynamic probability distribution functions conditioned on scene geometry, camera parameters, and material properties, CandelaDXR prioritises specular interactions and focuses sampling eêorts on perceptually relevant regions. To support this system, two auxiliary tools were developed: Anvil, a modular visual debugging platform for rendering pipelines, and Forge, an evaluation framework designed to facilitate reproducible, cross-system comparisons. These tools provide both insight and rigour in diagnosing rendering artefacts and validating algorithmic improvements. Additionally, the thesis introduces a spectral denoising pipeline tailored to the distinct characteristics of caustic signals, demonstrating the eêectiveness of Fourier, wavelet and curvelet-based transforms in preserving detail while reducing noise. Quantitative results across multiple scenes and viewpoints reveal signiacant performance gains, noise reduction, and perceptual improvements in CandelaDXR over baseline methods. Collectively, this work contributes a uniaed architecture for real-time caustic rendering, debugging, and evaluation, offering practical advances in both rendering theory and implementation for real-time physically based graphics.