Optimizing Light Samples for Real-Time Volumetric Clouds : Through Stochastic Rendering and Noise Reduction

Sammanfattning

Volumetric clouds are popular within real-time rendering applications such as games, as they provide realism to outdoor scenes. Therefore, rendering volumetric clouds at high quality is of interest to enhance user immersion. However, rendering volumetric clouds is known to be computationally demanding due to the high number of samples required per pixel during ray marching. To address this, this thesis investigates how stochastic rendering can reduce the rendering cost while preserving visual fidelity. The thesis focuses on optimizing the positions of a few sets of light samples for large volumes, such as the Walt Disney Animation Studios’ Moana cloud at half resolution, by randomly offsetting them and performing temporal and spatial filtering for noise reduction. The goal is to reduce the noise introduced by the stochastic rendering process. Various sources of noise, including White noise, Blue noise, Interleaved Gradient noise (IGN), scalar Spatiotemporal Blue noise (STBN), and scalar Filter-Adopted Spatio-Temporal noise (FAST), are evaluated along with denoising filters such as an Exponential Moving Average (EMA) temporal filter, and Gaussian sigma 1.0, Box 3x3/5x5, and Binomial 3x3/5x5 spatial filters. Performance is assessed through frames per second and image quality, measured using Root Mean Square Error (RMSE). Our results show that FAST noise, designed for specific temporal and spatial filters, consistently yields the lowest numerical error. While spatial filters offer only minor improvements post-convergence, they help reduce error in early frames. The findings highlight the importance of matching noise types to filters and suggest future user studies for perceptual validation in real-time scenarios.