Online Photon Guiding with 3D Gaussians for Caustics Rendering

TL;DR

This work addresses the challenge of rendering caustics with high fidelity in production renderers by introducing online photon guiding via a global 3D Gaussian mixture. It derives a novel directional transform to sample emission directions from a 3D Gaussian at any observation point, eliminating parallax issues common in prior directional models. The framework combines a scene-geometry-informed initializer with an adaptive light sampler and learns distributions online using gathered photons through KL-divergence optimization, enabling progressive convergence to high photon densities in caustic regions. Empirical results across multiple scenes show that the approach outperforms existing directional distributions and is more robust than state-of-the-art MCMC-based guiding in complex, production-scale scenarios. The method is compact, GPU-friendly, and integrates with a two-pass GPU path tracer, offering practical impact for high-quality caustic rendering in modern pipelines.

Abstract

In production rendering systems, caustics are typically rendered via photon mapping and gathering, a process often hindered by insufficient photon density. In this paper, we propose a novel photon guiding method to improve the photon density and overall quality for caustic rendering. The key insight of our approach is the application of a global 3D Gaussian mixture model, used in conjunction with an adaptive light sampler. This combination effectively guides photon emission in expansive 3D scenes with multiple light sources. By employing a global 3D Gaussian mixture, our method precisely models the distribution of the points of interest. To sample emission directions from the distribution at any observation point, we introduce a novel directional transform of the 3D Gaussian, which ensures accurate photon emission guiding. Furthermore, our method integrates a global light cluster tree, which models the contribution distribution of light sources to the image, facilitating effective light source selection. We conduct experiments demonstrating that our approach robustly outperforms existing photon guiding techniques across a variety of scenarios, significantly advancing the quality of caustic rendering.

Figures (10)

• Figure 1: Our method is integrated in a two-pass rendering process. After the photon pass and the path tracing pass, the distributions are refined using the collected photon data recorded in path tracing pass. With guiding of the learned distributions, the photon density gradually increases over iterations.
• Figure 2: The concept of photon guiding with 3D Gaussians: we fit a 3D Gaussian mixture with gathered photons and use it to guide photon emission. In (a), the light source uniformly emits multiple photons, however, only the yellow photon is actually gathered from the camera view. We use the first bounce location of yellow photon to fit a 3D Gaussian as shown in (b). Later when emitting photons, we use the 3D Gaussian to sample emission directions, so that more photons can be gathered, achieving higher density.
• Figure 3: Visualization of distributions for our directional transform of 3D Gaussian ($F_o$) and vMF ($F_v$) with different parameters. By choosing parameters meticulously, the two distributions achieve very similar shapes.
• Figure 4: The scene set we used for experiments: GlassOfWater, Dragon, Bistro, BistroStreet, Eyes, Kitchen, Ring, SunTemple.
• Figure 5: A portion of rendering results in comparison of rendering quality and efficiency across different distributions. The exposure of each inset is adjusted for better observation. (a) Similar results are achieved among G3D, H2D, and vMF under a point light scenario, indicating comparable guiding efficiency with the absence of parallax issues. (b) Under a small rectangular area light, results vary; G3D outperforms other distributions as it can accurately transform global positional distribution into directional distribution at any emission point without parallax issue. (c) When the distribution inside bounds of the caster mesh is not uniform, efficiency of bound-based guiding becomes low, while G3D can be optimized for a close approximation, achieving lower error.