GSCache: Real-Time Radiance Caching for Volume Path Tracing using 3D Gaussian Splatting

Qi Wu, David Bauer, Qi Wu 0015, Hamid Gadirov, Kwan-Liu Ma

Abstract

GSCache is a real-time radiance caching system designed for volume path tracing that leverages the efficiency of 3D Gaussian Splatting (3DGS). By dynamically storing and learning path-space radiance, the system significantly reduces Monte Carlo noise and enhances image quality at low sample counts without increasing the overall computational budget.

Introduction

Photorealistic volume rendering is a cornerstone of scientific visualization, aiding in the interpretation of complex 3D data from medical scans and astrophysical simulations. However, the gold standard—volume path tracing—is notoriously slow due to the infinite scattering of light within volumes. This results in high pixel variance, visible as distracting noise. To achieve real-time performance, developers often resort to low sample counts or aggressive denoising filters that may obscure critical scientific details. While radiance caching offers a solution by reusing previously calculated light values, existing methods often struggle with integration or the unique scattering properties of volumetric data.

Methods

The proposed GSCache system repurposes 3D Gaussian Splatting into a dynamic, multi-level radiance cache. Unlike traditional spatial caches, GSCache organizes radiance by path length, maintaining a hierarchy of levels ($L_1, L_2, L_3, \dots$) that store attenuated radiance for specific path lengths.

Overview of Volume Path Tracing and the Path Space Representation used in GSCache.

Each level is represented by a point cloud of 3D Gaussians. To optimize performance, the system employs a "MIP-level" approach where earlier path levels (which contribute more significantly to the final image) contain a higher density of Gaussians, while deeper levels are sub-sampled. The cache is trained on-the-fly using noisy samples from the path tracer, guided by a specialized HDR loss function and AdamW regularization to ensure stability. Additionally, an early path termination heuristic decides whether to continue tracing a ray or terminate it into the cache based on its expected contribution, saving significant compute time.

Results

GSCache was evaluated across several diverse scientific and medical datasets, including high-resolution CT scans and astrophysical simulations.

Visual comparison and performance metrics (PSNR and render time) of GSCache against standard Path Tracing (PT) and Next-Event Estimation (NEE) baselines.

At 1 sample per pixel (SPP), GSCache consistently outperformed baseline path tracers and state-of-the-art Neural Radiance Caching (NRC). For example, on the Carp dataset, GSCache achieved a PSNR of 16.9, compared to the baseline's 12.4. The system demonstrates remarkable adaptability, typically converging to a high-quality state within 16 frames after a change in lighting or transfer functions. While the Gaussian splatting adds a small overhead (~11ms), it often results in a net performance gain by enabling early ray termination.

Discussion and Limitations

Despite its advantages, GSCache has certain limitations. The 3D Gaussian Splatting representation is memory-intensive, requiring approximately 28MB of VRAM, which may be a constraint for lower-end hardware. Furthermore, as a caching method, it introduces a trade-off between variance (noise) and bias (potential blurring). While the authors found this bias acceptable for visualization, it may not suit applications requiring absolute physical accuracy. Future work will explore spatial data structures to help the cache retain information about the entire scene over time, as the current implementation primarily focuses on the active viewport.

Summary updated

BibTeX
@article{wu2025gscache,
  author = {Qi Wu and David Bauer and Qi Wu and Hamid Gadirov and Kwan-Liu Ma},
  title = {{GSCache: Real-Time Radiance Caching for Volume Path Tracing using 3D Gaussian Splatting}},
  journal = {IEEE Transactions on Visualization and Computer Graphics},
  year = {2025},
  doi = {10.1109/tvcg.2025.3634634},
}
Figure