Source-linked AI summary
Neural Caching of Prefiltered Radiance for Specular Lighting
Dmitrii Klepikov, Vladimir Frolov
TL;DR
Specular lighting is challenging for radiance caching because it depends strongly on viewing direction and roughness. The paper predicts roughness-dependent prefiltered incoming radiance using a reflection-direction parameterization and split-sum reconstruction; on Bunny and Specular Sponza, it reports faster convergence and improved output quality relative to evaluated NRC baselines while maintaining the same performance level.
Problem
Specular reflection depends strongly on viewing direction and surface roughness, while estimating the reflected-light integral is computationally expensive.
Source: §1 · p. 2Method
Results
On Bunny and Specular Sponza, the reported comparisons show faster convergence and improved output quality relative to evaluated NRC baselines while maintaining the same level of performance.
Source: §5 · p. 4Takeaways & Limitations
The online-trained cache adapts during rendering, and the reported comparisons support real-time specular-lighting approximation on Bunny and Specular Sponza.
Source: §5 · p. 4The split-sum approximation introduces error at grazing viewing angles because its prefilter assumes the viewing direction is perpendicular to the filtering surface.
Source: §2.2 · p. 2
Abstract
from arXiv · showhide
Neural Radiance Caching (NRC) provides an online neural representation of scene illumination for real-time path tracing. This paper presents an NRC variant tailored to specular lighting through a reflection-direction parameterization and a roughness-dependent radiance target. Our network predicts prefiltered incoming radiance at individual surface points, which is combined with a precomputed BRDF integration map using the split-sum approximation to estimate outgoing radiance. On Bunny and Specular Sponza, the reported comparisons indicate faster convergence to the accumulated path-tracing reference and improved specular-lighting quality relative to the evaluated NRC baselines. Our online-trained method maintains real-time performance while adapting its cache during rendering.
1. Introduction
Specular lighting is difficult to cache because reflection depends strongly on viewing direction and roughness, while estimating reflected radiance is computationally expensive. The paper adapts online neural radiance caching with reflection-direction inputs and roughness-dependent prefiltered radiance for split-sum reconstruction.
1. Introduction
- Estimating the light-transport integral is computationally expensive, motivating approaches that cache previously computed lighting information.For reflection, the paper considers the specular component of the BRDF.Source: §1 · p. 2
- Traditional radiance caching relies on difficult-to-implement heuristics and can produce artefacts requiring special suppression.The paper contrasts this with NRC’s neural-network abstraction of the internal radiance cache representation [MRNK21].Source: §1 · p. 2
- The method uses reflected viewing direction as a cache input and learns roughness-dependent prefiltered incoming radiance for split-sum reconstruction.The cache and BRDF reconstruction together approximate the reflected component of the light transport equation.Source: §1 · p. 2
2. Previous Work
This section reviews microfacet BRDF modeling, split-sum approximation, and neural radiance caching as background for the proposed specular-lighting approach.
2.1. Microfacet Specular BRDF
- A microfacet specular BRDF models surface scattering using a statistical facet distribution, shadowing-masking, and Fresnel reflectance.The implementation uses a modified Trowbridge-Reitz (GGX) distribution, Schlick Fresnel, and a modified GGX shadowing-masking function.Source: §2.1 · p. 2
2.2. Split sum approximation
- Split-sum approximation separates reflected-radiance estimation into prefiltered environment radiance and a BRDF integration map.The practical prefilter assumes the viewing direction is perpendicular to the filtering surface, introducing approximation error at grazing angles; Figure 2 plots FLIP, SSIM, and MSE across rendering frames for two scenes and the evaluated NRC variants.Source: §2.2 · pp. 2–3, Figure 2 · p. 3
2.3. Neural radiance cache
- The paper situates neural rendering among approaches using neural representations for geometry-aware reconstruction and novel-view synthesis, including Neural Radiance Fields [MST∗20].The cited line of work includes subsequent papers [BMT∗21] [VHM∗22].Source: §2.3 · p. 3
- Neural radiance caching uses online-trained neural networks to store and predict indirect illumination for real-time path tracing.Prior work also reports hash-grid parametric encoding as a development for NRC.Source: §2.3 · p. 3
3. Proposed method
The method trains an online neural cache to predict roughness-dependent prefiltered incoming radiance, then uses a BRDF integration lookup to reconstruct outgoing specular radiance.
3. Proposed method
- Rendering uses three passes: generate multi-bounce training paths, trace shorter rendering paths, then query the cache and reconstruct outgoing radiance.Both path-tracing passes use Monte Carlo integration and importance sampling for continuation directions.Source: §3 · p. 3
- Training targets combine continuation-path radiance and throughput into RGB incoming radiance, stored with each cache query.Different sampled paths can yield different targets for the same query.Source: §3 · p. 3
- The cache predicts prefiltered incoming radiance from a surface-position encoding, reflected viewing direction, and roughness.Roughness conditions cached lighting on the reflection lobe.Source: §3 · p. 3
- Under squared-error loss, the predictor estimates the conditional mean of the training target given the cache query.The sampling distribution and target weighting determine the radiance filter represented by the cache.Source: §3 · pp. 3–4
- An exponential moving average of network weights reduces fluctuations during online adaptation, while the BRDF integration map reconstructs outgoing specular radiance.The reconstruction retains the split-sum approximation and its grazing-angle limitations.Source: §3 · p. 4
4. Experiments and Results
Across Bunny and Specular Sponza, the reported comparisons show faster convergence and improved image quality relative to evaluated NRC baselines, without discernible performance degradation.
4. Experiments and Results
- Image quality is evaluated per frame using FLIP, SSIM, and MSE against a reference accumulated over 4000 path-tracing frames.MSE is plotted on a logarithmic scale.Source: §4 · p. 4
- The modifications improve convergence and output quality without discernible performance degradation.The BRDF integration map is precomputed and queried during rendering; network input changes do not negatively influence other major NRC components.Source: §4 · p. 4
- The comparison includes the NRC baseline and a variant trained only on initial hits, because removing multi-hit training targets reduced output quality in the authors’ tests.The two variants are distinguished in the comparison plots.Source: §4 · p. 4
- FLIP indicates closer agreement with the accumulated reference than evaluated NRC baselines, and Bunny image quality is better than the NRC baseline after 1000 frames.Figure 2 indicates faster convergence; Figure 1 presents the qualitative comparison after 1000 rendering frames.Source: §4 · p. 4
5. Conclusion
The paper concludes that its NRC variant improves convergence and output quality on Bunny and Specular Sponza relative to evaluated NRC baselines while maintaining the same performance level.
5. Conclusion
- On Bunny and Specular Sponza, the reported comparisons show faster convergence and improved output quality over evaluated NRC baselines while maintaining the same performance level.The method is presented as a real-time specular-lighting approximation.Source: §5 · p. 4
- The variant predicts roughness-dependent prefiltered incoming radiance and uses split-sum reconstruction for outgoing radiance.The evaluated method uses hash-grid position encoding and a reflection-direction input.Source: §5 · p. 4
- The method retains the split-sum approximation’s grazing-angle limitations, and the authors identify dynamic-scene behavior as a topic for further investigation.Online training enables the cache to adapt during rendering.Source: §5 · p. 4