Online Neural Path Guiding with Normalized Anisotropic Spherical Gaussians

TL;DR

This work tackles variance in unbiased physically-based rendering by presenting an online neural path guiding framework built around NASG, a closed-form, seven-parameter anisotropic spherical density. A compact 4-layer MLP predicts NASG parameters per shading point, and a KL-Divergence objective drives online learning from stochastic ray samples, with a learnable selection probability to blend NASG guidance with the BSDF. The NASG model offers analytic normalization and efficient sampling, enabling real-time, GPU-friendly online learning and integration with a wavefront path tracer. Empirical results show reduced variance and competitive performance against state-of-the-art neural and statistical guiding methods, highlighting practical potential for GPU-based production rendering with limited resources.

Abstract

The variance reduction speed of physically-based rendering is heavily affected by the adopted importance sampling technique. In this paper we propose a novel online framework to learn the spatial-varying density model with a single small neural network using stochastic ray samples. To achieve this task, we propose a novel closed-form density model called the normalized anisotropic spherical gaussian mixture, that can express complex irradiance fields with a small number of parameters. Our framework learns the distribution in a progressive manner and does not need any warm-up phases. Due to the compact and expressive representation of our density model, our framework can be implemented entirely on the GPU, allowing it produce high quality images with limited computational resources.

Figures (9)

• Figure 1: Overview of proposed framework. Area in blue frame is computation of a classic path tracer. Area in yellow frame shows how we learn the light distribution with a neural network and use the estimated explicit density model to guide further sampling.
• Figure 2: Visualization of NASG component $G$ with different parameters. Note that $G$ agrees with spherical Gaussian when $a = 0$.
• Figure 3: Training process of proposed network. We heavily apply automatic differentiation to adjust network weights based on estimated distribution and rendering estimation $\left< L\right>$ from the sampled scattering direction $\omega_i$.
• Figure 4: Eight scenes used to evaluate proposed GPU implementation. From left to right: Ajar, Bathroom, Bidir, Corridor, Kitchen, Glossy Tube, Box, Pool.
• Figure 5: A subset of rendering results in variance reduction comparison, rendered at 1024 samples per pixel (SPP). (a) Compared with incident radiance guiding, our full product guiding significantly improves results on specular surface: our "Full Product" result shows clear reflection of the metallic shelf, while "Li" or cosine weighted guiding only produces noise. MDMA and Kent distribution fail to learn an accurate distribution for narrow incident light, leading to higher variance. (b) Compared to traditional distributions such as SG, our NASG is more expressive and can learn a more accurate distribution for anisotropic lighting conditions, leading to lower variance.