Stochastic Ray Tracing for the Reconstruction of 3D Gaussian Splatting

Abstract

Ray-tracing-based 3D Gaussian splatting (3DGS) methods overcome the limitations of rasterization -- rigid pinhole camera assumptions, inaccurate shadows, and lack of native reflection or refraction -- but remain slower due to the cost of sorting all intersecting Gaussians along every ray. Moreover, existing ray-tracing methods still rely on rasterization-style approximations such as shadow mapping for relightable scenes, undermining the generality that ray tracing promises. We present a differentiable, sorting-free stochastic formulation for ray-traced 3DGS -- the first framework that uses stochastic ray tracing to both reconstruct and render standard and relightable 3DGS scenes. At its core is an unbiased Monte Carlo estimator for pixel-color gradients that evaluates only a small sampled subset of Gaussians per ray, bypassing the need for sorting. For standard 3DGS, our method matches the reconstruction quality and speed of rasterization-based 3DGS while substantially outperforming sorting-based ray tracing. For relightable 3DGS, the same stochastic estimator drives per-Gaussian shading with fully ray-traced shadow rays, delivering notably higher reconstruction fidelity than prior work.

Figures (11)

• Figure 1: We introduce a differentiable stochastic formulation for ray-traced 3DGS, enabling efficient reconstruction and rendering of both standard and relightable 3DGS scenes. In this figure, we show re-renderings of our reconstructed standard 3DGS model (a) as well as relightable ones under local point light (b), area light (c), and image-based environmental illumination (d).
• Figure 2: Our stochastic gradient estimation (\ref{['alg:ours']}) works by drawing a Gaussian $g_I$ along each camera ray (illustrated in dark blue) followed by another $g_K$ behind it. Then, the colors $c^+$ and $c^-$ of these Gaussians are computed for evaluating \ref{['Eq:ours_dCdbalpha']}. For standard 3DGS, these colors can be obtained by evaluating the associated spherical harmonics (SH). For relightable 3DGS, on the contrary, we compute $c^+$ and $c^-$ using Monte Carlo by tracing additional shadow rays (shown in yellow) toward a light source.
• Figure 3: PSNR vs. optimization time for sorted alpha blending (3DGRT) and our method.
• Figure 4: Reconstruction quality comparison between our method and StochasticSplats kv2025stochasticsplats.
• Figure 5: Equal-time comparison between our method and baselines. All methods run for the same wall-clock time. Our method produces comparable visual quality to 3DGS and outperforms 3DGRT.