Stochastic RaySplats

• Stochastic RaySplats are rendering algorithms that reformulate 3D Gaussian Splatting as unbiased Monte Carlo estimators to achieve efficient volume rendering.
• They replace deterministic per-ray sorting with randomized acceptance tests, ensuring scalable performance on modern GPUs and differentiable pipelines.
• Empirical studies show up to 3–4× speedup in rendering and training while maintaining high fidelity through effective noise reduction with sample accumulation.

Stochastic RaySplats are a class of rendering and reconstruction algorithms for 3D Gaussian Splatting (3DGS) that employ stochastic, sorting-free Monte Carlo estimators to achieve unbiased, efficient, and parallelizable volume rendering. By replacing deterministic per-ray or per-pixel sorting with randomized acceptance and importance sampling, Stochastic RaySplats address fundamental bottlenecks in both rasterization-based and ray-tracing-based 3DGS, providing scalability, enhanced control over cost–fidelity trade-offs, and compatibility with modern hardware and differentiable pipelines (Kheradmand et al., 31 Mar 2025, Xu et al., 24 Mar 2026, Sun et al., 9 Apr 2025).

1. Mathematical Foundations and Monte Carlo Estimators

Traditional 3DGS alpha compositing is defined as: $$C = \sum_{i=1}^M c_i \left[ \alpha_i \prod_{k < i}(1-\alpha_k) \right],$$ where $\alpha_i$ and $c_i$ denote the opacity and color of the $i$-th splat, assuming front-to-back sorting. The corresponding volume rendering integral along a ray is: $$C = \int_0^\infty \sigma(\mathbf{x}(t)) \exp\left(-\int_0^t \sigma(\mathbf{x}(s))\,ds\right) L(\mathbf{x}(t))\,dt,$$ with $\sigma$ as the opacity field and $L$ as the radiance (Kheradmand et al., 31 Mar 2025, Xu et al., 24 Mar 2026).

Stochastic RaySplats reinterpret these summations and integrals as expectations over discrete distributions defined by per-splat weights: $w_i = \alpha_i \prod_{k < i} (1 - \alpha_k),$ so that $C = \sum_i w_i c_i$. An unbiased estimator is obtained by sampling a single index $I$ with probability $\alpha_i$0 and returning $\alpha_i$1. Averaging $\alpha_i$2 i.i.d. samples, variance decays as $\alpha_i$3: $\alpha_i$4 This stochastic formulation is directly extensible to path tracing by analogous constructions for continuous opacity fields and supports both color and gradient (for optimization) estimation (Kheradmand et al., 31 Mar 2025, Xu et al., 24 Mar 2026, Sun et al., 9 Apr 2025).

2. Sorting-Free Stochastic Blending and Transparency

Stochastic RaySplats achieve sorting-free compositing through randomized acceptance tests. In rasterization, each splat fragment is assigned a uniform sample $\alpha_i$5; it is accepted if $\alpha_i$6, otherwise discarded. The Z-buffer resolves frontmost contributions among accepted fragments, yielding correct alpha blending over repeated sampling. In stochastic ray tracing, each candidate ray–Gaussian intersection performs a Bernoulli ($\alpha_i$7) test at the computed intersection, accepting at most one hit per ray in the single-sample variant (Kheradmand et al., 31 Mar 2025, Sun et al., 9 Apr 2025). This process implements unbiased importance sampling proportional to each splat's transmittance-weighted opacity.

A streamlined variant for $\alpha_i$8-sample rendering stores arrays of accepted hits, gathering multiple stochastic samples per BVH traversal and reducing estimator variance: $\alpha_i$9 Even with $c_i$0, results show that noise is minimal and controllable, especially as temporal or spatial sample accumulation is introduced (Sun et al., 9 Apr 2025).

3. Algorithmic Structure and Implementation

Stochastic RaySplats provision two primary algorithmic frameworks: rasterization-based stochastic compositor and ray-tracing-based stochastic intersection samplers.

Rasterization Pipeline (Kheradmand et al., 31 Mar 2025):

• Store each 3D Gaussian as $c_i$1 in a GPU buffer.
• Emit quads in the vertex shader; in the fragment shader, test $c_i$2.
• Z-buffer acceptance ensures per-sample compositing of the closest fragment.
• Repeat over $c_i$3 passes or draw buffers, accumulating results.

Ray Tracing Pipeline (Xu et al., 24 Mar 2026, Sun et al., 9 Apr 2025):

• Build a BVH over Gaussian AABBs, enabling sublinear intersection queries.
• Per ray, traverse BVH once, conducting randomized acceptance at candidate intersections with only small per-ray payloads (e.g., hit-depth, ID).
• Accept and shade the first (or $c_i$4) intersection(s).
• Extend to optimization/differentiable pipelines using a two-sample Monte Carlo gradient estimator, which remains unbiased: $c_i$5 where $c_i$6 are independent samples from $c_i$7.

Hardware and GPU Considerations:

• Aggressively minimize per-ray state and maximize warp occupancy.
• Use stateless RNG (cheap trigonometric hash seeded by intersection position).
• Support mixed-precision accumulation and fused forward/backward kernels for training.

4. Performance Characteristics and Empirical Results

Stochastic RaySplats demonstrate substantial performance gains over traditional sorting-based 3DGS in both rasterization and ray tracing:

Method	PSNR (dB)	Render/Train time	Key Notes
StochasticSplats, SPP=1	18.0	1.85 ms	3–4× faster
StochasticSplats, SPP=16	26.3	6.71 ms	Approaches 3DGS
Sorted rasterization 3DGS	29.0	5.60 ms	No cost/fidelity control
Stochastic RaySplats, 2-sample	30.34	0.37 s/iter	3× faster backward pass
3DGRT (sorted ray-tracing)	30.07	0.9 s/iter	Higher memory usage

On RTX 4090 and MipNeRF360, Stochastic RaySplats provide comparable or better PSNR with up to $c_i$8–$c_i$9 speedup in rendering and training (Kheradmand et al., 31 Mar 2025, Xu et al., 24 Mar 2026, Sun et al., 9 Apr 2025). Streaming single-sample ray tracing matches mesh-ray intersection performance and scales to millions of Gaussians on both high-end and memory-constrained GPUs. Variance declines as $i$0 ($i$1=samples per pixel); even $i$2 yields plausible results, with noise in the rendered image vanishing rapidly by $i$3 (Sun et al., 9 Apr 2025).

5. Extensions: Differentiable Ray Tracing and Relightable Pipelines

A central innovation of Stochastic RaySplats is sorting-free differentiable ray tracing, enabling efficient gradient estimation for learnable 3DGS representations. The two-sample estimator circumvents the need to recompute probability masses ($i$4) for all primitives, significantly reducing backward-pass time and memory overhead.

For relightable 3DGS, each Gaussian incorporates spherical harmonics coefficients or a compact neural material module. Per-Gaussian shading is performed after randomized index selection, with shadow- and visibility-randomization performed via the same two-sample Monte Carlo estimators. This brings physically consistent shadowing, reflection, and refraction effects to 3DGS pipelines absent in rasterization-based approaches (Xu et al., 24 Mar 2026).

Stochastic RaySplats naturally integrate into conventional path tracers, supporting seamless interaction with mesh objects and full global illumination. The Russian-roulette opacity tests at each hit replicate continuous semi-transparency and are compatible with both primary and shadow rays (Sun et al., 9 Apr 2025).

6. Limitations, Variance, and Future Directions

Stochastic RaySplats introduce noise for low per-pixel sample counts. Although temporal anti-aliasing and sample accumulation mitigate this, further variance reduction is an active area, with importance and stratified sampling or lightweight neural denoisers as plausible enhancements (Kheradmand et al., 31 Mar 2025, Sun et al., 9 Apr 2025).

Extensions towards fully volumetric integration (e.g., direct free-flight distance sampling within Gaussians) are possible but increase fragment-shader and intersection complexity. The present approach extends to arbitrary transparent primitives—including triangles and particles—by applying per-hit stochastic acceptance (Kheradmand et al., 31 Mar 2025).

A plausible implication is that the stochastic, sorting-free paradigm presented in Stochastic RaySplats may become foundational for scalable, high-fidelity, and hardware-efficient differentiable rendering pipelines in both photorealistic and neural scene reconstruction settings.