Moment-Based OIT in 3D Gaussian Splatting

• Moment-Based OIT is a transmittance computation technique that uses statistical moments to achieve order-independent blending of semi-transparent volumes.
• It replaces traditional alpha blending and per-pixel sorting with efficient moment accumulation and closed-form Gaussian integrals.
• Empirical results show reduced artifacts and improved reconstruction quality in complex 3D scenes compared to conventional methods.

Moment-based order-independent transparency (OIT) is a transmittance computation technique designed for real-time rendering of semi-transparent and overlapping volumetric objects, particularly within 3D Gaussian splatting frameworks. By characterizing the entire distribution of volumetric density along each camera ray via statistical moments, moment-based OIT achieves physically faithful, order-independent compositing of translucent geometry, offering a significant advancement over traditional order-dependent alpha blending and per-pixel sorting approaches. Notably, the method avoids the need for ray tracing or sample sorting, providing both efficiency and high reconstruction quality in complex scenes (Müller et al., 12 Dec 2025).

1. Theoretical Foundations and Transmittance Calculation

Order-independent transparency in volumetric rendering aims to compute, for each camera ray $r(t) = o + t\,d$ (with $t \in [t_n, t_f]$), the fraction of incident radiance that remains unabsorbed after passing through all overlapping media. For a scene represented by $M$ weighted 3D Gaussians, the total extinction-density along the ray is

$$o(t) = \sum_{i=1}^M w_i \frac{1}{\sqrt{2\pi\sigma_i^2}} \exp\left[-\frac{(t-\mu_i)^2}{2\sigma_i^2}\right]$$

A key challenge is to accumulate contributions without imposing a rendering order. Rather than compute raw moments $\int t^k o(t) \,dt$ (which are numerically unstable at large $k$), the domain is warped via $t \mapsto z = g(t) \in [0,1]$ and compactly encoded as "warped-distance moments":

$$m_k = \int_{t_n}^{t_f} g(t)^k o(t) dt = \sum_i m_{k,i}$$

Each $m_{k,i}$ is computed via closed-form 1D Gaussian integrals, with the base cases and a three-term recurrence relation supporting efficient moment computation up to order $2n+1$. This per-pixel sequence $\{m_k\}$ succinctly summarizes the aggregated density profile along the ray.

From these moments, the cumulative absorbance $A(t)$ and thus the transmittance

$$T(t) = \exp\bigl[-A(t)\bigr], \quad A(t) = \int_{t_n}^t o(s)\,ds$$

are reconstructed using a classical moment-bounds scheme (Münstermann et al., 2018). The Hankel matrix $H_{ij} = m_{i+j}$ is assembled, and its kernel polynomial yields a quadrature rule $\{(x_j, w_j)\}$. Absorbance bounds $L(t), U(t)$ are thus defined in terms of $g(t)$ and these roots/weights, and the final empirical transmittance is computed as

$$\widehat{T}(t) = (1-\beta)\,\exp[-L(t)] + \beta\,\exp[-U(t)], \qquad \beta = 0.25$$

This approach ensures continuous, order-independent, and physically plausible transmittance estimation per pixel.

2. Per-Gaussian Volume Integration Without Sorting

Once the per-pixel transmittance profile $\widehat{T}(t)$ is reconstructed, rendering proceeds by evaluating each Gaussian's contribution independently. For Gaussian $i$, the ray segment $[t_n, t_f]$ is subdivided into $N$ importance-weighted intervals using

$$t_j = \mu_i + K\,\sigma_i\,\Phi^{-1}\left( \frac{j - \frac{1}{2}}{N} \right), \quad K \approx 3, \quad j = 1\ldots N+1$$

where $\Phi^{-1}$ is the inverse standard normal CDF. The local contribution $L_i$ to the pixel radiance is approximated as

$$L_i \approx \sum_{j=1}^N \left[\widehat{T}(t_n \to t_j) - \widehat{T}(t_n \to t_{j+1})\right] o_i(t_j) (Le)_i(d)$$

with $o_i(t_j)$ denoting the Gaussian density at each sample. Under the piecewise-constant approximation,

$$o_i(t_j) = -\frac{1}{t_{j+1}-t_j} \ln \left( \frac{\widehat{T}(t_{j+1})}{\widehat{T}(t_{j})} \right)$$

This step eliminates all sorting, as each particle is rendered independently with respect only to the globally accumulated per-pixel moments.

3. Order-Independent Rendering Pipeline

The rendering pipeline for moment-based OIT in 3D Gaussian splatting proceeds in a strictly order-independent fashion:

1. Frustum and sphere culling: All 3D Gaussians are culled using conservative geometric proxies.
2. Moment pass (additive blend): Each visible Gaussian projects a confidence-interval proxy. The pixel shader efficiently computes and accumulates its moment contributions $m_{k,i}(u,v)$ into “moment textures” $m_k(u,v)$.
3. Quadrature pass (additive blend): Each Gaussian is re-rasterized; for each pixel covered, its $N$ intervals $\{t_j\}$ are sampled, per-pixel moments $m_k(u,v)$ are fetched, $\widehat{T}(t_j)$ is reconstructed, and particle contributions $L_i$ and small consistency penalties $P_i$ are accumulated.
4. Opacity-correction pass: The zeroth moment $m_0$ yields true scene opacity $1-\exp(-m_0)$, allowing the radiance sum to be rescaled to match this value precisely, consistent with prior moment-OIT approaches.

Because each Gaussian’s radiance is derived solely from summed per-pixel moments, the final image is inherently order-independent, requiring neither depth-peeling nor per-pixel sorting.

4. Empirical Performance and Ablative Analysis

Moment-based order-independent transparency has been quantitatively evaluated against several competing methods across the MipNeRF-360, Tanks & Temples (train & truck), and DeepBlending datasets:

Method	MipNeRF-360 PSNR	MipNeRF-360 SSIM	MipNeRF-360 LPIPS	Tanks&Temples PSNR	Tanks&Temples SSIM	Tanks&Temples LPIPS	DeepBlending PSNR	DeepBlending SSIM	DeepBlending LPIPS
3DGS	27.43	0.813	0.218	23.72	0.846	0.178	29.46	0.900	0.247
StopThePop	27.31	0.814	0.213	23.16	0.843	0.173	29.92	0.905	0.234
Vol3DGS	27.44	0.820	0.201	23.67	0.851	0.174	29.61	0.905	0.242
EVER	25.60	0.772	0.299	22.59	0.842	0.199	28.12	0.891	0.353
Ours (MB3DGS)	25.96	0.760	0.245	22.18	0.825	0.194	29.14	0.900	0.248

On synthetic tests with intersecting Gaussians, only the moment-based OIT reproduces ground-truth volumetric blending. Alternative methods (alpha-blend or front-to-back) exhibit artifacting or color bias.

Ablation on Tanks & Temples (truck):

Ablation	PSNR	Time [h]	#points
w/o regularizer	24.14	3.45	$1.08 \times 10^6$
EWA-proxy (vs. ours)	24.10	6.20	$1.59 \times 10^6$
default ADC	23.50	12.88	$3.82 \times 10^6$
full MB3DGS	24.25	2.83	$1.06 \times 10^6$

Disabling the moment-continuity regularizer or employing the EWA 2D proxy both lower PSNR and increase computational cost. Opacity-based ADC increases point count and runtime, degrading reconstruction quality (Müller et al., 12 Dec 2025).

5. Relation to Classical and Contemporary OIT Techniques

Moment-based OIT extends the approach of screen-space moment analysis, previously introduced in rasterization-based order-independent transparency contexts, to the setting of 3D Gaussian splatting. Prior techniques rely on depth-peeling, per-pixel sorting, or coarse approximate blending which suffer from artifacts in scenes with complex volumetric occlusion and semi-transparency.

Unlike standard alpha blending or sorted splatting, moment-based OIT provides a closed-form, order-independent characterization of accumulated extinction—mitigating rendering artifacts and yielding closer alignment to physically-based light transport models. The analytic moment construction and associated recurrence support efficient accumulation and bounded memory usage even when rendering complex volumes.

6. Practical Considerations and Limitations

Moment-based OIT as implemented in MB3DGS is tailored for scenes representable by millions of 3D Gaussians, leveraging GPU rasterization for interactive performance. The method introduces minor overhead from the moment computation and the necessity for quadrature per Gaussian, but significantly reduces the number of particles and required runtime vis-à-vis sorting-based pipelines. The regularizer ensures moment continuity across pixels, further enhancing image fidelity.

A plausible implication is that moment-based OIT is best suited for moderate to high-density semi-transparent volumes where conventional splatting fails to resolve occlusion without visual artifacts. However, the technique currently depends on accurate per-pixel moment evaluation; numerical instability may arise at high moment orders or in cases with extreme density overlap. Future work could explore adaptive moment order selection or incorporate hybridizations with ray-casting for edge cases.

7. Summary and Outlook

Moment-based order-independent transparency in the context of 3D Gaussian splatting enables analytic, physically-aligned, and order-agnostic transmittance estimation for high-quality volumetric rendering. By gathering and utilizing per-pixel statistical moments, the method bridges the gap between fast splatting and physically accurate volume integration, with empirical results demonstrating superior artifact suppression and efficiency compared to state-of-the-art alternatives. Its adoption is likely to influence future works in real-time view synthesis, translucent media rendering, and physically-based graphics (Müller et al., 12 Dec 2025).