3DGS Rendering Equation

• 3DGS Rendering Equation is a paradigm where scenes are modeled using anisotropic 3D Gaussians to enable continuous control over radiance and transmittance.
• It approximates volumetric integration by projecting 3D Gaussians into 2D splats, blending them front-to-back for accurate depth and opacity handling.
• Extensions include GPU-efficient order-independent methods, physically-based lighting models, and dynamic, frequency-adaptive techniques for realistic scene synthesis.

3DGS Rendering Equation

3D Gaussian Splatting (3DGS) defines a rendering paradigm wherein a scene is represented as a collection of anisotropic 3D Gaussian primitives. Each Gaussian is parameterized by its spatial center, anisotropic covariance, opacity, orientation, and color (often via spherical harmonics), enabling continuous, explicit control over radiance and transmittance in a manner computationally efficient for real-time neural scene synthesis. The 3DGS rendering equation is central to the construction of both static and dynamic neural radiance fields via splatting-based rendering, and recent literature has advanced its efficiency, quality, and physical accuracy across diverse application domains.

1. Mathematical Formulation of the 3DGS Rendering Equation

In its classical formulation, the 3DGS rendering equation proceeds by approximating the volumetric rendering integral through discrete compositing of projected 3D Gaussian ellipsoids. Each Gaussian $i$ is projected from world to screen space via an affine transformation and Jacobian-based covariance linearization, resulting in a 2D elliptical "splat" whose spatial footprint and opacity inform per-pixel blending weights.

Given a ray or pixel intersection, all Gaussians overlapping the pixel are collected and sorted by camera depth, forming an ordered list $i=1,...,N$. The standard compositing equation is:

$$C(\mathbf{x}) = \sum_{i=1}^{N} c_i\,\alpha_i\,\prod_{j<i}(1-\alpha_j)$$

where $c_i \in \mathbb{R}^3$ is the (typically view-dependent) color of Gaussian $i$, and $\alpha_i$ the effective opacity at pixel location $\mathbf{x}$. The product term implements front-to-back "OVER" alpha blending, enforcing correct physical light transmittance ordering. Each opacity $\alpha_i$ is computed as:

$$\alpha_i(\mathbf{x}) = t_i \exp\left(-\frac{1}{2} (\mathbf{x}-\mu_i)^\top \Sigma_{2D,i}^{-1} (\mathbf{x}-\mu_i)\right)$$

where $t_i$ denotes maximum opacity and $\Sigma_{2D,i}$ is the screen-space covariance. Color $c_i$ is often defined as a spherical harmonics expansion evaluated in the direction from camera to primitive (Hou et al., 2024, Li et al., 2024). This pipeline originates from the volume rendering integral:

$$C(\ell) = \int_{t_n}^{t_f} T(t)\,\sigma(\ell(t))\,c(\ell(t))\,dt,\quad T(t) = \exp\left(-\int_{t_n}^{t} \sigma(\ell(s))ds\right)$$

with densities $\sigma(\mathbf{x})$ discretized as sums of anisotropic 3D Gaussians. The integral along a ray is efficiently approximated by single splat evaluations instead of dense sampling.

2. Approximations, Exact Variants, and Limitations

The standard 3DGS rendering equation relies on a critical approximation: the integral of a 3D Gaussian along the camera ray is replaced by evaluating its 2D projection at the pixel location. This "flatten-along-the-ray" approach, while computationally efficient, incurs accuracy limitations under wide field-of-view or non-pin-hole optics due to higher-order distortions of the true projected ellipsoid. For applications demanding higher photometric or geometric fidelity, exact ray-Gaussian integrals have been formulated (Huang et al., 29 May 2025):

$$\alpha_i = \sigma_i \exp\left(-\frac{1}{2} D_i^2\right)$$

$$C(o,d) = \sum_{i=1}^N c_i\,\alpha_i\,\prod_{j<i}(1-\alpha_j)$$

with $D_i^2$ denoting the squared perpendicular distance from the ray to the Gaussian mean in a "whitened" canonical basis.

Such exact volumetric integration admits closed-form solutions for the integral $\int G_{I,0}(o_u + t d_u) dt$, and outperforms 2D projection approximations in edge sharpness, aliasing reduction, and FOV invariance (Huang et al., 29 May 2025). The approach, though, involves more computational complexity in view culling and ray-Gaussian association, which is efficiently mitigated by the particle bounding frustum (PBF) and bipolar equiangular projection (BEAP) mechanisms.

3. GPU-Efficient and Order-Independent Extensions

A bottleneck of the original 3DGS equation is the non-commutative nature of alpha blending, necessitating per-pixel or per-tile front-to-back sorting. This incurs both CPU/GPU synchronization overhead and frame-to-frame "popping" artifacts. Weighted Sum Rendering (WSR) (Hou et al., 2024) introduces a commutative, order-independent transparency (OIT) approximation:

$$C = \frac{c_B w_B + \sum_{i=1}^N c_i \alpha_i w(d_i)}{w_B + \sum_{i=1}^N \alpha_i w(d_i)}$$

where $w(d_i)$ is a learned, monotonically decreasing function of depth $d_i$, and $c_B,w_B$ the background color and weight. This replaces the product term with a depth-based weighting, removing the sorting requirement, decreasing memory usage, and accelerating rendering (by 1.23× on mobile GPUs) without sacrificing image quality (PSNR, SSIM, LPIPS are preserved or improved). Parameterizations for $w(d)$ (direct, exponential, linear-correction) enable flexible modeling of attenuation and occlusion.

This formulation generalizes to device-efficient implementations, especially leveraging modern rasterization hardware (e.g., Vulkan), as blending becomes a per-pixel commutative accumulation without read/write hazards. The commutative nature further guarantees temporal stability, eliminating reordering-induced artifacts.

4. Physically-Based and Illumination-Aware Extensions

Recent extensions of the 3DGS equation integrate physically-based rendering semantics. Normal-GS (Wei et al., 2024) imbues each Gaussian with a local normal and models outgoing radiance via Kajiya's rendering equation, explicitly encoding the interaction of local normals with incident direct/diffuse lighting:

$$L_{\text{out}}(\omega_o) = \int_{\Omega^+} L_{\text{in}}(\omega_i) (\omega_i \cdot \mathbf{n}) f_r(\omega_i, \omega_o) d\omega_i$$

In the practical 3DGS context, the color per primitive is reparameterized as:

$$c_i = k_{D,i} (\mathbf{n}_i \cdot \mathbf{l}_i) + \theta_S(\phi_{\text{IDE}}(\omega_{r,i}), \mathbf{n}_i, \mathbf{f}_{v(i)})$$

where $k_{D,i}$ is the diffuse albedo, $\mathbf{n}_i$ the Gaussian normal, $\mathbf{l}_i$ the integrated directional illumination vector (IDIV), and $\theta_S$ predicts the specular term using integrated directional encoding (IDE). The anchor-based sharing of IDIV ensures compactness and computational efficiency. This enables direct interpolation between geometry and appearance, as well as improved normal recovery.

For remote sensing with variable illumination and shadows, ShadowGS (Luo et al., 4 Jan 2026) adapts the rendering equation to explicitly accumulate direct (sun-lit), ambient (sky), and near-surface secondary bounce contributions per pixel:

$C(u) = F(u) \cdot L_{\text{total}}(u)$

$$L_{\text{total}}(u) = S(u) + [1 - S(u)][L_{\text{sky}}(u) + L_n(u)]$$

where $S(u)$ is the order-independent sun shadow mask from hardware-accelerated ray marching, $L_{\text{sky}}(u)$ and $L_n(u)$ are global and local incident radiance from SH expansions, and $F(u)$ is the per-pixel albedo. The pipeline blends these components via discrete alpha blending of Gaussian attributes.

5. Extensions to Dynamic Scenes and Frequency-Adaptivity

Dynamic 3DGS representations (e.g., (Katsumata et al., 2023)) express time-varying centers and rotations for each Gaussian via Fourier and linear quaternion bases, leaving color, scale, and opacity time-invariant for memory efficiency:

$$\boldsymbol{\mu}(t) = \text{Fourier basis} \qquad \mathbf{R}(t) = \text{linear quaternion interpolation}$$

$$I(\mathbf{u}) = \sum_{i=1}^N \mathbf{c}_i \alpha_i \prod_{j<i}(1-\alpha_j)$$

This compactness enables real-time dynamic rendering at high framerates while maintaining rendering fidelity.

Frequency-adaptive extensions (e.g., 3DGabSplat (Zhou et al., 7 Aug 2025)) generalize the 3DGS equation from isotropic low-pass Gaussian kernels to Gabor-modulated primitives:

$$\text{Gabor}_k(\mathbf{x}) = G_k(\mathbf{x}) \left[1-\sum_{i=1}^F \omega_{k,i} + \sum_{i=1}^F \omega_{k,i} \cos(2\pi \mathbf{f}_{k,i}^\top(\mathbf{x} - \boldsymbol{\mu}_k))\right]$$

with corresponding modification in both rasterization and optimization, capturing high-frequency texture without resorting to a proliferation of primitives. The compositing equation remains identical, ensuring both compatibility and increased expressive power.

6. Implementation Considerations, Scaling, and Impact

Implementation of the 3DGS rendering equation at scale necessitates parallelization over image space and world space: RetinaGS (Li et al., 2024) partitions Gaussians spatially across GPUs and merges partial α-blending results hierarchically to recover the standard rendering outcome. This enables dense scenes with over a billion Gaussians to be rendered and trained efficiently.

The 3DGS rendering equation, in all its forms, generalizes both classical volume rendering (of continuous MLP-based fields) and point-based splatting to a theory that is physically plausible, differentiable, and compatible with modern graphics hardware. The equation underpins advances in real-time radiance field synthesis, efficient inverse rendering, dynamic avatars, remote-sensed shadow analysis, and neural relighting pipelines, supporting a diverse range of applications from neural view synthesis to physically accurate satellite reconstruction (Li et al., 2024, Huang et al., 29 May 2025, Wei et al., 2024, Hou et al., 2024, Luo et al., 4 Jan 2026, Mir et al., 4 Jan 2026, Zhou et al., 7 Aug 2025, Katsumata et al., 2023).

7. Summary Table of Core 3DGS Rendering Variants

Variant	Key Equation / Technique	Use Case / Benefit
Standard 3DGS	$C = \sum c_i \alpha_i \prod_{j < i} (1-\alpha_j)$	Efficient, accurate rendering with front-to-back order
Weighted Sum Rendering	$$C = (c_B w_B + \sum c_i \alpha_i w(d_i))/(w_B + ...)$$	Sort-free, order-independent, efficient compositing (Hou et al., 2024)
Volumetric (3DGEER)	$\alpha_i = \sigma_i \exp(-\frac{1}{2} D_i^2)$	Exact ray-Gaussian integration, enhanced quality in extreme FOVs (Huang et al., 29 May 2025)
Physically-based (Normal-GS)	$$c_i = k_{D,i} (\mathbf{n}_i \cdot \mathbf{l}_i) + ...$$	Accurate lighting/shading, geometry-appearance coupling (Wei et al., 2024)
Shadow-aware (ShadowGS)	$C(u) = F(u) L_{\text{total}}(u)$, elaborate blending	Explicit direct/ambient separation, shadow consistency (Luo et al., 4 Jan 2026)
Dynamic (Compact Dynamic)	Dynamic $\mu(t)$, $R(t)$ in $C = \sum c_i \alpha_i ...$	Real-time dynamic synthesis, memory reduction (Katsumata et al., 2023)
Gabor-augmented (3DGabSplat)	Gabor-mod. $G_k$ in standard blend	Frequency-adaptive, sharper reconstruction (Zhou et al., 7 Aug 2025)

Each formulation inherits the core splatting pipeline while extending expressiveness, accuracy, or computational efficiency for specific application domains.