SGGX Orientation Distributions

• SGGX orientation distributions are anisotropic microfacet normal distribution functions that generalize the classic GGX model using a symmetric shape matrix.
• They enable precise modeling of ellipsoidal anisotropy in microgeometry, vital for capturing directional scattering and supporting hierarchical level-of-detail aggregation.
• By leveraging efficient sampling and clustering techniques, SGGX offers a compact, physically-plausible representation for both surface rendering and volumetric light scattering.

Symmetrical GGX (SGGX) orientation distributions define a class of three-dimensional, anisotropic microfacet normal distribution functions (NDFs) that generalize the isotropic GGX distribution to arbitrary ellipsoidal spreads of directions. SGGX is parameterized by a symmetric positive-definite matrix $S \in \mathbb{R}^{3 \times 3}$ (the "shape matrix"), which determines the orientation and anisotropy of the distribution. Designed to provide a highly compact and physically accurate fit to orientation statistics encountered in complex microgeometry, SGGX is central both as a foundational NDF in microfacet theory and as the basis for hierarchical level-of-detail (LoD) aggregation in rendering of voxellized microgeometry (Fabre et al., 14 Apr 2026).

1. Mathematical Formulation

The SGGX distribution maps a unit-length normal vector $n \in \mathbb{R}^3$ to a probability density on the unit sphere $S^2$, parameterized by the symmetric positive-definite matrix $S$. The most widely employed closed form is: $$f(n; S) = \frac{1}{\pi \sqrt{\det S}\, (n^\top S^{-1} n)^2}$$ where $\det S$ is the determinant of $S$, and $(n^\top S^{-1} n) > 0$ given $S \succ 0$. The normalization by $\pi\sqrt{\det S}$ ensures that

$n \in \mathbb{R}^3$0

A small regularizer $n \in \mathbb{R}^3$1 is sometimes added to avoid singularities for highly anisotropic $n \in \mathbb{R}^3$2. This distribution reduces to the classic isotropic GGX for $n \in \mathbb{R}^3$3, where $n \in \mathbb{R}^3$4 is the identity matrix and $n \in \mathbb{R}^3$5 is the scalar roughness.

The exponent $n \in \mathbb{R}^3$6 in $n \in \mathbb{R}^3$7 imparts the characteristic long tails of GGX, supporting the accurate modeling of grazing-angle scattering. The parameter $n \in \mathbb{R}^3$8 encodes both the dominant orientation and anisotropic spread of microfacet normals, allowing SGGX to fit arbitrary ellipsoidal orientation distributions.

2. From Isotropic GGX to Full Anisotropy

Isotropic GGX is a special case of SGGX, achieved by setting $n \in \mathbb{R}^3$9, yielding: $S^2$0 which can also be expressed in SGGX notation as $S^2$1. For more general anisotropy, the matrix $S^2$2 is diagonalized as $S^2$3 with eigenvalues $S^2$4 and orthonormal $S^2$5. The principal roughnesses are $S^2$6, $S^2$7 (in the tangent plane), and directions aligned with the mean normal (typically axis $S^2$8) have the smallest eigenvalue $S^2$9. This parameterization deforms the GGX lobe into an arbitrary ellipsoid, providing the flexibility required to model microgeometry with strongly directional or aligned features.

3. Hierarchical SGGX Clustering for Level-of-Detail

Level-of-detail hierarchies leveraging SGGX are constructed by agglomerative clustering over sets of per-voxel SGGX distributions $S$0. The hierarchical algorithm merges pairs of SGGX distributions at each step to reduce the set size toward a target count $S$1 (often $S$2), preserving visually salient anisotropic features per voxel:

• For each candidate pair, a histogram $S$3 of $S$4 samples from $S$5 is computed.
• Pairs are selected for merging by minimizing a symmetric histogram distance, such as the 1-Wasserstein distance.
• The merged histogram $S$6 is fitted to a new SGGX distribution $S$7 by matching covariance.
• This process iterates until $S$8.

The resulting cluster per voxel encodes the dominant anisotropic lobes, and acts as a compact approximant of the local NDF ensemble. Errors arise from limited histogram statistics (scaling as $S$9) and from the progressive merging's softening of anisotropy—using $$f(n; S) = \frac{1}{\pi \sqrt{\det S}\, (n^\top S^{-1} n)^2}$$0 clusters maintains fidelity. LoD error at each merge is bounded by the difference (e.g., in $$f(n; S) = \frac{1}{\pi \sqrt{\det S}\, (n^\top S^{-1} n)^2}$$1 metric) between the original and merged NDFs, with cumulative error growing sub-additively under Wasserstein distance.

4. Sampling Procedures

Efficient sampling of microfacet normals from SGGX is based on the anisotropic GGX slope-mapping (Heitz 2015):

1. Decompose $$f(n; S) = \frac{1}{\pi \sqrt{\det S}\, (n^\top S^{-1} n)^2}$$2.
2. Draw two independent uniform random variables $$f(n; S) = \frac{1}{\pi \sqrt{\det S}\, (n^\top S^{-1} n)^2}$$3.
3. Compute slope magnitude $$f(n; S) = \frac{1}{\pi \sqrt{\det S}\, (n^\top S^{-1} n)^2}$$4, and azimuth $$f(n; S) = \frac{1}{\pi \sqrt{\det S}\, (n^\top S^{-1} n)^2}$$5.
4. Form the local microfacet direction $$f(n; S) = \frac{1}{\pi \sqrt{\det S}\, (n^\top S^{-1} n)^2}$$6.
5. Rotate back: $$f(n; S) = \frac{1}{\pi \sqrt{\det S}\, (n^\top S^{-1} n)^2}$$7.

The corresponding density in slope coordinates is

$$f(n; S) = \frac{1}{\pi \sqrt{\det S}\, (n^\top S^{-1} n)^2}$$8

with the Jacobian yielding $$f(n; S) = \frac{1}{\pi \sqrt{\det S}\, (n^\top S^{-1} n)^2}$$9. For cluster mixtures (SGGX LoD sets), mixture sampling operates by randomly selecting a component (weighted, e.g., by $\det S$0) and sampling $\det S$1 from its $\det S$2.

In volumetric media (participating media), sampling is further constrained by visibility, using Visible-Normal Distribution Function (VNDF) inversion combined with anisotropic stretching. This restricts sampled normals to directions visible to the incident direction.

5. Integration in Microgeometry Rendering and Path Tracing

SGGX orientation distributions, including their hierarchical LoD aggregates, are directly integrated into path tracing of voxelized microgeometry and explicit-fiber volumetric fabrics:

• At each shading point, the corresponding SGGX cluster ($\det S$3 at LoD level $\det S$4) is retrieved.
• A microfacet normal $\det S$5 is sampled from the mixture, and the half-vector $\det S$6 is computed.
• The microfacet BSDF is then: $\det S$7 with $\det S$8 and $\det S$9 a shadow-masking term.
• For volumetric scattering, SGGX serves as the orientation PDF for the phase function, with outgoing direction $S$0 sampled conditioned on the local distribution.

Level selection uses a mip-mapping metric based on projected screen-space voxel diameter: $S$1 and selects the appropriate SGGX cluster for the current LoD.

The SGGX-based approach delivers highly compact (as little as 6 bytes per SGGX instance via a compact encoding), anisotropic, and easily sampleable orientation representations, with a tunable LoD hierarchy that preserves directional detail substantially better than naive (e.g., arithmetic) averaging approaches.

6. Implications and Limitations

SGGX clustering for LoD preserves dominant anisotropic lobes within each voxel, but merging introduces softening of anisotropy proportional to the number of merges. Fitting SGGX to histograms with finite samples introduces statistical bias, while the LoD error can be quantified by NDF distances at each merge. Using multiple (e.g., $S$2) clusters per voxel mitigates the loss of detail and maintains accuracy for highly anisotropic fibers and structures. This suggests that in practical applications, cluster size adjustment offers an effective tradeoff between memory footprint and rendering fidelity.

A plausible implication is that hierarchical SGGX LoD aggregation can extend beyond path tracing to other domains requiring directional texture compression or analysis of orientation distributions in 3D datasets, given its compactness and physical conservatism.

7. Context Within Microfacet Theory and Analytical Rendering

SGGX orientation distributions generalize GGX to support arbitrary anisotropy and offer a unified, physically-plausible NDF for use in microfacet theory as applied to both surface and volumetric rendering. This generalization supports the physically-correct modeling of complex microstructure, including but not limited to aligned fibers, brushed metals, and volumetric fabrics.

The representation is seamlessly adapted for LoD hierarchies, supporting efficient path tracing implementations for triangle meshes and volumetric data. The method is validated on CUDA-based implementations and provides a rendering system with per-voxel, cluster-based SGGX LoD for both accurate light scattering and scalable data aggregation in physically-based rendering pipelines (Fabre et al., 14 Apr 2026).