3D Skew Gaussian Splatting: A Detailed Overview

• 3D Skew Gaussian Splatting is a technique that replaces symmetric Gaussian primitives with asymmetric skewed kernels to better capture complex boundaries and discontinuities.
• It employs a skew-normal formulation that modulates opacity and adjusts primitive centroids for improved visualization in real-time rendering pipelines.
• The method optimizes parameters via a decoupled strategy, enhancing structural fidelity and reducing redundancy compared to traditional symmetric approaches.

Searching arXiv for papers on 3D Skew Gaussian Splatting and closely related skew-normal splatting formulations. 3D Skew Gaussian Splatting (3DSGS) is an extension of 3D Gaussian Splatting that replaces symmetric Gaussian primitives with asymmetric skewed kernels in order to improve real-time novel view synthesis and interactive visualization, particularly near object boundaries, thin structures, one-sided surfaces, and shape or color discontinuities. In the recent literature, one formulation adopts the Azzalini Skew-Normal distribution as the fundamental primitive and is termed Skew-Normal Splatting (SNS), while another introduces a general skew Gaussian primitive together with enhanced opacity modeling, depth-aware densification, and a decoupled free-camera visualization engine (Wu et al., 14 May 2026, Zhao et al., 18 May 2026).

1. Motivation and relation to symmetric 3DGS

3D Gaussian Splatting (3DGS) is described as a leading representation for real-time novel view synthesis because Gaussian primitives provide favorable mathematical and computational properties. The limitation identified by skew-based extensions is that, under a finite primitive budget, the symmetric shape of each primitive directly affects representation compactness, especially near asymmetric structures such as object boundaries and one-sided surfaces. The visualization-oriented formulation states the same issue in rendering terms: symmetric kernels struggle to capture shape and color discontinuities, which cause blurriness and primitive redundancy that mislead human perception during visual analysis (Wu et al., 14 May 2026, Zhao et al., 18 May 2026).

The central intervention is therefore not merely to enlarge the kernel family, but to introduce intrinsic asymmetry while preserving rasterization efficiency. One paper emphasizes continuous interpolation between symmetric Gaussians and Half-Gaussian-like shapes; the other emphasizes structural fidelity, compactness, and compatibility with real-time rendering. A common misconception is that any asymmetric primitive can be treated as a hard truncation of a Gaussian. The skew-normal formulation explicitly distinguishes itself from approaches that rely on hard truncation, arguing that such approaches limit continuous shape control and introduce distributional discontinuities (Wu et al., 14 May 2026).

2. Skew primitive and distributional structure

In the skew-normal formulation, a random variable $X\in\mathbb R^d$ follows Azzalini’s Skew-Normal distribution with location $\mu$, scale matrix $\Omega$, and slant vector $\alpha$, written $X\sim\mathcal{SN}_d(\mu,\Omega,\alpha)$, with unnormalized density

$$f(x;\mu,\Omega,\alpha) = \exp\Bigl(-\tfrac12(x-\mu)^\top\Omega^{-1}(x-\mu)\Bigr) \;\Phi\!\bigl(\alpha^\top\,\Omega^{-1/2}(x-\mu)\bigr),$$

where $\Phi$ is the CDF of a standard univariate normal, $\Omega\in\mathbb R^{d\times d}$ is positive-definite, $\mu\in\mathbb R^d$ is the location parameter, and $\alpha\in\mathbb R^d$ is the slant vector. If $\mu$0, the density reduces, up to a constant factor, to the $\mu$1-variate Gaussian. As $\mu$2, the distribution approaches a “half-Gaussian” truncated on one side (Wu et al., 14 May 2026).

The paper also stresses that the location parameter is not the centroid. The mean is shifted by a skew-dependent term,

$\mu$3

so the primitive’s effective center differs from $\mu$4. This matters directly for rasterization, culling, and tile assignment. The same formulation measures overall skewness using Mardia’s multivariate skewness and reports that it saturates at $\mu$5, which ensures each primitive remains spatially compact (Wu et al., 14 May 2026).

A closely related formulation writes the 3DSGS primitive as a skew Gaussian obtained by modulating a standard anisotropic Gaussian with a one-dimensional CDF along a skewness direction: $\mu$6 Here the factor of $\mu$7 ensures normalization to a true PDF when $\mu$8, and the construction recovers the symmetric Gaussian when $\mu$9. This suggests that the naming difference between “skew-normal” and “skew Gaussian” reflects two presentations of the same basic idea: a Gaussian core modulated by a univariate normal CDF to induce directed asymmetry (Zhao et al., 18 May 2026).

3. Projection, rasterization, and compositing

A decisive technical property of the skew-normal primitive is closure under affine transforms and marginalization. Let $\Omega$0 denote the local affine approximation of the world-to-screen map. If $\Omega$1, then the projected variable remains skew-normal in two dimensions,

$\Omega$2

with $\Omega$3 and $\Omega$4. The paper also gives a reparameterized 2D kernel

$\Omega$5

where $\Omega$6 is a closed-form function of $\Omega$7. This analytical projection is what allows seamless integration into existing Gaussian Splatting rasterization pipelines (Wu et al., 14 May 2026).

Compositing follows front-to-back alpha blending. For each pixel $\Omega$8, the renderer collects the front-to-back sorted set of primitives intersecting that pixel and evaluates

$\Omega$9

Because the skew-normal location parameter is not the centroid, the tile-bounding-box center is shifted to the true 2D mean rather than kept at $\alpha$0. This is a nontrivial correction: without it, screen-space support estimation would be biased for strongly skewed splats (Wu et al., 14 May 2026).

The generalized 3DSGS CUDA pipeline describes the same rendering logic at the systems level. The image is partitioned into tiles; each tile stores a list of overlapping primitives; each primitive is affinely projected; the $\alpha$1 projected covariance is computed; and the forward-splat kernel evaluates a Gaussian fall-off $\alpha$2, a skew term via $\alpha$3, and then $\alpha$4. The resulting opacity is $\alpha$5, after which color and transmittance are accumulated exactly as in standard front-to-back splatting. The backward pass remains analytical, with closed-form gradients for covariance terms, opacity, and the skew vector (Zhao et al., 18 May 2026).

4. Parameterization and optimization

One source of instability in skewed splatting is the coupling between scale, rotation, and skewness. The skew-normal paper addresses this through a decoupled parameterization. As in 3DGS, the covariance is factorized as

$\alpha$6

with $\alpha$7 and diagonal $\alpha$8. Instead of optimizing $\alpha$9 directly, the method introduces a latent intrinsic skew vector $X\sim\mathcal{SN}_d(\mu,\Omega,\alpha)$0 and sets

$X\sim\mathcal{SN}_d(\mu,\Omega,\alpha)$1

Under this construction, $X\sim\mathcal{SN}_d(\mu,\Omega,\alpha)$2 alone controls skewness in the primitive’s canonical frame, while $X\sim\mathcal{SN}_d(\mu,\Omega,\alpha)$3 and $X\sim\mathcal{SN}_d(\mu,\Omega,\alpha)$4 only orient and scale that intrinsic shape (Wu et al., 14 May 2026).

The same work further decomposes $X\sim\mathcal{SN}_d(\mu,\Omega,\alpha)$5 into magnitude and direction: $X\sim\mathcal{SN}_d(\mu,\Omega,\alpha)$6 where $X\sim\mathcal{SN}_d(\mu,\Omega,\alpha)$7 and $X\sim\mathcal{SN}_d(\mu,\Omega,\alpha)$8 are unconstrained. This cleanly separates “how much skew” from “which direction.” The primitive parameter set is

$X\sim\mathcal{SN}_d(\mu,\Omega,\alpha)$9

The reported training schedule is two-phase. During warm-up, all shape and skew parameters are updated jointly via Adam. After $$f(x;\mu,\Omega,\alpha) = \exp\Bigl(-\tfrac12(x-\mu)^\top\Omega^{-1}(x-\mu)\Bigr) \;\Phi\!\bigl(\alpha^\top\,\Omega^{-1/2}(x-\mu)\bigr),$$0, the optimizer switches to cyclic Block-Coordinate Descent on $$f(x;\mu,\Omega,\alpha) = \exp\Bigl(-\tfrac12(x-\mu)^\top\Omega^{-1}(x-\mu)\Bigr) \;\Phi\!\bigl(\alpha^\top\,\Omega^{-1/2}(x-\mu)\bigr),$$1 versus $$f(x;\mu,\Omega,\alpha) = \exp\Bigl(-\tfrac12(x-\mu)^\top\Omega^{-1}(x-\mu)\Bigr) \;\Phi\!\bigl(\alpha^\top\,\Omega^{-1/2}(x-\mu)\bigr),$$2: for the first $$f(x;\mu,\Omega,\alpha) = \exp\Bigl(-\tfrac12(x-\mu)^\top\Omega^{-1}(x-\mu)\Bigr) \;\Phi\!\bigl(\alpha^\top\,\Omega^{-1/2}(x-\mu)\bigr),$$3 steps of each cycle, $$f(x;\mu,\Omega,\alpha) = \exp\Bigl(-\tfrac12(x-\mu)^\top\Omega^{-1}(x-\mu)\Bigr) \;\Phi\!\bigl(\alpha^\top\,\Omega^{-1/2}(x-\mu)\bigr),$$4 are frozen while $$f(x;\mu,\Omega,\alpha) = \exp\Bigl(-\tfrac12(x-\mu)^\top\Omega^{-1}(x-\mu)\Bigr) \;\Phi\!\bigl(\alpha^\top\,\Omega^{-1/2}(x-\mu)\bigr),$$5 are updated; for the remaining $$f(x;\mu,\Omega,\alpha) = \exp\Bigl(-\tfrac12(x-\mu)^\top\Omega^{-1}(x-\mu)\Bigr) \;\Phi\!\bigl(\alpha^\top\,\Omega^{-1/2}(x-\mu)\bigr),$$6 steps, $$f(x;\mu,\Omega,\alpha) = \exp\Bigl(-\tfrac12(x-\mu)^\top\Omega^{-1}(x-\mu)\Bigr) \;\Phi\!\bigl(\alpha^\top\,\Omega^{-1/2}(x-\mu)\bigr),$$7 are frozen while $$f(x;\mu,\Omega,\alpha) = \exp\Bigl(-\tfrac12(x-\mu)^\top\Omega^{-1}(x-\mu)\Bigr) \;\Phi\!\bigl(\alpha^\top\,\Omega^{-1/2}(x-\mu)\bigr),$$8 are updated. Positional parameters $$f(x;\mu,\Omega,\alpha) = \exp\Bigl(-\tfrac12(x-\mu)^\top\Omega^{-1}(x-\mu)\Bigr) \;\Phi\!\bigl(\alpha^\top\,\Omega^{-1/2}(x-\mu)\bigr),$$9 are updated with SGHMC to promote global exploration, all other parameters use Adam, and the only explicit image-space loss is the per-pixel $\Phi$0 color loss

$\Phi$1

No additional regularizers are needed beyond bounded skewness and parameter decoupling (Wu et al., 14 May 2026).

The visualization-oriented formulation reports complementary optimization details rather than a different objective. Its listed hyperparameters include learning rate $\Phi$2 for mean, covariance, and color, learning rate $\Phi$3 for skewness $\Phi$4, densification thresholds $\Phi$5 and $\Phi$6, and an opacity-bisect regularizer $\Phi$7 (Zhao et al., 18 May 2026).

5. Opacity modeling, densification, and visualization engine

Beyond asymmetric spatial support, 3DSGS can alter the opacity model itself. In the 3DSGS visualization framework, screen-space opacity for primitive $\Phi$8 is

$\Phi$9

so skewness modulates opacity directly after projection. The same work introduces a bisected-region opacity model in which a single primitive carries two base opacities, $\Omega\in\mathbb R^{d\times d}$0 and $\Omega\in\mathbb R^{d\times d}$1, separated by a directional partition $\Omega\in\mathbb R^{d\times d}$2 in screen space. The stated purpose is to represent a sharp in-kernel discontinuity analytically, rather than by overlapping multiple Gaussians (Zhao et al., 18 May 2026).

The same paper augments standard 3DGS densification with a depth-aware criterion. In addition to the 2D screen-space position gradient norm $\Omega\in\mathbb R^{d\times d}$3, it tracks a depth-gradient

$\Omega\in\mathbb R^{d\times d}$4

If either $\Omega\in\mathbb R^{d\times d}$5 or $\Omega\in\mathbb R^{d\times d}$6, the Gaussian is split along its principal scale axis: parameters $\Omega\in\mathbb R^{d\times d}$7 are cloned, the scale is reduced by $\Omega\in\mathbb R^{d\times d}$8 on that axis, skew and base opacities are adjusted by heuristic, two new Gaussians are inserted, and the original is removed. The reported interpretation is that large $\Omega\in\mathbb R^{d\times d}$9 indicates under-resolved depth discontinuities such as thin occluders and floaters (Zhao et al., 18 May 2026).

At the implementation level, the skew-normal method is a PyTorch extension of the 3DGS rasterizer in which the new 2D skew-normal kernel is inserted wherever 3DGS originally used a Gaussian. Each skew-normal evaluation requires one call to $\mu\in\mathbb R^d$0, described as an erf-based overhead that adds a small constant slowdown while remaining real-time on tile-based GPU splatting. Training on an NVIDIA RTX A6000 for 30,000 iterations on the “drjohnson” scene takes approximately 38 minutes, reported as on par with Student-t splatting (SSS), and the asymptotic per-frame complexity remains $\mu\in\mathbb R^d$1 (Wu et al., 14 May 2026).

The 3DSGS visualization engine re-derives the CUDA rasterization flow so that both symmetric and skew Gaussians are natively supported with minimal branching. The implementation stores primitives in Structure-of-Arrays for coalesced memory loads, performs bounding-box checks and covariance untangling once per tile-primitive pair and stores them in shared memory, approximates $\mu\in\mathbb R^d$2 via a fast polynomial, places intrinsics $\mu\in\mathbb R^d$3 and the coordinate-alignment matrix $\mu\in\mathbb R^d$4 in constant memory, and assigns one CUDA block to each tile. During free-camera exploration in Blender via VisEngine, the reported runtime is $\mu\in\mathbb R^d$5–$\mu\in\mathbb R^d$6 FPS depending on resolution, with render-time remaining above $\mu\in\mathbb R^d$7 FPS even under $\mu\in\mathbb R^d$8 M primitives (Zhao et al., 18 May 2026).

6. Empirical behavior and reported gains

The skew-normal paper evaluates on Mip-NeRF360 with $\mu\in\mathbb R^d$9 scenes, Tanks & Temples with $\alpha\in\mathbb R^d$0 scenes, and Deep Blending with $\alpha\in\mathbb R^d$1 scenes, using PSNR$\alpha\in\mathbb R^d$2, SSIM$\alpha\in\mathbb R^d$3, and LPIPS$\alpha\in\mathbb R^d$4. Its reported averages are: 3DGS $\alpha\in\mathbb R^d$5, GES $\alpha\in\mathbb R^d$6, 3D-HGS $\alpha\in\mathbb R^d$7, SSS $\alpha\in\mathbb R^d$8, and SNS $\alpha\in\mathbb R^d$9. On all three benchmarks, 3DSGS achieves the highest PSNR, with average $\mu$00 dB over SSS, and ties or exceeds SSIM and LPIPS. The same report states that on every single scene, SNS outperforms 3DGS, GES, 3D-HGS, and SSS by at least $\mu$01 dB; one cited example is “Room,” which improves from $\mu$02 dB with 3D-HGS to $\mu$03 dB with SNS. Its ablation study starts from a “vanilla” skew-normal primitive and reports $\mu$04 dB on Mip-NeRF360 from $\mu$05-decomposition alone, $\mu$06 dB from alternating optimization alone, and $\mu$07 dB from combining both, yielding the full $\mu$08 dB result. Qualitatively, the paper highlights sharper castle towers, curtain folds, thin poles, and box lips (Wu et al., 14 May 2026).

The visualization-engine paper reports a different set of absolute numbers but a similar trend. On the Mip-NeRF-360 outdoor dataset with $\mu$09 scenes, vanilla 3DGS is reported at PSNR $\mu$10, SSIM $\mu$11, LPIPS $\mu$12, kernels $\mu$13 M, and $\mu$14 FPS, while 3DSGS reaches PSNR $\mu$15, SSIM $\mu$16, LPIPS $\mu$17, kernels $\mu$18 M, and $\mu$19 FPS. On Tanks & Temples with $\mu$20 scenes, PSNR rises from $\mu$21 to $\mu$22 and SSIM from $\mu$23 to $\mu$24, while kernel count decreases from $\mu$25 to $\mu$26 M at $\mu$27 FPS. On Deep Blending with $\mu$28 scenes, PSNR increases from $\mu$29 to $\mu$30 and LPIPS decreases from $\mu$31 to $\mu$32. The qualitative description emphasizes crisper edges and fewer “ghost-glows” in railings, wires, window frames, and foliage, with lower absolute pixel-error heatmaps along those features (Zhao et al., 18 May 2026).

Taken together, these results indicate a consistent empirical pattern: asymmetric primitives improve reconstruction quality and structural compactness most clearly at sharp boundaries, thin geometry, one-sided surfaces, and regions with complex transparency. A plausible implication is that the main advantage of 3DSGS is not uniformly higher fidelity everywhere in the scene, but more efficient allocation of representational capacity where symmetry is intrinsically mismatched to the underlying geometry or appearance.