3D Gabor Splatting: Frequency-Adaptive Rendering

• The paper presents 3DGabSplat, a novel radiance-field rendering approach that enhances traditional Gaussian splatting with multiple directional 3D frequency responses.
• It employs a Gabor-based filter bank that combines a low-frequency Gaussian response with localized oscillatory modulation to better capture high-frequency details and mitigate aliasing.
• Empirical evaluations demonstrate improved PSNR and efficiency, with reduced primitive counts compared to conventional 3DGS, confirming its practical benefits in detail representation.

3D Gabor Splatting denotes a line of explicit splat-based radiance-field representations that replace or augment the low-pass Gaussian primitive of 3D Gaussian Splatting (3DGS) with Gabor-style localized oscillatory structure. In the recent literature, the designation is used most directly for "3DGabSplat: 3D Gabor Splatting for Frequency-adaptive Radiance Field Rendering" (Zhou et al., 7 Aug 2025), which introduces a 3D Gabor-based primitive with multiple directional 3D frequency responses for radiance field rendering. A related but distinct use appears in "3D Gabor Splatting: Reconstruction of High-frequency Surface Texture using Gabor Noise" (Watanabe et al., 15 Apr 2025), which augments 2DGS-style planar splats with oscillatory local color fields for high-frequency diffuse texture reconstruction. Both formulations arise from a common diagnosis established in the 3DGS literature: anisotropic Gaussian splats are efficient, explicit, and rasterizable, but their low-pass character makes high-frequency detail, anti-aliasing, and representation efficiency persistent technical issues (Wu et al., 2024).

1. Context within Gaussian splatting

3DGS represents a scene as a set of anisotropic 3D Gaussian ellipsoids rather than as a viewpoint-conditioned neural field. In the standard formulation, each primitive carries a center, covariance, opacity, and color, and rendering proceeds by projecting the primitive to screen space and compositing projected splats front to back. The surveys on 3DGS emphasize two central properties of this representation: efficient rasterization and explicit editability. They also highlight the canonical parameterization

$\Sigma = R S S^T R^T,$

its projected form

$\Sigma' = J W \Sigma W^T J^T,$

and opacity-weighted front-to-back accumulation as the structural scaffold of the method family (Wu et al., 2024).

The same surveys also identify the limitations that motivate Gabor-based variants. Standard Gaussian splats are repeatedly characterized as smooth, low-pass carriers that can blur high-frequency texture, require substantial densification, and exhibit aliasing or scale sensitivity under changing sampling rate, focal length, or camera distance. Anti-aliasing and frequency-aware variants such as Mip-Splatting and Analytic-Splatting were introduced precisely because projected Gaussian evaluation is not inherently alias-free across scales (Yu et al., 2023, Liang et al., 2024). The 2024 surveys do not mention 3DGabSplat explicitly, but they locate its natural design space in primitive redesign, splatting extension, frequency-aware supervision, and anti-aliasing (Bao et al., 2024).

Within that context, the Gabor construction is motivated by a standard signal-processing contrast. A Gaussian envelope provides spatial locality, whereas sinusoidal modulation provides frequency selectivity. This makes a Gabor primitive localized yet band-pass, rather than purely low-pass. In 3DGabSplat, that combination is used to retain the projectable, rasterizable structure of Gaussian splatting while increasing the ability of each primitive to encode localized directional detail (Zhou et al., 7 Aug 2025).

2. Core formulations in the literature

The term has been used for two closely related formulations in splat-based 3D reconstruction.

Formulation	Base geometry	Gabor mechanism
3DGabSplat (Zhou et al., 7 Aug 2025)	3DGS-style anisotropic 3D primitive	One Gaussian low-frequency component plus multiple 3D Gabor band-pass components
3D Gabor Splatting (Watanabe et al., 15 Apr 2025)	2DGS-style planar splat in 3D	Gaussian alpha profile with multi-wave oscillatory color in local $(u,v)$ coordinates

In the radiance-field formulation of 3DGabSplat, the baseline Gaussian primitive is

$$\mathcal{G}_k(\mathbf{x}) = \exp{\left(-\frac{1}{2}(\mathbf{x} - \boldsymbol{\mu}_k)^{\top} \mathbf{\Sigma}_k^{-1} (\mathbf{x} - \boldsymbol{\mu}_k)\right)},$$

with

$$\mathbf{\Sigma}_k = \mathbf{R}_k \mathbf{S}_k \mathbf{S}_k^\top\mathbf{R}_k^\top.$$

The paper then defines a 3D Gabor-based primitive as a per-primitive filter bank,

$$\{Gabor\}_k(\mathbf{x})=\mathcal{G}_k(\mathbf{x})\cdot \Big[\big(1-\sum_{i=1}^F \omega_{k,i}\big) + \sum_{i=1}^F \omega_{k,i}\cos\big(2\pi \boldsymbol{f}_{k,i}^\top(\mathbf{x}-\boldsymbol{\mu}_k)\big)\Big].$$

Here, $\boldsymbol{f}_{k,i}\in\mathbb{R}^3$ are learnable 3D frequency vectors and $\omega_{k,i}$ are learnable weights. The residual term $\big(1-\sum_{i=1}^F \omega_{k,i}\big)$ is the weight on the pure Gaussian component. The authors state that Gaussian splatting is recovered as a special case when $\boldsymbol{f}_{k,i}=0$ and $\Sigma' = J W \Sigma W^T J^T,$0, and that unidirectional Gabor modulation is likewise a degenerate case (Zhou et al., 7 Aug 2025).

This construction changes the role of a primitive. A standard 3DGS primitive is a single low-pass kernel. A 3DGabSplat primitive is a weighted sum of one low-frequency Gaussian response and multiple localized directional cosine-modulated responses. The default implementation uses $\Sigma' = J W \Sigma W^T J^T,$1, and each Gabor kernel adds three frequency components plus one weight, for $\Sigma' = J W \Sigma W^T J^T,$2 additional parameters per primitive. The frequencies are not globally shared; each primitive is assigned distinct frequencies for its own region. The paper also notes that there is no explicit learnable phase parameter in this formulation, since phase is handled implicitly through $\Sigma' = J W \Sigma W^T J^T,$3 (Zhou et al., 7 Aug 2025).

By contrast, the planar high-frequency texture formulation is built on 2DGS. A splat is parameterized on a local plane as

$\Sigma' = J W \Sigma W^T J^T,$4

with Gaussian profile

$\Sigma' = J W \Sigma W^T J^T,$5

Instead of modifying opacity or geometry with a 3D frequency bank, it defines a local oscillatory color field using multiple fixed-orientation waves. The phase of wave $\Sigma' = J W \Sigma W^T J^T,$6 is

$\Sigma' = J W \Sigma W^T J^T,$7

and the color is

$\Sigma' = J W \Sigma W^T J^T,$8

This earlier formulation explicitly learns phase shifts and frequencies, but fixes the wave orientations and focuses on diffuse surface texture rather than full view-dependent radiance-field rendering (Watanabe et al., 15 Apr 2025).

3. Projection, compositing, and optimization

The radiance-field form of 3DGabSplat preserves the standard 3DGS projection-and-compositing pipeline while replacing the primitive kernel. Covariance projection remains

$\Sigma' = J W \Sigma W^T J^T,$9

and alpha compositing in baseline 3DGS is written as

$(u,v)$0

3DGabSplat extends this by projecting not only covariance but also frequency content. The projected frequency vector is

$(u,v)$1

and after integrating along the $(u,v)$2-axis the 2D frequency components are corrected by entries of $(u,v)$3, rather than being obtained by a naive $(u,v)$4 slice (Zhou et al., 7 Aug 2025).

The final screen-space primitive is

$(u,v)$5

and rendered color becomes

$(u,v)$6

Thus the overall renderer remains a rasterization-based, depth-ordered, front-to-back compositing system; the principal modification is the evaluation of a projected Gabor-based primitive rather than a pure projected Gaussian (Zhou et al., 7 Aug 2025).

Training preserves the standard photometric objective

$(u,v)$7

The method initializes from an SfM point cloud, inherits densification and pruning from 3DGS, and jointly optimizes position, covariance, opacity, color, Gabor frequencies, and Gabor weights. Its distinctive optimization mechanism is frequency-adaptive rather than regularization-based in analytic form. Frequencies are initialized to $(u,v)$8, weights to $(u,v)$9, child primitives created by densification reset both quantities to small values, and all Gabor weights are reset every $$\mathcal{G}_k(\mathbf{x}) = \exp{\left(-\frac{1}{2}(\mathbf{x} - \boldsymbol{\mu}_k)^{\top} \mathbf{\Sigma}_k^{-1} (\mathbf{x} - \boldsymbol{\mu}_k)\right)},$$0 epochs together with opacity reset. The paper argues that these heuristics suppress the accumulation of overly high-frequency child primitives during densification (Zhou et al., 7 Aug 2025).

Implementation remains close to the 3DGS software stack. The method uses a differentiable CUDA-based rasterizer adapted from the tile-based 3DGS renderer, constrains opacity and Gabor weights with sigmoid to $$\mathcal{G}_k(\mathbf{x}) = \exp{\left(-\frac{1}{2}(\mathbf{x} - \boldsymbol{\mu}_k)^{\top} \mathbf{\Sigma}_k^{-1} (\mathbf{x} - \boldsymbol{\mu}_k)\right)},$$1, and is presented as a plug-and-play kernel that can be integrated into Scaffold-GS, Mip-Splatting, and AbsGS by replacing the Gaussian kernel while keeping the larger framework structure (Zhou et al., 7 Aug 2025).

4. High-frequency surface texture reconstruction variant

The earlier high-frequency texture formulation addresses a narrower problem: reconstructing objects whose geometry is comparatively simple but whose diffuse surface texture contains fine stripes, repetitive textile patterns, or similar local oscillations. Its central claim is that a Gaussian or planar Gaussian splat effectively carries one color per primitive, so reconstructing stripe-like patterns requires many separate kernels, often with very high aspect ratio. The paper therefore augments the internal color field of each planar splat with Gabor-inspired multi-wave structure, while leaving the 2DGS geometry and alpha profile essentially intact (Watanabe et al., 15 Apr 2025).

Its representation explicitly optimizes

$$\mathcal{G}_k(\mathbf{x}) = \exp{\left(-\frac{1}{2}(\mathbf{x} - \boldsymbol{\mu}_k)^{\top} \mathbf{\Sigma}_k^{-1} (\mathbf{x} - \boldsymbol{\mu}_k)\right)},$$2

together with

$$\mathcal{G}_k(\mathbf{x}) = \exp{\left(-\frac{1}{2}(\mathbf{x} - \boldsymbol{\mu}_k)^{\top} \mathbf{\Sigma}_k^{-1} (\mathbf{x} - \boldsymbol{\mu}_k)\right)},$$3

In the reported configuration, $$\mathcal{G}_k(\mathbf{x}) = \exp{\left(-\frac{1}{2}(\mathbf{x} - \boldsymbol{\mu}_k)^{\top} \mathbf{\Sigma}_k^{-1} (\mathbf{x} - \boldsymbol{\mu}_k)\right)},$$4, corresponding to orientations sampled every $$\mathcal{G}_k(\mathbf{x}) = \exp{\left(-\frac{1}{2}(\mathbf{x} - \boldsymbol{\mu}_k)^{\top} \mathbf{\Sigma}_k^{-1} (\mathbf{x} - \boldsymbol{\mu}_k)\right)},$$5. The wave orientations are fixed, while weights, phases, and frequencies are learned. Relative to a homogeneous-color splat, this introduces an additional color and $$\mathcal{G}_k(\mathbf{x}) = \exp{\left(-\frac{1}{2}(\mathbf{x} - \boldsymbol{\mu}_k)^{\top} \mathbf{\Sigma}_k^{-1} (\mathbf{x} - \boldsymbol{\mu}_k)\right)},$$6 scalar wave parameters per primitive. For $$\mathcal{G}_k(\mathbf{x}) = \exp{\left(-\frac{1}{2}(\mathbf{x} - \boldsymbol{\mu}_k)^{\top} \mathbf{\Sigma}_k^{-1} (\mathbf{x} - \boldsymbol{\mu}_k)\right)},$$7, the paper identifies 18 appearance scalars associated with the Gabor augmentation, excluding opacity and geometry. Unlike the later radiance-field 3DGabSplat, this method explicitly ignores view-dependent color change; the paper states that its main target is detailed texture of the diffuse surface, such as garments (Watanabe et al., 15 Apr 2025).

The optimization setup follows 2DGS closely. Positions are initialized from COLMAP/SfM points, remaining parameters are randomly initialized, the method uses the same loss function and the same number of training iterations ($$\mathcal{G}_k(\mathbf{x}) = \exp{\left(-\frac{1}{2}(\mathbf{x} - \boldsymbol{\mu}_k)^{\top} \mathbf{\Sigma}_k^{-1} (\mathbf{x} - \boldsymbol{\mu}_k)\right)},$$8) as 2DGS, and the reported loss spike at the 7001st iteration comes from the introduction of normal consistency loss inherited from 2DGS. For fair comparison, densification is turned off so that the number of resulting primitives matches 2DGS. The rendering change is therefore localized to primitive shading: ray-plane intersection yields local $$\mathcal{G}_k(\mathbf{x}) = \exp{\left(-\frac{1}{2}(\mathbf{x} - \boldsymbol{\mu}_k)^{\top} \mathbf{\Sigma}_k^{-1} (\mathbf{x} - \boldsymbol{\mu}_k)\right)},$$9, alpha is evaluated from the Gaussian envelope, color is evaluated from the multi-wave Gabor model, and the final image is obtained by alpha blending similar to 2DGS (Watanabe et al., 15 Apr 2025).

This formulation also establishes a distinct practical lesson about Gabor splatting. A single learnable wave with free orientation and frequency was found difficult to optimize; the paper therefore fixes orientations to a small set of uniformly sampled directions and learns only frequencies, phases, and mixture weights. That design makes the model less orientation-adaptive than a fully free Gabor basis, but the paper presents it as a deliberate stability tradeoff (Watanabe et al., 15 Apr 2025).

5. Empirical behavior, efficiency, and ablations

The radiance-field 3DGabSplat evaluates on Mip-NeRF360, Tanks and Temples, Deep Blending, and NeRF Synthetic using PSNR, SSIM, and LPIPS, with additional reporting of training time, FPS, memory usage, and primitive count. Relative to 3DGS, the reported PSNR gains are $$\mathbf{\Sigma}_k = \mathbf{R}_k \mathbf{S}_k \mathbf{S}_k^\top\mathbf{R}_k^\top.$$0 dB on Mip-NeRF360, $$\mathbf{\Sigma}_k = \mathbf{R}_k \mathbf{S}_k \mathbf{S}_k^\top\mathbf{R}_k^\top.$$1 dB on Tanks and Temples, $$\mathbf{\Sigma}_k = \mathbf{R}_k \mathbf{S}_k \mathbf{S}_k^\top\mathbf{R}_k^\top.$$2 dB on Deep Blending, and $$\mathbf{\Sigma}_k = \mathbf{R}_k \mathbf{S}_k \mathbf{S}_k^\top\mathbf{R}_k^\top.$$3 dB on NeRF Synthetic. The corresponding table entries are 27.85 versus 27.41 on Mip-NeRF360, 24.49 versus 23.69 on Tanks and Temples, 30.09 versus 29.55 on Deep Blending, and 33.722 versus 33.347 on NeRF Synthetic. The paper also reports that 3DGabSplat+Mip-Splatting+AbsGS achieves SSIM $$\mathbf{\Sigma}_k = \mathbf{R}_k \mathbf{S}_k \mathbf{S}_k^\top\mathbf{R}_k^\top.$$4, PSNR $$\mathbf{\Sigma}_k = \mathbf{R}_k \mathbf{S}_k \mathbf{S}_k^\top\mathbf{R}_k^\top.$$5, and LPIPS $$\mathbf{\Sigma}_k = \mathbf{R}_k \mathbf{S}_k \mathbf{S}_k^\top\mathbf{R}_k^\top.$$6 on Mip-NeRF360, and identifies this as a new state-of-the-art in its comparison table (Zhou et al., 7 Aug 2025).

Efficiency claims are tied to primitive expressiveness rather than cheaper per-primitive computation. In the ablation table, baseline 3DGS uses $$\mathbf{\Sigma}_k = \mathbf{R}_k \mathbf{S}_k \mathbf{S}_k^\top\mathbf{R}_k^\top.$$7M primitives and $$\mathbf{\Sigma}_k = \mathbf{R}_k \mathbf{S}_k \mathbf{S}_k^\top\mathbf{R}_k^\top.$$8 MB, whereas 3DGabSplat with $$\mathbf{\Sigma}_k = \mathbf{R}_k \mathbf{S}_k \mathbf{S}_k^\top\mathbf{R}_k^\top.$$9 uses $$\{Gabor\}_k(\mathbf{x})=\mathcal{G}_k(\mathbf{x})\cdot \Big[\big(1-\sum_{i=1}^F \omega_{k,i}\big) + \sum_{i=1}^F \omega_{k,i}\cos\big(2\pi \boldsymbol{f}_{k,i}^\top(\mathbf{x}-\boldsymbol{\mu}_k)\big)\Big].$$0M primitives and $$\{Gabor\}_k(\mathbf{x})=\mathcal{G}_k(\mathbf{x})\cdot \Big[\big(1-\sum_{i=1}^F \omega_{k,i}\big) + \sum_{i=1}^F \omega_{k,i}\cos\big(2\pi \boldsymbol{f}_{k,i}^\top(\mathbf{x}-\boldsymbol{\mu}_k)\big)\Big].$$1 MB. FPS is reported as $$\{Gabor\}_k(\mathbf{x})=\mathcal{G}_k(\mathbf{x})\cdot \Big[\big(1-\sum_{i=1}^F \omega_{k,i}\big) + \sum_{i=1}^F \omega_{k,i}\cos\big(2\pi \boldsymbol{f}_{k,i}^\top(\mathbf{x}-\boldsymbol{\mu}_k)\big)\Big].$$2 for 3DGS and $$\{Gabor\}_k(\mathbf{x})=\mathcal{G}_k(\mathbf{x})\cdot \Big[\big(1-\sum_{i=1}^F \omega_{k,i}\big) + \sum_{i=1}^F \omega_{k,i}\cos\big(2\pi \boldsymbol{f}_{k,i}^\top(\mathbf{x}-\boldsymbol{\mu}_k)\big)\Big].$$3 for 3DGabSplat with $$\{Gabor\}_k(\mathbf{x})=\mathcal{G}_k(\mathbf{x})\cdot \Big[\big(1-\sum_{i=1}^F \omega_{k,i}\big) + \sum_{i=1}^F \omega_{k,i}\cos\big(2\pi \boldsymbol{f}_{k,i}^\top(\mathbf{x}-\boldsymbol{\mu}_k)\big)\Big].$$4. The paper attributes this to the reduction in primitive count outweighing the added cost of evaluating cosine-modulated kernels. Ablation over the number of Gabor kernels shows best PSNR at $$\{Gabor\}_k(\mathbf{x})=\mathcal{G}_k(\mathbf{x})\cdot \Big[\big(1-\sum_{i=1}^F \omega_{k,i}\big) + \sum_{i=1}^F \omega_{k,i}\cos\big(2\pi \boldsymbol{f}_{k,i}^\top(\mathbf{x}-\boldsymbol{\mu}_k)\big)\Big].$$5: $$\{Gabor\}_k(\mathbf{x})=\mathcal{G}_k(\mathbf{x})\cdot \Big[\big(1-\sum_{i=1}^F \omega_{k,i}\big) + \sum_{i=1}^F \omega_{k,i}\cos\big(2\pi \boldsymbol{f}_{k,i}^\top(\mathbf{x}-\boldsymbol{\mu}_k)\big)\Big].$$6 for $$\{Gabor\}_k(\mathbf{x})=\mathcal{G}_k(\mathbf{x})\cdot \Big[\big(1-\sum_{i=1}^F \omega_{k,i}\big) + \sum_{i=1}^F \omega_{k,i}\cos\big(2\pi \boldsymbol{f}_{k,i}^\top(\mathbf{x}-\boldsymbol{\mu}_k)\big)\Big].$$7, $$\{Gabor\}_k(\mathbf{x})=\mathcal{G}_k(\mathbf{x})\cdot \Big[\big(1-\sum_{i=1}^F \omega_{k,i}\big) + \sum_{i=1}^F \omega_{k,i}\cos\big(2\pi \boldsymbol{f}_{k,i}^\top(\mathbf{x}-\boldsymbol{\mu}_k)\big)\Big].$$8 for $$\{Gabor\}_k(\mathbf{x})=\mathcal{G}_k(\mathbf{x})\cdot \Big[\big(1-\sum_{i=1}^F \omega_{k,i}\big) + \sum_{i=1}^F \omega_{k,i}\cos\big(2\pi \boldsymbol{f}_{k,i}^\top(\mathbf{x}-\boldsymbol{\mu}_k)\big)\Big].$$9, $\boldsymbol{f}_{k,i}\in\mathbb{R}^3$0 for $\boldsymbol{f}_{k,i}\in\mathbb{R}^3$1, $\boldsymbol{f}_{k,i}\in\mathbb{R}^3$2 for $\boldsymbol{f}_{k,i}\in\mathbb{R}^3$3, and $\boldsymbol{f}_{k,i}\in\mathbb{R}^3$4 for $\boldsymbol{f}_{k,i}\in\mathbb{R}^3$5, while training time and memory both increase as $\boldsymbol{f}_{k,i}\in\mathbb{R}^3$6 grows. Comparison to fixed unidirectional Gabor modulation gives PSNR $\boldsymbol{f}_{k,i}\in\mathbb{R}^3$7 for 2DGabSplat and $\boldsymbol{f}_{k,i}\in\mathbb{R}^3$8 for full 3DGabSplat, isolating the contribution of learnable 3D directional frequency vectors. The component ablation further reports $\boldsymbol{f}_{k,i}\in\mathbb{R}^3$9 without periodic frequency reset, $\omega_{k,i}$0 without densification frequency reset, and $\omega_{k,i}$1 for the full model. When all trained frequencies and weights are set to zero, PSNR drops to $\omega_{k,i}$2 from $\omega_{k,i}$3, which the paper describes as about 10% PSNR degradation (Zhou et al., 7 Aug 2025).

The surface-texture formulation reports a different profile because it compares against 2DGS on object-centric textured datasets. On Sweat, Boots, and Shirt, reported metrics are 0.872/25.69/0.232, 0.867/23.59/0.305, and 0.849/24.12/0.262 for 3D Gabor Splatting, versus 0.852/25.15/0.276, 0.853/23.23/0.329, and 0.815/22.13/0.311 for 2DGS in SSIM/PSNR/LPIPS. Training time is roughly 43 minutes versus roughly 30 minutes for 2DGS on a GeForce RTX 3090, and rendering is reported at 65–95 FPS versus 130–176 FPS. Ablations show that a single wave produces blurrier results, multiple waves in only one fixed direction lead to false colors, and fixing phase to zero slightly degrades quantitative scores. The paper also reports that an approximately half-primitive-count variant can still outperform 2DGS at similar data size, supporting the claim that a more suitable local texture basis can improve representation efficiency (Watanabe et al., 15 Apr 2025).

6. Relation to adjacent methods, misconceptions, and open directions

In the broader 3DGS ecosystem, Gabor splatting occupies the intersection of primitive redesign, splatting extension, frequency-aware modeling, and anti-aliasing. The surveys identify closely neighboring lines of work: Mip-Splatting, Multi-scale 3DGS, Analytic-Splatting, and SA-GS for anti-aliasing; FreGS and TRIPS for high-frequency recovery and blur reduction; GES for non-Gaussian primitive redesign; Spec-Gaussian for more complex anisotropic view-dependent appearance; and 2DGS for surface-aligned primitives (Wu et al., 2024, Bao et al., 2024). This situates 3DGabSplat less as an isolated anomaly than as a concrete answer to a recognized pressure point in 3DGS: the need to enhance the representational capacity of individual primitives without abandoning explicit rasterizable structure.

A common misconception is to treat all such extensions as variants of view-dependent shading. That is inaccurate. VoD-3DGS, for example, retains Gaussian primitives and adds a quadratic directional modulation to opacity,

$\omega_{k,i}$4

to suppress or boost Gaussians depending on viewpoint. Its target is specular highlights and reflections, not localized spatial-frequency encoding. Gabor splatting, by contrast, modifies the local basis carried by each primitive, either through directional 3D frequency responses in radiance-field rendering or through oscillatory local color fields on planar splats (Nowak et al., 29 Jan 2025).

Open directions differ between the two main formulations. The radiance-field 3DGabSplat paper explicitly notes per-primitive computational overhead, sensitivity to the number of Gabor kernels and reset heuristics, added implementation complexity in the rasterizer, and a validation scope limited to static radiance-field rendering rather than dynamic scenes, very large-scale scenes, special failure textures, or temporal stability (Zhou et al., 7 Aug 2025). The planar texture paper explicitly notes the absence of view-dependent color, the possibility that filtering may become necessary even though aliasing was not observed in the reported experiments, and the need to extend the method to more complicated scenes (Watanabe et al., 15 Apr 2025). The anti-aliasing literature suggests a further implication: once splats carry explicit frequency content, footprint-aware rendering becomes even more central, since point-sampled projected kernels are vulnerable to scale-dependent artifacts (Yu et al., 2023, Liang et al., 2024).

Taken together, these works define 3D Gabor Splatting as a frequency-aware extension of explicit splat rendering. Its central technical idea is consistent across formulations: retain the local, differentiable, rasterizable support of Gaussian splats, but replace purely smooth per-primitive behavior with a localized oscillatory basis so that fine structure can be represented more directly than by repeated densification alone (Zhou et al., 7 Aug 2025, Watanabe et al., 15 Apr 2025).