Universal Beta Splatting¶
- Conference: ICLR 2026
- arXiv: 2510.03312
- Code: Project Page
- Area: 3D Vision / Neural Rendering
- Keywords: 3D Gaussian Splatting, Beta Kernel, N-Dimensional, View-Dependent, Dynamic Scene, Real-Time Rendering
TL;DR¶
This paper proposes Universal Beta Splatting (UBS), which generalizes 3D Gaussian Splatting to an N-dimensional anisotropic Beta kernel. By enabling per-dimension shape control, UBS unifies spatial geometry, view-dependent appearance, and scene dynamics within a single representation, achieving interpretable scene decomposition and state-of-the-art rendering quality.
Background & Motivation¶
3D Gaussian Splatting (3DGS) achieves real-time rendering via explicit primitives, yet the fixed bell-shaped profile of the Gaussian kernel imposes fundamental limitations:
Spatial dimensions: Sharp boundaries require large numbers of small primitives, resulting in low efficiency.
Angular dimensions: View-dependent effects rely on additional spherical harmonic encoding (48 parameters), leading to a fragmented representation.
Temporal dimensions: Dynamic scenes require auxiliary deformation networks, increasing overall complexity.
Core Insight: Different scene properties demand different kernel behaviors—spatial geometry requires adaptive sharpness, angular appearance ranges from diffuse to specular, and temporal dynamics span from static to fast-moving elements. The Gaussian kernel enforces the same symmetric profile across all dimensions, whereas a Beta kernel can provide per-dimension shape control.
Method¶
N-Dimensional Beta Kernel¶
The core density function is defined as:
where \(\mathbf{x} \in \mathbb{R}^3\) denotes spatial coordinates, \(\mathbf{q} \in \mathbb{R}^{N-3}\) encodes additional dimensions (view direction / time), and \(\mathbf{b} \in \mathbb{R}^{N-2}\) controls the Beta shape parameter for each dimension. The Beta exponent for each dimension is \(\beta_i = 4\exp(b_i)\): - Negative \(b_i\): flat profile (suitable for smooth surfaces, static elements, diffuse reflection) - Positive \(b_i\): sharp peak (suitable for fine-grained textures, fast motion, specular reflection)
Spatial-Orthogonal Cholesky Parameterization¶
The covariance matrix is decomposed as:
- \(\mathbf{R}_x \in SO(3)\): preserves spatial orthogonal structure (first-order Taylor approximation)
- \(\mathbf{L}_{qx}\): encodes cross-dimensional correlations
- Guarantees backward compatibility with the rotation-scale parameterization of 3DGS
Beta-Modulated Conditional Slice¶
Conditional mean and covariance:
where \(\text{diag}(\tilde{\boldsymbol{\beta}}_q)\) applies Beta modulation to non-spatial dimensions.
Beta-modulated opacity:
Here \(d_i = \tanh(d_i^{raw}) \in [0,1)\) maps the per-dimension Mahalanobis distance to a bounded value.
Universal Compatibility¶
| \(\mathbf{b}\) Setting | Equivalent Method |
|---|---|
| \(N=3\), \(b_x=0\) | ≈ 3DGS |
| \(N=3\), \(b_x \neq 0\) | ≈ DBS |
| \(N=6\), \(\mathbf{b}=\mathbf{0}\) | ≈ 6DGS |
| \(N=7\), \(\mathbf{b}=\mathbf{0}\) | ≈ 7DGS |
Interpretable Scene Decomposition¶
The learned Beta parameters naturally enable unsupervised scene decomposition: - Spatial \(b_x\): negative → smooth surfaces; positive → fine-grained textures - Angular \(b_d\): negative → diffuse; positive → specular - Temporal \(b_t\): negative → static elements; positive → dynamic elements
Loss & Training¶
Opacity regularization ensures effective MCMC densification, while the scale penalty encourages primitive relocation.
Parameter Efficiency¶
- Static scenes: 41% fewer parameters than 3DGS (no 48-parameter spherical harmonics required)
- Dynamic scenes: 73% fewer parameters than 4DGS
Key Experimental Results¶
Static Scenes¶
NeRF Synthetic (UBS-6D vs. 3DGS vs. 6DGS):
| Scene | 3DGS PSNR | 6DGS PSNR | UBS-6D PSNR |
|---|---|---|---|
| chair | 35.60 | 35.55 | 36.72 |
| ficus | 35.49 | 34.62 | 36.90 |
| materials | 30.50 | 30.63 | 32.90 |
| lego | 36.06 | 35.22 | 36.95 |
PSNR improvements reach up to +8.27 dB on the 6DGS-PBR dataset.
Dynamic Scenes¶
D-NeRF and 7DGS-PBR (UBS-7D vs. 4DGS vs. 7DGS): - PSNR improvements reach up to +2.78 dB - Particularly pronounced advantages on complex spatio-temporal-angular correlation scenes (cardiac motion, daylight variation, translucent deformation)
Key Findings¶
- Initializing Beta parameters to zero guarantees convergence starting from the Gaussian limit.
- The first-order approximation of spatial-orthogonal Cholesky achieves accuracy comparable to exact rotation, with lower computational cost.
- The MCMC optimization strategy remains effective for Beta kernels.
- Real-time rendering performance is on par with 3DGS.
Training Efficiency¶
- 30K iterations
- Static: single RTX 4090, approximately 8–10 minutes
- Dynamic: single V100, consistent with 4DGS/7DGS baselines
Highlights & Insights¶
- Unified framework: A single Beta kernel primitive jointly handles spatial, angular, and temporal dimensions, replacing multiple specialized encodings.
- Backward compatibility: Setting Beta=0 degrades to the Gaussian case, providing a guaranteed performance lower bound.
- Unsupervised decomposition: Learned Beta parameters naturally disentangle geometry, appearance, and motion.
- Substantial parameter reduction: 41–73% fewer parameters with simultaneous performance gains.
- Real-time rendering: Full CUDA-accelerated implementation.
Limitations & Future Work¶
- The number of parameters in the N-dimensional Cholesky parameterization grows with dimensionality.
- The bounded support of the Beta kernel may require more primitives in far-field regions.
- Validation is currently limited to 7 dimensions; effectiveness at higher dimensionalities remains to be investigated.
- Dynamic scene training requires a batch size of 4, resulting in relatively high GPU memory consumption.
- The first-order rotation approximation may be insufficiently accurate under extreme rotation angles.
Related Work & Insights¶
- Alternative kernel designs: GES, TNT-GS, Half-GS, and Disc-GS modify the 3D Gaussian; DBS introduces a 3D Beta kernel.
- High-dimensional Gaussians: 6DGS (spatial + view) and 7DGS (spatial + view + time) model conditional distributions over extended dimensions.
- Dynamic scenes: D-NeRF employs deformation fields; 4DGS directly extends the temporal dimension.
- Alternative primitives: 3D Convex Splatting, Triangle Splatting, and others.
Rating¶
- Novelty: ⭐⭐⭐⭐⭐ — The N-dimensional Beta kernel unified framework constitutes an elegant theoretical contribution.
- Practicality: ⭐⭐⭐⭐⭐ — Plug-and-play compatibility combined with real-time rendering.
- Clarity: ⭐⭐⭐⭐ — Mathematical derivations are clear, though notation is dense.
- Significance: ⭐⭐⭐⭐⭐ — Establishes a universal primitive framework for radiance field rendering.