One-sentence Summary

University College London and University of Leeds propose Student Splatting and Scooping (SSS), a mixture model for novel view synthesis that replaces Gaussian Splatting with flexible Student's t distributions featuring both positive (splatting) and negative (scooping) densities, paired with a principled sampling approach for optimization, achieving comparable rendering quality with up to $82\%$82% fewer components.

Key Contributions

• Student Splatting and Scooping (SSS) replaces the Gaussian splatting formulation of 3DGS with a mixture of flexible 3D Student's t distributions that can represent both positive splatting and negative scooping densities for novel view synthesis.
• SSS introduces a principled SGHMC-based sampling and optimization method for learning this positive and negative Student's t mixture model, enabling more expressive scene representations than standard 3DGS formulations.
• Across multiple datasets, settings, and metrics, SSS delivers higher rendering quality and parameter efficiency, matching or exceeding existing methods with similar component counts while maintaining comparable quality with up to 82% fewer components.

Introduction

The authors leverage 3D Gaussian Splatting as a fast neural rendering foundation used in geometry reconstruction and autonomous driving, where real-time image synthesis and parameter efficiency matter. Prior methods retain Gaussians and positive-density splatting, giving practical speed but limited expressivity and often requiring many components to fit complex radiance fields, while related NeRF and 3DGS variants also face tradeoffs between rendering efficiency and primitive expressivity. Their main contribution is Student Splatting and Scooping, which replaces Gaussians with Student's t distributions, extends splatting to positive and negative densities that add and subtract representation capacity, and uses SGHMC sampling to learn the resulting model with higher rendering quality and fewer components.

Dataset

• The authors build their experimental dataset from 11 scenes across three established 3D scene benchmarks: Mip-NeRF 360, Tanks & Temples, and Deep Blending.

• The dataset subset structure is:

  • Mip-NeRF 360: 7 public scenes.
  • Tanks & Temples: 2 outdoor scenes.
  • Deep Blending: 2 indoor scenes.

• The source selection follows prior research in the field. The paper does not describe additional filtering rules, scene splitting, or metadata construction beyond using these standard benchmark scenes.

• The authors use the dataset to evaluate 3D Gaussian splatting and related novel view synthesis methods. They compare performance against prior baselines and report common image quality metrics.

• The evaluation metrics are PSNR, SSIM, and LPIPS. The authors provide average scores for each dataset and refer to the supplementary material for detailed per-scene results.

• No cropping strategy, schema, or preprocessing pipeline is specified in the provided text beyond the use of standard public scenes and standard evaluation metrics.

Method

The proposed method builds on the mixture model interpretation of 3D Gaussian splatting and replaces Gaussian components with more flexible Student t components, while also introducing signed components and a principled sampling-based optimizer. The authors view the scene representation as an unnormalized 3D distribution assembled from many local basis functions. A basic component is defined by a location, a shape, an opacity, and a color. The overall radiance field is written as a weighted mixture, and each ray through an image pixel integrates the components intersected by that ray.

For standard 3D Gaussian splatting, the radiance field is an unnormalized Gaussian mixture:

Each Gaussian is associated with a center $\mu$μ, a covariance matrix $\Sigma = R S S^T R^T$Σ=RSSTRT, opacity, and color. During rendering, 3D Gaussians are projected onto the 2D image plane and composited along rays:

This formulation is additive and monotonic because component weights are positive. It is numerically convenient because Gaussians are closed under affine transformations and marginalization, which makes projection and gradient computation straightforward.

The authors then replace each Gaussian basis function with an unnormalized Student t basis function. A Student t component has a mean $\mu \in \mathbb{R}^3$μ∈R3, a covariance matrix $\Sigma = R S S^T R^T \in \mathbb{R}^{3 \times 3}$Σ=RSSTRT∈R3×3, and a degree of freedom parameter $\nu \in [1,+\infty)$ν∈[1,+∞). The resulting mixture is

with component density

The normalizing scalar of the classical t distribution is omitted because the mixture is unnormalized and the opacity, color, and compositing weights can absorb scale during learning.

The key design choice is that $\nu$ν controls tail fatness. As $\nu \to \infty$ν→∞, the Student t component approaches a Gaussian, while as $\nu \to 1$ν→1, it approaches a Cauchy distribution. Because Cauchy components are heavier tailed than Gaussians, they can cover broader regions with higher density, potentially reducing the number of components needed to explain a scene. Since $\nu$ν, $\mu$μ, and $\Sigma$Σ are all learnable, the model can adapt not only the location and scale of each component, but also the underlying shape family. Refer to the figure below:

This behavior is important for representational capacity. The model can express localized Gaussian-like features, fat-tailed regions, and intermediate shapes by adjusting $\nu$ν. It therefore provides a continuum between compact Gaussian modeling and broader heavy-tailed coverage without requiring a fixed mixture of hand-designed basis types.

A second reason for using Student t components is that they retain computational tractability under projection. Rendering requires mapping a 3D component into 2D by an affine transformation, an approximate projective transformation, and integration along a ray. The projected Student t density also has a closed form:

where

and

Here $W$W and $t$t encode affine scale and translation, while $J$J is the approximated Jacobian of the projective transformation. This closed form allows the model to compute stable forward passes and exact enough gradients without relying on approximations for the basis component itself.

The rendering formula follows the same alpha compositing structure as Gaussian splatting, but substitutes the projected Student t component:

The authors then extend the model beyond purely positive mixtures. A fully nonmonotonic mixture can be formed by squaring the sum of components:

Although this introduces negative contributions through products of weights, it also creates pairwise interactions between all components, increasing evaluation complexity to $O(n^2)$O(n2) and making optimization substantially slower. Instead, the authors keep the additive mixture form but allow signed contributions. In this setting, a negative density is interpreted as subtracting color from the rendered result. The signed weight of a component along a ray can be understood through its color, opacity, and accumulated transparency:

Because signed opacity can make the mixture ill-defined as a probability distribution, the learning objective is instead parameterized in an energy-based form. The opacity is also bounded using a $\tanh$tanh mapping, allowing components to shift between positive and negative contributions while remaining numerically controlled. The role of negative components is illustrated below:

In the torus example, a positive component can cover the overall shape but may also fill the central hole. A co-located negative component can remove density from the middle, making the topology easier to express with fewer primitives. This subtractive mechanism increases parameter efficiency and helps the model capture holes, contours, and thin structures without relying on dense positive coverage alone.

The training procedure is formulated as sampling rather than only vanilla stochastic gradient descent. The authors observe that the Student t model introduces tighter coupling between parameters because changing $\nu$ν effectively changes the distribution family in which $\mu$μ and $\Sigma$Σ are optimized. To handle this coupling, they use a Stochastic Gradient Hamiltonian Monte Carlo optimizer over the key geometric parameters. Since the signed mixture is not a normalized posterior, they define an energy-based posterior:

where $\theta$θ contains the learnable parameters, $L_\theta(x)$Lθ​(x) is the rendering loss, $I$I is an identity matrix, and $r$r is a momentum-like auxiliary variable. The energy function defines high-density regions corresponding to good parameter values without requiring a normalized mixture density.

The continuous sampling dynamics are

where $U(\theta)=L_\theta(x)-\frac{1}{2}r^T I r$U(θ)=Lθ​(x)−21​rTIr, $M$M is a mass matrix, and $C$C controls friction and injected noise. The momentum term creates dimension-wise inertia, enabling adaptive movement through the parameter space rather than relying on a single uniform step size.

The loss combines image reconstruction terms and regularizers that encourage compactness:

The first two terms drive photometric and structural alignment with training images. The opacity regularizer penalizes unnecessary components, while the eigenvalue regularizer encourages compact covariance support, effectively pushing components to become localized and spiky unless they are needed to represent the scene.

In practice, the sampling optimizer is applied primarily to the mean parameters $\mu$μ, while other parameters are updated using standard gradient-based optimization. The authors use an adaptive version of the SGHMC update for $\mu$μ:

where

and

The term $\sigma(o)$σ(o) acts as a gate that activates friction and noise primarily for low-opacity components. The authors initialize the model from a sparse set of points, run a burn-in phase without friction to encourage exploration, and then use the full sampled dynamics for exploitation. During burn-in, the noise is scaled by the component covariance to preserve the anisotropic shape of the Student t component. After burn-in, the momentum term maintains the anisotropy profile while the adaptive friction and noise help low-opacity components move more freely.

Component maintenance is also handled through recycling rather than simple deletion. During optimization, some components may become nearly transparent. The authors relocate these low-opacity components toward higher-opacity components while preserving the rendered color field. The relocation problem is formulated as

A key constraint enforcing color preservation is

where $N$N is the total number of components after relocation. The covariance is adjusted in a degree-of-freedom-aware way, ensuring that the local density and rendered color remain close to their pre-relocation values. The authors limit relocation to a small fraction of components at each step to maintain stability, and they also introduce new zero-opacity components that can later be recycled into useful regions. The final trainable variables for each component include the mean $\mu$μ, the scaling and rotation parameters defining $\Sigma$Σ, the color $c$c, the opacity $o$o, and the degree-of-freedom parameter $\nu$ν.

Experiment

The experiments evaluate the proposed SSS on multiple scenes from three datasets using standard image reconstruction metrics against representative 3DGS variants and NeRF baselines. The results show that SSS generally delivers the best overall visual and quantitative performance while preserving high fidelity in challenging scenes. Additional experiments with reduced numbers of components and qualitative comparisons demonstrate that SSS can capture both homogeneous regions and fine details more effectively with fewer components than competing methods, highlighting its improved representation ability and parameter efficiency.

Across the benchmark datasets, SSS is the strongest overall method, placing first on most image quality metrics and second on the rest except one close perceptual quality case. It also uses fewer components than baselines while preserving quality, and it degrades less as component counts are reduced. Visual comparisons indicate that it handles both large uniform regions and detailed structures more consistently under constrained budgets. SSS wins most quality metrics among compared methods Its largest advantage over the closest baseline appears on the Train scene The remaining gap is mainly a small perceptual quality loss in Deep Blending With fewer components, SSS remains more stable than methods relying on more Gaussians

The experiments evaluate SSS on benchmark datasets against Gaussian-based baselines, focusing on image quality, perceptual quality, and robustness as component budgets are reduced. SSS is consistently the strongest overall, winning most quality metrics and showing its largest advantage on the Train scene, while remaining more stable than baseline methods even with fewer components despite a small perceptual quality gap on Deep Blending. Visual comparisons further indicate that it preserves both large uniform regions and detailed structures more reliably under constrained budgets.