Mini-Splatting: Efficient 3D Scene Synthesis

• The paper introduces targeted densification, geometric simplification, and pruning methods to optimize Gaussian splatting, reducing redundancy and speeding up convergence.
• Mini-Splatting models 3D scenes as collections of anisotropic Gaussian splats, using depth reinitialization and importance sampling to enhance rendering fidelity and efficiency.
• It employs aggressive pruning and visibility culling strategies that reduce memory usage from 7 GB to under 3 GB and boost rendering speed from 100 FPS to over 400 FPS.

The Mini-Splatting framework is an approach to 3D scene representation and novel-view synthesis that applies targeted densification, geometrically guided simplification, and aggressive pruning to achieve high-fidelity rendering with a small, computationally efficient set of Gaussian primitives. Emerging from core limitations in traditional 3D Gaussian Splatting (3DGS)—notably its high redundancy and slow optimization—Mini-Splatting reconfigures the process to optimize convergence speed, memory efficiency, and rendering quality. The framework has since been extended by techniques such as aggressive densification, visibility culling, and integration with structured and multimodal scene representations.

1. Mathematical Foundations of Mini-Splatting

Mini-Splatting models a 3D scene as a collection of $N$ anisotropic Gaussian "splats." Each splat is defined by its center $\mathbf{p}_i\in\mathbb{R}^3$, covariance $\Sigma_i\in\mathbb{R}^{3\times3}$, opacity $\alpha_i\in [0,1]$, and view-dependent color $c_i$ parameterized through low-order spherical harmonics. Given a camera projection, each 3D Gaussian is projected into screen space as a 2D elliptical Gaussian $G_i^{2D}(u)$.

Per-pixel blending weights are computed as:

$$w_i(u) = T_i(u) \alpha_i G_i^{2D}(u),\qquad T_i(u) = \prod_{j<i}(1 - w_j(u))$$

Pixels are rendered by compositing:

$c(u) = \sum_{i=1}^N w_i(u) c_i(u)$

This rasterization formulation closely follows differentiable alpha-compositing pipelines used in real-time graphics and 3DGS, but Mini-Splatting imposes hard constraints and adaptive distribution on the number and spatial arrangement of Gaussians (Fang et al., 2024).

2. Densification and Reinitialization Strategies

Early approaches to 3DGS employed progressive densification, incrementally splitting Gaussians with high residual error every few hundred optimization iterations, culminating in millions of Gaussians and extended training time. Mini-Splatting introduces targeted densification using two main techniques:

• Blur Split: Gaussians that dominate large, smooth regions (measured by $S_i$ = number of pixels where $i$ is top contributor) are split along their principal axis to avoid over-blurring and improve local detail.
• Depth Reinitialization: Periodically, a rendered depth map is used to sample new Gaussian centers in view space, ensuring uniform coverage and mitigating clustering artifacts. The midpoint of ray-Gaussian intersection is computed for each pixel, providing robust geometric reinitialization.

Subsequent extensions (Mini-Splatting2) introduced aggressive Gaussian densification: rather than pacing out splits, the framework identifies "critical" Gaussians using a max-weight criterion and injects a large proportion of final geometry-supporting Gaussians (often ∼30%) within the first 3K iterations. This is done via a "smoothed" clone rule for opacity and covariance to ensure training stability, and empirical results demonstrate convergence 3–5× faster (e.g., reaching target PSNR and Chamfer metrics in ∼3K rather than ∼15K iterations) (Fang et al., 2024).

3. Simplification, Pruning, and Visibility Culling

Mini-Splatting incorporates multiple mechanisms for model size constraint and computational efficiency:

• Intersection-Preserving Pruning: Only Gaussians providing ray intersections (i.e., actually contributing to at least one pixel) are retained, eliminating floaters and noise.
• Importance-Weighted Sampling: Each Gaussian is assigned an importance score (e.g., total blending weight), and a target budget of $M\ll N$ Gaussians is retained via probabilistic sampling, followed by detail-preserving fine-tuning.
• Visibility Gaussian Culling: In Mini-Splatting2, per-view importance $I_i^k$ is accumulated for each Gaussian and camera view; visibility masks $V_i^k$ are computed using a high quantile threshold (e.g., $\tau=0.99$). Gaussians that are not important for any batch-selected view are pruned from rasterization, reducing the per-frame Gaussian set by up to 40% beyond frustum culling.

These strategies maintain rendering fidelity while achieving major reductions in memory (e.g., from 7 GB to <3 GB) and improving rendering speed (∼100 FPS to >400 FPS) (Fang et al., 2024).

4. Mini-Splatting Optimization Pipeline

The pipeline is organized as follows:

1. Initialization: Gaussians are seeded from Structure-from-Motion (SfM) or Multi-View Stereo (MVS) point clouds, with typical counts around 10K–100K.
2. Densification: For a fixed window (e.g., iter 500–3,000), densification occurs via blur splitting or aggressive max-weight cloning.
3. Depth Reinitialization: At a mid-point (e.g., iter 2,000), depth maps are sampled and new Gaussians are injected at depth-derived positions.
4. Visibility Culling and Pruning: Starting from early iterations (500+) to a mid-to-late cutoff (e.g., 13K), visibility masks are used to cull unimportant Gaussians per view; intersection-preserving and importance pruning trim the set to the final Gaussian budget.
5. Optimization: Photometric and geometric losses are minimized with per-iteration rasterization and backpropagation.
6. Periodic Simplification: At strategic iterations (e.g., 3K, 8K), clustering and importance-based pruning further rebalance the Gaussian count (e.g., targeting 0.5–0.6M).
7. Termination: The compact, final set of Gaussians—often an order of magnitude smaller than vanilla 3DGS—yields the optimized, high-fidelity representation.

The principal hyperparameters include the quantile threshold for culling (e.g., $\tau=0.99$), clone factor (default 2), densification window, and simplification milestones (Fang et al., 2024).

5. Performance, Quantitative Results, and Benchmarks

Mini-Splatting and its extensions have been benchmarked on large-scale 360° datasets and standard view-synthesis testbeds:

Method	SSIM	PSNR	LPIPS	#Gaussians (M)	Train Time
3DGS-accel	0.811	27.38	0.225	2.39	10 min 02 s
Mini-Splatting	0.822	27.34	0.217	0.49	11 min 33 s
Mini-Splatting2	0.821	27.33	0.215	0.62	3 min 34 s

Mini-Splatting2 achieves convergence in minutes rather than tens of minutes on high-resolution, full 360° scenes, with substantial reduction in Gaussian count and memory, and only negligible differences (<0.01 PSNR) compared to more resource-intensive methods (Fang et al., 2024).

6. Extensions and Integration with Related Frameworks

Several lines of research have built upon or incorporated the principles of Mini-Splatting:

• Structured Representations: CUS-GS incorporates a voxelized anchor structure with per-voxel multimodal semantic features, achieving competitive performance and strong multimodal alignment in as few as 6M parameters (Ming et al., 22 Nov 2025).
• Modular/Minimal Frameworks: GauStudio provides a decomposable infrastructure for Gaussian Splatting pipelines, delineating minimal Mini-Splatting configurations (core: data loading, Gaussian init, splat optimization) as well as plug-in surface reconstruction or environmental background modules (Ye et al., 2024).
• Fine-Detail Regularization: Micro-splatting introduces compactness constraints and gradient-driven adaptive densification, recovering more high-frequency detail via trace penalties and localized Gaussian splitting (Lee et al., 8 Apr 2025).
• Textured Gaussians: Textured-GS extends Mini-Splatting to allow spatially varying color and opacity within each Gaussian via spherical harmonics texture fields. Compared at constant Gaussian count, this improves visual fidelity (e.g., for Mip-NeRF360, SSIM 0.825 vs. 0.822 for Mini-Splatting), at the cost of increased per-frame compute (Huang et al., 2024).
• Real-Time/VR Rendering: VRSplat uses Mini-Splatting's compact Gaussian set, augments with VR-specific rasterization enhancements (hierarchical sorting, tangent-plane projection, foveated rendering), and achieves >72 FPS with elimination of temporal and projection artifacts (Tu et al., 15 May 2025).

7. Limitations, Open Problems, and Outlook

Major limitations of Mini-Splatting include:

• Depth-based reinitialization is brittle in textureless or "infinite" backgrounds (e.g., sky). This motivates integration of multiview or semantic priors.
• Current schedules for splitting, pruning, and simplification are hand-tuned; adaptive or learned policies remain an open research area.
• While the hard budget enables memory and speed gains, fixed allocation may miss fine-scale structures, suggesting possible benefit from hybrid or multiscale splatting (Fang et al., 2024).

The Mini-Splatting framework has effectively established new baselines for rapid, resource-efficient optimization of 3D Gaussian scene representations, and its algorithmic innovations continue to inform structured, semantic, and real-time rendering research across novel-view synthesis and high-fidelity 3D modeling (Fang et al., 2024, Tu et al., 15 May 2025).