FastGS: Accelerated Gaussian Summation

Updated 2 June 2026

FastGS is a family of methods that accelerate Gaussian kernel summation for both scientific computing and visual rendering tasks.
Key techniques include plane-wave expansion, hierarchical decomposition, and data-driven densification/pruning for enhanced efficiency.
FastGS frameworks achieve significant speedups (up to 10–15×) and reduced VRAM usage, enabling real-time 3D Gaussian splatting and novel view synthesis.

FastGS encompasses a family of methods and frameworks for accelerating Gaussian kernel manipulations in large-scale computational tasks. The term “FastGS” designates two closely related but distinct lines of research: (1) the fast Gauss transform (FGT) and its generalizations for scalable kernel summation in scientific computing and machine learning, and (2) a suite of recent methods for real-time, resource-efficient 2D and 3D Gaussian Splatting in novel view synthesis, surface reconstruction, and related computer vision tasks. While the original FGT stems from kernel summation for N-body and heat/diffusion problems, the contemporary FastGS frameworks focus on optimized training, adaptive densification/pruning, memory footprint, and multi-view consistency in visual computing. This article presents a unified treatment, organizing FastGS research along the following dimensions: algorithmic principles and kernel expansion, density and pruning strategies, memory/computational optimizations, empirical performance, and extensions to specialized domains.

1. Core Algorithmic Principles of FastGS

The foundational FastGS algorithms adapt the fast Gauss transform from computational physics into a general kernel summation and rendering engine. These methods accelerate evaluation of sums/integrals of the form

$u_i = \sum_{j=1}^N e^{-\|x_i-y_j\|^2/h^2}\,q_j,\quad u(x) = \int_B e^{-\|x-y\|^2/h^2}\,\sigma(y)\,dy,$

where the kernel is the isotropic or anisotropic Gaussian $G(x,y;\delta) = \exp(-\|x-y\|^2/\delta)$ .

Contemporary FastGS works generalize this to:

Large-scale sums in high-dimensional feature spaces (e.g., in kernel density estimation and spectral learning)
High-throughput 2D and 3D Gaussian rasterization and α-blending for visual computing

Key techniques include:

Plane-wave expansion: Expanding the Gaussian kernel using Fourier representations to enable locality and efficient kernel summation.
Hierarchical decomposition: Adaptive quad/octree partitioning with far-field and near-field interactions, leveraging multipole/local expansions or fast Fourier transforms (FFT/NUFFT) depending on source/target distribution (Greengard et al., 2023, Wang et al., 2017).
Densification/pruning via importance: In graphical applications, optimizing the number and configuration of Gaussian primitives by measuring per-primitive contribution across multiple contexts (views, neighborhoods), replacing heuristic or uniform adaptive control with data-driven criteria (Ren et al., 6 Nov 2025, Hahlbohm et al., 10 Feb 2026).

2. FastGS for Visual Computing: Accelerated 3D Gaussian Splatting

A dominant application of FastGS is acceleration of 3D Gaussian Splatting (3DGS) for novel-view rendering, scene reconstruction, and surface modeling. The baseline problem involves millions of anisotropic 3D Gaussians projected and composited over many images. The naive approach results in high computational and memory cost due to uncontrolled growth of splats and inefficient gradient computation.

FastGS-type frameworks introduce several innovations:

Multi-view-consistent densification and pruning: Key metrics assign scores to Gaussians based on their average contribution to photometric error or image-space loss over multiple views. Densification (split/clone) and pruning are both executed using this score, ensuring only geometrically and photometrically relevant Gaussians remain active (Ren et al., 6 Nov 2025).
Efficient parameter updates: Fused per-Gaussian backward passes allow gradient accumulation to be performed with a drastically reduced number of atomics, leveraging shared memory (Hahlbohm et al., 10 Feb 2026).
Memory and data locality optimizations: Z-ordering/Morton key sorting, kernel fusion for activations, and parameter packing minimize VRAM, kernel launches, and cache misses.
Tight axis-aligned bounding box computation and culling: Radically reduce update, culling, and sorting overhead by computing tight bounds for each Gaussian's influence region (Hahlbohm et al., 10 Feb 2026).
Parallel, distributed implementations: For city-scale scenes, index-based sharding and synchronized AllToAll communication enable efficient multi-GPU scaling (BlitzGS) (Wang et al., 13 May 2026).

Pseudocode Outline for FastGS Training Loop

for iter in range(N):
    # Sample K views and acquire corresponding images
    # Forward render: fused rasterizer with AABB, tile culling, per-tile sort
    # Compute per-pixel errors and construct multi-view error masks
    if iter % densification_period == 0:
        # Compute importance/pruning scores using multi-view error maps
        densify/prune Gaussians as dictated by scores
    # Compute loss (L1+SSIM) and backpropagate with fused gradient/optimizer update
    # Periodically reorder buffers by Morton code

3. Adaptive Densification, Pruning, and Split Operators

FastGS frameworks distinguish themselves by adaptive density control (ADC) mechanisms:

Multi-view error-based metrics: For each Gaussian, per-view error maps are accumulated and binarized to form "hard-to-reconstruct" pixel sets. Gaussians are split only if their average score $s_i^+$ exceeds a threshold and pruned if their redundancy score $s_i^-$ exceeds another threshold (Ren et al., 6 Nov 2025).
Error-driven adaptive splitting: AdpSplit utilizes connected-component analysis of high-error pixel regions to determine not only whether but how many times to split a Gaussian. Initialization is performed using region centroids and principal component statistics, and redundant splits across views are merged (Lee et al., 7 May 2026).
Plug-and-play budget-free strategies: Unlike classical controller-based regulation (e.g., minimum backward budget), FastGS often forgoes such mechanisms, relying on data-driven consistency and error statistics for managing splat count.

A typical FastGS split/prune workflow is summarized below:

Stage	Trigger	Metric	Output
Densification	ADC update	Multi-view error score $s_i^+$	Split/clone if above τ₊
Pruning	ADC update	Multi-view redundancy score $s_i^-$	Prune if above τ₋

This approach ensures that the number of active splats is proportional to the actual reconstruction need, rather than preset schedules or heuristic budgets.

4. Computational and Memory Efficiency

Acceleration is achieved not only by reducing the active set size but also through deep computational optimizations targeting GPU utilization:

Fused per-Gaussian backward/Adam: Gradient and parameter updates are merged into a single kernel, minimizing global memory contention and maximizing arithmetic intensity.
Two-stage sort and warp-level bucketization: Allows sorting-by-depth and then by tile in two passes, halving the VRAM and sort time for large sets of splats.
Opacity-independent truncation and tight AABB: Ensures that only relevant pixels/tiles are processed, achieving a substantial reduction in rendering and gradient computation workload (Hahlbohm et al., 10 Feb 2026).
Performance: On standard benchmarks (Mip-NeRF 360, Deep Blending, Tanks & Temples), FastGS achieves up to 4–5× training speedup and ∼30% lower VRAM usage, with negligible PSNR/SSIM/LPIPS degradation (Hahlbohm et al., 10 Feb 2026, Ren et al., 6 Nov 2025, Lee et al., 7 May 2026).

Method	Speedup vs. Vanilla 3DGS	Memory Saving	Quality Change
FastGS	10×–15×	~30%	≤0.1 dB PSNR drop
AdpSplit+FastGS	12.6×	similar	Full-schedule quality
BlitzGS	6–10× (city-scale)	Multi-GPU	Within top 3 overall

5. Applications, Generalization, and Downstream Impact

FastGS and its derivatives are deployed broadly in:

Fast multi-view scene reconstruction: Substantial reduction in training time for photorealistic novel view synthesis, city-scale mapping, and dense mesh extraction (Wang et al., 13 May 2026, Huang et al., 8 Jan 2025).
Real-time and resource-constrained inference: Mobile AR/VR and edge devices benefit from ultra-compact Gaussian representations (see also CF3 (Lee et al., 7 Aug 2025)).
Sparse-view and surface reconstruction: FatesGS extends FastGS to rapid, sparse-input surface and mesh modeling, combining photometric, depth, and multi-view feature-losses for improved geometry recovery with fewer Gaussians (Huang et al., 8 Jan 2025).
CT reconstruction: The FaCT-GS system adapts FastGS-inspired kernels and pruning criteria to volumetric CT, reporting 4–13× acceleration over previous methods at comparable accuracy (Pieta et al., 2 Apr 2026).
2DGS/image compression: Fast-2DGS extends fast splatting and density control to 2D image encoding with deep learned priors, achieving >1000 FPS rendering and low-latency/fine-tuned real-time reconstruction (Wang et al., 14 Dec 2025).

The broad generality of FastGS methodologies enables their integration into dynamic scene modeling (4DGS), foundation model feature compression (CF3), accelerated SLAM, large-scale distributed reconstruction (BlitzGS), and compressed storage/streaming.

6. Limitations and Outlook

Complexity and implementation: FastGS requires custom kernel development, fused CUDA/Adam, memory layout management, and careful numerical handling; generic frameworks are not sufficient.
Pruning/densification trade-off: Over-aggressive pruning may sacrifice fine detail; adaptive schemes (e.g., AdpSplit) require careful tuning.
Scalability: While effective to d = 4, for much higher dimensions the curse of dimensionality limits the benefit of tensor-product expansions and adaptivity (Greengard et al., 2023).
Ecosystem adaptation: Plug-and-play compatibility with budget controllers, distributed pipelines, and non-Gaussian (e.g., implicit neural) representations continues to evolve, with recent research focusing on hybrid integration.

Future research directions include further fusion of FastGS with foundation models, end-to-end differentiable feature lifting, real-time video and temporal dimension extensions, and exploration of compression/quantization schemes designed specifically for splat-based representations.

References

(Ren et al., 6 Nov 2025) "FastGS: Training 3D Gaussian Splatting in 100 Seconds"
(Hahlbohm et al., 10 Feb 2026) "Faster-GS: Analyzing and Improving Gaussian Splatting Optimization"
(Lee et al., 7 May 2026) "AdpSplit: Error-Driven Adaptive Splitting for Faster Geometry Discovery in 3D Gaussian Splatting"
(Wang et al., 13 May 2026) "BlitzGS: City-Scale Gaussian Splatting at Lightning Speed"
(Greengard et al., 2023) "A new version of the adaptive fast Gauss transform for discrete and continuous sources"
(Pieta et al., 2 Apr 2026) "FaCT-GS: Fast and Scalable CT Reconstruction with Gaussian Splatting"
(Wang et al., 14 Dec 2025) "Fast 2DGS: Efficient Image Representation with Deep Gaussian Prior"
(Lee et al., 7 Aug 2025) "CF3: Compact and Fast 3D Feature Fields"
(Huang et al., 8 Jan 2025) "FatesGS: Fast and Accurate Sparse-View Surface Reconstruction using Gaussian Splatting with Depth-Feature Consistency"
(Wang et al., 2017) "An adaptive fast Gauss transform in two dimensions"