[论文解读] Faster-GS: Analyzing and Improving Gaussian Splatting Optimization

Florian Hahlbohm, Linus Franke+2 more|arXiv (Cornell University)|Feb 10, 2026

Computer Graphics and Visualization Techniques被引用 0

一句话总结

Faster-GS 将有效的三维高斯点画（3D Gaussian Splatting）训练优化进行整合并增强，在保持重建质量的前提下实现高达 5× 的更快训练，并将该收益扩展到 4D（动态）场景。

ABSTRACT

Recent advances in 3D Gaussian Splatting (3DGS) have focused on accelerating optimization while preserving reconstruction quality. However, many proposed methods entangle implementation-level improvements with fundamental algorithmic modifications or trade performance for fidelity, leading to a fragmented research landscape that complicates fair comparison. In this work, we consolidate and evaluate the most effective and broadly applicable strategies from prior 3DGS research and augment them with several novel optimizations. We further investigate underexplored aspects of the framework, including numerical stability, Gaussian truncation, and gradient approximation. The resulting system, Faster-GS, provides a rigorously optimized algorithm that we evaluate across a comprehensive suite of benchmarks. Our experiments demonstrate that Faster-GS achieves up to 5$ imes$ faster training while maintaining visual quality, establishing a new cost-effective and resource efficient baseline for 3DGS optimization. Furthermore, we demonstrate that optimizations can be applied to 4D Gaussian reconstruction, leading to efficient non-rigid scene optimization.

研究动机与目标

• Survey and consolidate performance-enhancing strategies in 3D Gaussian Splatting (3DGS) prior work.
• Identify bottlenecks in training and propose broadly applicable optimizations.
• Integrate novel optimizations to form a fast, modular 3DGS training pipeline (Faster-GS).
• Demonstrate that optimizations generalize to multi-dimensional (4D) Gaussian reconstruction.

提出的方法

• Review and categorize memory-access bottlenecks in 3DGS training.
• Integrate techniques: tighter, opacity-aware 2D bounding boxes; per-tile splat lists; two-stage sorting with stable sort; per-Gaussian backward pass; fused activations; memory-coalesced layouts; fused Adam updates; z-ordering of Gaussians during densification; fusion of backward pass with optimizer steps.
• Apply all optimizations to the original CUDA-based differentiable rasterization pipeline while preserving compatibility with common 3DGS codebases.
• Extend the optimized pipeline to 4D Gaussian Splatting for dynamic scene reconstruction.
• Provide a refactored basis implementation to improve numerical stability and VRAM handling; enable easy integration into existing 3DGS frameworks.

实验结果

研究问题

• RQ1 How much faster can training become when applying a unified set of memory- and computation-centric optimizations to 3DGS without compromising reconstruction quality?
• RQ2 Which optimizations yield the largest speedups and memory savings when evaluated across standard 3DGS benchmarks and datasets?
• RQ3 Can the optimization techniques for 3DGS training be extended to 4D (dynamic) Gaussian Splatting without loss of fidelity or performance?
• RQ4 What is the effect of fused updates and per-Gaussian backward passes on training stability and VRAM usage?

主要发现

• Faster-GS achieves up to 5× faster training compared with the original 3DGS implementation on benchmark scenes while maintaining reconstruction quality.
• Average reconstruction time for the 3DGS benchmark with full quality and all Gaussians is 163 seconds under the Faster-GS pipeline.
• Various optimizations reduce memory traffic and improve cache locality, enabling substantial speedups across multiple datasets (Mip-NeRF360, Tanks and Temples, Deep Blending).
• A refined basis implementation already yields sizable gains (~15% faster than the baseline) before additional optimizations are applied.
• Fusing gradient computations and parameter updates, along with per-Gaussian backward passes and activated kernel fusion, yields notable reductions in training time and VRAM usage (examples shown in ablations across outdoor/indoor scenes).
• The framework extends to 4D Gaussian reconstruction, enabling efficient non-rigid dynamic scene optimization.