Faster-GS: Analyzing and Improving Gaussian Splatting Optimization

TL;DR

This work tackles fragmentation and training inefficiency in 3D Gaussian Splatting (3DGS) by surveying existing performance improvements and integrating them into Faster-GS, a high-performance training pipeline. It preserves reconstruction quality while achieving up to 5× faster training and up to 30% VRAM reductions on standard benchmarks, using memory-coalescing techniques, per-Gaussian backward passes, and kernel fusion. The authors also extend the framework to 4D Gaussians for dynamic scenes and provide an open-source testbed to enable fair comparisons across evolving methods. Overall, Faster-GS offers a rigorous, cost-effective baseline that accelerates 3DGS research and enables broader adoption, including non-rigid scene optimization. These advancements address practical bottlenecks and lay groundwork for future enhancements such as second-order optimization and fused training pipelines.

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$\times$ 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.

Figures (3)

• Figure 1: Our method, Faster-GS, substantially accelerates training and reduces GPU memory (VRAM) usage compared to the original 3DGS algorithm kerbl3Dgaussians without altering quality or number of Gaussians (left). Averaged over all Mip-NeRF360 scenes on an RTX 4090 GPU (right), we train 4.1× faster with 30% less VRAM than 3DGS. On the Deep Blending dataset (not depicted), the speedup is more than 5.2×. We also outperform improved implementations from prior works taming3dgsradl2024stopthepophanson2025speedysplat. $^\dagger$Baseline modified to keep quality$\slash$#Gaussians unchanged.
• Figure 2: Runtime comparison for the basis and full versions of our optimized 3DGS framework. We measure the time it takes to compute the forward$\slash$backward pass and the optimizer step respectively during iteration 500, 5000, and 20000 when training four scenes from the Mip-NeRF360 dataset barron2022mipnerf360.
• Figure 3: We show renderings of two of our models. The first one (top) was trained without anti-aliasing techniques, the second one (bottom) has them enabled. It is clearly visible that rendering at a different resolution compared to the one used during training (1×) leads to aliasing artifacts (top). The implemented anti-aliasing techniques significantly reduce these artifacts leading to higher visual fidelity (bottom).