Low-Frequency First: Eliminating Floating Artifacts in 3D Gaussian Splatting

TL;DR

This work identifies floating artifacts in 3D Gaussian Splatting (3DGS) as primarily arising from under-optimized Gaussians during low-quality initialization. It introduces Eliminating-Floating-Artifacts Gaussian Splatting (EFA-GS), which uses Low-Frequency-Come-First (LFCF) updates plus depth-based and scale-based strategies to expand under-optimized Gaussians and learn low-frequency content first, thereby suppressing artifacts while preserving high-frequency detail. The approach yields state-of-the-art results on real-world low-quality data (e.g., RWLQ) with PSNR gains and improves downstream 3D editing performance, while maintaining computational efficiency. These findings demonstrate that frequency-aware Gaussian manipulation can robustly mitigate artifacts in 3DGS without compromising detail, enabling more reliable reconstruction and editing in challenging real-world scenarios.

Abstract

3D Gaussian Splatting (3DGS) is a powerful and computationally efficient representation for 3D reconstruction. Despite its strengths, 3DGS often produces floating artifacts, which are erroneous structures detached from the actual geometry and significantly degrade visual fidelity. The underlying mechanisms causing these artifacts, particularly in low-quality initialization scenarios, have not been fully explored. In this paper, we investigate the origins of floating artifacts from a frequency-domain perspective and identify under-optimized Gaussians as the primary source. Based on our analysis, we propose \textit{Eliminating-Floating-Artifacts} Gaussian Splatting (EFA-GS), which selectively expands under-optimized Gaussians to prioritize accurate low-frequency learning. Additionally, we introduce complementary depth-based and scale-based strategies to dynamically refine Gaussian expansion, effectively mitigating detail erosion. Extensive experiments on both synthetic and real-world datasets demonstrate that EFA-GS substantially reduces floating artifacts while preserving high-frequency details, achieving an improvement of 1.68 dB in PSNR over baseline method on our RWLQ dataset. Furthermore, we validate the effectiveness of our approach in downstream 3D editing tasks. Project Website: https://jcwang-gh.github.io/EFA-GS

Figures (10)

• Figure 1: Illustration of our EFA-GS. Ordinary 3DGS frameworks (such as Mip-splatting yu2024mip) sometimes have a frequency bias in the training process and low-quality initialization exacerbate this phenomenon, resulting in more over-shrunk Gaussians. Our EFA-GS successfully mitigate this issue and improve the performance by selectively expanding Gaussians.
• Figure 2: We train Mip-splatting models yu2024mip on ship and kitchen mildenhall2020nerfbarron2022mip for 30k iterations and results are shown in (a), (b) and (c). We inject noise into Gaussian scales of ship and Gaussian coordinates of kitchen respectively. (a),(b) and (d) are the rendered images from testing view whereas (c) are the rendered images from training view(close to the testing view). Obviously, noisy initialization introduces floating artifacts, and (d) demonstrates that EFA-GS (simple) effectively eliminates these floating artifacts.
• Figure 3: We record the numbers of manipulated Gaussians of ship mildenhall2020nerf in the densification process(up to 15,000 iterations) and the average scaling of Gaussians in the whole optimization process. The results demonstrate that 3DGS also learns low frequency information first before diving into high frequency information. Furthermore, when noisy initialization were used, the model would split less Gaussians, clone more Gaussians, and most Gaussians become over-shrunk.
• Figure 4: An demonstration diagram about optimizing Gaussians. According to the Nyquist sampling theorem, normal Gaussians bigger than corresponding sampling intervals are over-optimized, whereas smaller or deeper Gaussians whose scales cannot surpass the sampling intervals are under-optimized.
• Figure 5: Reconstruction results of bicycle model(outdoors). The results shows that erosion of the buildings would arise if the strategies were not used.