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

Nanjing University of Aeronautics and Astronautics
Teaser Banner Teaser Banner

Illustration of our EFA-GS in 3D Reconstruction/Editing tasks. Ordinary 3DGS frameworks (such as Mip-splatting and GaussianEditor) 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.

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 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 theeffectiveness of our approach in downstream 3D editing tasks.

Performance

\[ \begin{array}{c|ccc} \hline & \textrm{PSNR}\uparrow & \textrm{SSIM}\uparrow & \textrm{LPIPS}\downarrow \\ \hline \textrm{2DGS} & \textbf{28.00} & \textbf{0.95} & 0.16 \\ \textrm{GOF} & 27.92 & 0.95 & \textbf{0.15} \\ \textrm{eRank-GS} & 23.16 & 0.91 & 0.18 \\ \hline \textrm{Vanilla 3DGS} & 27.71 & 0.95 & 0.15 \\ \textrm{EFA-GS(3DGS)} & \textbf{28.70} & \textbf{0.95} & \textbf{0.14} \\ \hline \textrm{Mip-splatting} & 26.67 & 0.94 & 0.16 \\ \textrm{EFA-GS(Mip, default)} & 28.35 & 0.95 & 0.15 \\ \hline \end{array} \]

Reconstruction results on RWLQ dataset.

\[ \begin{array}{c|ccc} \hline & \textrm{PSNR}\uparrow & \textrm{SSIM}\uparrow & \textrm{LPIPS}\downarrow \\ \hline \textrm{2DGS} & 27.00 & 0.81 & 0.24 \\ \textrm{GOF} & 27.33 & 0.82 & 0.20 \\ \textrm{eRank-GS} & 27.69 & 0.84 & 0.20 \\ \hline \textrm{Vanilla 3DGS} & 27.58 & 0.82 & 0.21 \\ \textrm{EFA-GS(3DGS)} & 27.52 & 0.82 & 0.21 \\ \hline \textrm{Mip-splatting} & \textbf{27.92} & \textbf{0.84} & \textbf{0.18} \\ \textrm{EFA-GS(Mip, default)} & \textbf{27.94} & \textbf{0.84} & \textbf{0.18} \\ \hline \end{array} \]

Reconstruction results on Mip-NeRF 360 dataset. There are not many floating artifacts in reconstruction results of Mip-NeRF 360 dataset, so our EFA-GS does not bring significant improvement on this dataset.

\[ \begin{array}{c|ccc|ccc} \hline & \textrm{PSNR}\uparrow & \textrm{SSIM}\uparrow & \textrm{LPIPS}\downarrow\\ \hline \textrm{2DGS} & 21.17 & 0.78 & 0.32 \\ \textrm{GOF} & 19.80 & 0.77 & 0.30 \\ \textrm{eRank-GS} & 18.43 & 0.72 & 0.37 \\ \hline \textrm{Vanilla 3DGS} & \textbf{21.51} & \textbf{0.79} & \textbf{0.28} \\ \textrm{EFA-GS(3DGS)} & \textbf{21.69} & \textbf{0.80} & \textbf{0.28} \\ \hline \textrm{Mip-splatting} & 20.63 & 0.78 & 0.29 \\ \textrm{EFA-GS(Mip, default)} & 21.31 & 0.79 & 0.28 \\ \hline \end{array} \]

Reconstruction results on TanksandTemples dataset (Normal initialization).

\[ \begin{array}{c|ccc|ccc} \hline & \textrm{PSNR}\uparrow & \textrm{SSIM}\uparrow & \textrm{LPIPS}\downarrow\\ \hline \textrm{2DGS} & 18.98 & \textbf{0.72} & 0.41 \\ \textrm{GOF} & 17.84 & 0.67 & 0.39 \\ \textrm{eRank-GS} & 16.37 & 0.62 & 0.45 \\ \hline \textrm{Vanilla 3DGS} & \textbf{19.11} & 0.69 & \textbf{0.37} \\ \textrm{EFA-GS(3DGS)} & \textbf{19.45} & 0.69 & \textbf{0.36} \\ \hline \textrm{Mip-splatting} & 18.15 & 0.67 & 0.39 \\ \textrm{EFA-GS(Mip, default)} & 19.07 & 0.69 & 0.37 \\ \hline \end{array} \]

Reconstruction results on TanksandTemples dataset (Low Quality initialization).

BibTeX

@misc{wang2025lowfrequencyfirsteliminatingfloating,
      title={Low-Frequency First: Eliminating Floating Artifacts in 3D Gaussian Splatting}, 
      author={Jianchao Wang and Peng Zhou and Cen Li and Rong Quan and Jie Qin},
      year={2025},
      eprint={2508.02493},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
    }