EFA-GS: Eliminating Floating Artifacts in 3DGS

• The paper introduces a novel EFA-GS method that integrates frequency-domain analysis, depth-based expansion, and adaptive scaling to eliminate floating artifacts.
• It refines Gaussian optimization by prioritizing low-frequency learning and dynamically adjusting Gaussian parameters to prevent misaligned visual structures.
• Experimental results show improved PSNR and consistent artifact suppression across diverse benchmarks, ensuring high-fidelity and real-time 3D renderings.

Eliminating-Floating-Artifacts Gaussian Splatting (EFA-GS) is a class of methodologies in 3D Gaussian Splatting (3DGS) that specifically target the suppression and removal of “floating artifacts”—spurious or detached visual structures that misalign with true scene geometry in photorealistic neural rendering. These artifacts are typically caused by spectral or geometric misalignment, density control failures, over-shrinking of Gaussians, or suboptimal optimization regimes, and can significantly degrade depth accuracy and the overall fidelity of 3D reconstructions. EFA-GS approaches are distinguished by explicit algorithmic interventions, including frequency-domain analysis, geometry-aware optimization, and refined adaptive control, to ensure robust, artifact-free rendering across diverse benchmarking environments and practical deployments.

1. Origins and Frequency-Domain Diagnosis of Floating Artifacts

Floating artifacts in 3DGS emerge predominantly from under-optimized or over-shrunk Gaussian primitives, often accentuated in low-quality/null or noisy initialization scenarios or when high-frequency learning is prematurely prioritized. Each Gaussian, represented as: $$\mathcal{G}(x) = \frac{1}{(2\pi)^{3/2} |\Sigma|^{1/2}} \exp\Big[ -\frac{1}{2}(x-p)^{\top} \Sigma^{-1} (x-p) \Big],$$ has a Fourier transform magnitude: $$|\mathcal{F}(\mathcal{G}(x), \omega)| = \exp\left(-\frac{1}{2}\omega^\top \Sigma \omega\right).$$ Small-scale (over-shrunk) Gaussians have a wide frequency spectrum but insufficient coverage of low-frequency components, violating the Nyquist–Shannon criterion necessary for reconstructing correct scene structure. This misallocation leads to structural “floaters,” often exacerbated by noisy initialization which aggravates the nonuniform sampling and optimization pathologies (Wang et al., 4 Aug 2025).

2. Core Methodology of EFA-GS

The EFA-GS paradigm modifies the Gaussian optimization process to prioritize low-frequency learning and systematically detect under-optimized Gaussians. The central mechanism is the Low-Frequency-Come-First (LFCF) algorithm:

• Training monitors the gradient magnitude $\mathrm{Grad}(i)$ of each Gaussian.
• If $\mathrm{Grad}(i) > \tau$ and $\mathrm{Grad}(i) > \mathrm{PGrad}(i)$ (where $\tau$ is a threshold and $\mathrm{PGrad}(i)$ is the previous gradient), scale expansion is performed: $s' = s \cdot c$ ($c > 1$). Otherwise, shrinking or splitting may occur.
• Gaussians with low opacity ($\alpha(i) < \epsilon$) are pruned.

This procedure adaptively encourages Gaussians to capture low-frequency information, thus preventing premature collapse to high-frequency, isolated features which manifest as floaters. Expansion and splitting decisions are further moderated based on scale and recent gradient history to avoid oscillatory or degenerate behavior (Wang et al., 4 Aug 2025).

3. Depth- and Scale-Based Refinements

Naive global expansion could erode high-frequency scene details. EFA-GS strategies refine this with:

• Depth-based Expansion: Each Gaussian’s sampling rate $\nu_i$ (derived, e.g., from camera focal length and depth: $$\nu_i \propto \max_k(1_k(p_i) \cdot \tfrac{f_k}{d_k})$$) is used to interpolate an expansion coefficient $$c(i) = \theta(i) c_{\text{max}} + (1-\theta(i)) c_{\text{min}}$$, with $c_{\text{min}}=1,\, c_{\text{max}}>1$. This reduces excessive expansion for deeper (hard-to-sample) Gaussians, preserving relevance of high frequencies.
• Anisotropic, Volume-Preserving Scaling: Expansion along the shortest axis only, ensuring $\prod c_i = 1$, allows local adjustment to sampling anisotropy without grossly inflating the overall kernel volume, thus mitigating directional blur and further loss of structural detail.

These adjustments balance artifact suppression with detail preservation, targeting floating artifacts only where they statistically arise while maintaining photorealistic fidelity (Wang et al., 4 Aug 2025).

4. Experimental Evidence and Comparative Performance

The efficacy of EFA-GS is supported by controlled experiments on both synthetic and real-world benchmarks. On the RWLQ dataset, EFA-GS improves PSNR by 1.68 dB over a Mip-splatting baseline, substantially reducing the PSNR drop between training and test views. Floating artifacts are shown visually and quantitatively to be markedly reduced, including in highly challenging setups with reflective or fine-detail scenes (Wang et al., 4 Aug 2025). Complementary ablation studies demonstrate the necessity of both depth- and scale-based strategies: disabling either leads to incomplete artifact suppression or unnecessary over-smoothing.

Dataset	Baseline PSNR	EFA-GS PSNR	Floating Artifacts Suppressed
RWLQ (noisy init)	XX.XX	XX.XX	Yes
Tanks and Temples	XX.XX	XX.XX	Yes
Mip-NeRF360	XX.XX	XX.XX	Yes

Table: Summary of PSNR improvements and qualitative artifact suppression (see detailed tables in the source for exact values).

Further, EFA-GS transfers to 3D editing tasks using pipelines such as GaussianEditor; edits such as “Turn it into a polar bear” maintain realistic, artifact-free results in edited scenes—clear evidence that EFA-GS not only benefits static photorealistic rendering but also consistency in downstream manipulation.

5. Theoretical Implications and Algorithmic Integration

EFA-GS situates itself as a frequency-domain regularization technique, complementing (rather than replacing) structure-aware, geometry-encoding, or density-adaptive reforms in 3DGS. The LFCF philosophy can be integrated with:

• Structure-aware graph optimization (Ververas et al., 29 Apr 2024)
• Adaptive density control improvements (Grubert et al., 18 Mar 2025)
• Spectral entropy-based splitting (Huang et al., 19 Sep 2024)
• Event-based or hybrid blending architectures (Yu et al., 29 May 2024, Hou et al., 24 Oct 2024, Liu et al., 5 Aug 2025)

The methodology enhances the Gaussian parameter update schedule, serving as a frequency-based regularizer atop standard photometric and geometric supervision.

6. Implementation Considerations

Practical integration of EFA-GS requires:

• Periodic application of the LFCF schedule during densification.
• Maintenance of per-Gaussian gradient histories.
• Efficient online computation of depth and per-axis sampling rates.
• Compatibility with standard 3DGS frameworks such as Vanilla 3DGS, Mip-splatting, or advanced sort-free renderers.

Further computational overhead is minimized via probabilistic splitting and annealed expansion schedules. The implementation and reproducible codebase are available from the authors at https://jcwang-gh.github.io/EFA-GS (Wang et al., 4 Aug 2025).

7. Applications and Impact

EFA-GS is highly relevant for:

• High-fidelity photorealistic 3D reconstruction from sparse, noisy, or low-quality data.
• Robust scene representations for neural 3D editing pipelines.
• Real-time and interactive rendering in applications where floating artifacts are intolerable, such as VR/AR or digital content creation.
• Integration as a regularization backbone for spectral, graph-based, or density-adaptive neural rendering methods.

This suite of techniques systematically addresses the core limitations in Gaussian Splatting, providing a principled and experimentally verified path toward artifact-free, real-time, and reliable 3D scene representation.