Floating Artifacts in 3D Gaussian Splatting

• Floating artifacts are spurious Gaussian primitives detached from true surfaces, often due to under-constrained photometric optimization and projection errors.
• Adaptive pruning, delayed densification, and geometry priors are employed to suppress these artifacts while preserving essential scene details.
• Balancing image-based supervision with multi-view consistency and specialized rendering techniques is key to mitigating floaters in 3DGS.

Floating artifacts in 3D Gaussian Splatting (3DGS) are spurious Gaussian primitives or rendered structures that are detached from the true scene or object surface. Across the literature, they are described as “large-size tiny-opacity Gaussians” drifting near a human surface, isolated blobs or clusters in front of the camera or in free space, unsupported nearly transparent elements away from the actual surface, and high-opacity primitives wandering in free space (Liu et al., 2023, Hou et al., 11 Mar 2025, Yang et al., 14 Jan 2026, Wang et al., 14 Apr 2026). They arise in multiple settings—few-shot reconstruction, text-driven generation, surface reconstruction, inverse rendering, virtual reality, underwater imaging, and sparse-view restoration—and are linked to under-constrained photometric optimization, unstable densification, distractors and transients, approximate projection and compositing, and medium- or reflection-induced ambiguity (Chung et al., 2023, Fu et al., 3 Jun 2025, Tu et al., 15 May 2025, Kweon et al., 8 Jan 2026).

1. Definitions, terminology, and observable forms

The term floating artifacts does not denote a single visual pattern. In "HumanGaussian" they are a failure mode of SDS-guided 3DGS in which optimization creates “large-size tiny-opacity Gaussians” that drift near the human surface and produce blurry, vaporous blobs disconnected from the actual body (Liu et al., 2023). In "MVGSR" they are spurious Gaussians that do not belong to the actual scene surface but instead appear as isolated blobs or clusters in front of the camera, near distractors, or along uncertain depth regions (Hou et al., 11 Mar 2025). In "TIDI-GS" a floater is an unsupported, nearly transparent Gaussian or cluster of Gaussians that sits in empty space away from the true scene surface (Yang et al., 14 Jan 2026).

Some papers emphasize specific perceptual contexts. "VRSplat" treats floaters as view-inconsistent, stereo-disrupting floating blobs, clouds, or elongated splats that become especially noticeable under the large field of view and frequent head motion of VR (Tu et al., 15 May 2025). "StableGS" describes floaters as sporadic color artifact chunks caused by floating translucent Gaussians in free space (Wang et al., 24 Mar 2025). "ArtifactWorld" places floaters inside a broader taxonomy of sparse-view 3DGS degradations: within the Geometric Structure domain, floaters are “high-opacity primitives wandering in free space,” alongside dilation, needles, and cracks (Wang et al., 14 Apr 2026).

The literature also distinguishes floaters from adjacent artifact classes. "Sort-free Gaussian Splatting via Weighted Sum Rendering" focuses on popping and temporal instability caused by order-dependent alpha compositing rather than on floating artifacts specifically (Hou et al., 2024). "Gaussian Blending" addresses erosion-induced blurring, dilation-induced staircase artifacts, blurred boundaries, occlusion errors, and mentions “floaters” only indirectly in the context of unseen-resolution mipmaps (Koo et al., 19 Nov 2025). This suggests that floating artifacts are best understood as one member of a larger family of geometry, visibility, and compositing failures rather than as a single, uniform defect.

2. Why vanilla 3DGS is prone to floaters

Standard 3DGS represents a scene with anisotropic Gaussians and renders them through alpha compositing. A representative formulation writes the Gaussian density and rendered color as

$$f(\mathbf{x} \mid \boldsymbol{\mu}, \boldsymbol{\Sigma}) = e^{-\frac{1}{2}(\mathbf{x}-\boldsymbol{\mu})^T \boldsymbol{\Sigma}^{-1} (\mathbf{x}-\boldsymbol{\mu})}$$

and

$$\mathbf{C} = \sum_{i=1}^{N} \alpha_i \prod_{j=1}^{i-1}(1-\alpha_j)\mathbf{c}_i,$$

with training commonly driven by photometric losses such as

$$\mathcal{L} = (1-\lambda)\mathcal{L}_1 + \lambda \mathcal{L}_{SSIM}.$$

Several papers argue that this combination is geometrically under-constrained, so Gaussians may move far from their initialization or settle into photometrically convenient but physically implausible positions (Ververas et al., 2024, Jäger et al., 29 Jan 2025).

A recurring diagnosis is that image-only supervision can fit appearance while failing to lock the splats onto a consistent 3D surface. In few-shot settings, "Depth-Regularized Optimization for 3D Gaussian Splatting in Few-Shot Images" states that limited multi-view evidence allows Gaussians to be placed at incorrect depths or scattered in free space, yielding floating artifacts and poor geometry (Chung et al., 2023). "FeatureGS" similarly argues that Gaussian centers and surfaces do not directly represent the object surface under standard photometric optimization, which is why floater artifacts appear and why the number of Gaussians grows unnecessarily (Jäger et al., 29 Jan 2025).

Other papers emphasize optimization pathologies internal to the renderer. "StableGS" attributes floaters to coupled opacity-color optimization that frequently converges to local minima; the paper argues that the model can explain away errors by adjusting color rather than eliminating outlier Gaussians, and that gradient vanishing can then preserve translucent free-space clusters (Wang et al., 24 Mar 2025). "SAGS" attributes floaters to geometry-agnostic per-Gaussian optimization, noting that histograms of displacements show many Gaussians moving far from initialization (Ververas et al., 2024).

Generative and low-quality-initialization settings expose additional mechanisms. In "HumanGaussian," high-variance SDS gradients, together with the original gradient-based densification and pruning strategy, incur blurry results with floating artifacts (Liu et al., 2023). In "Low-Frequency First," the central claim is that floating artifacts arise because some Gaussians become over-shrunk and under-optimized, so they fail to learn reliable low-frequency structure before training moves on to finer detail (Wang et al., 4 Aug 2025). This suggests that floaters are not only a geometric ambiguity but also an optimization-dynamics failure.

3. Densification, pruning, and representation control

A major research direction treats floating artifacts as a consequence of how 3DGS grows, shrinks, clones, splits, or deletes primitives during training. The baseline densification process is repeatedly described as a double-edged mechanism: it is necessary for detail recovery, but under noisy or ambiguous supervision it can allocate capacity to artifacts.

"HumanGaussian" introduces a particularly explicit fix. After structure-aware optimization and adaptive density control, it enters a Gaussian size-based prune-only phase that removes Gaussian instances whose scaling factor exceeds a threshold; in the implementation, the threshold is $0.008$ (Liu et al., 2023). The rationale is that the target artifacts have large scaling factors, tiny opacity, and little contribution to rendered alpha blending, while the SMPL-X-based initialization is already dense and redundant around the body.

"RobustSplat" reframes the problem as premature growth under transient inconsistency. It delays Gaussian splitting and cloning until 10K iterations, pushes opacity reset to start at 15000 iterations with the same 3000-iteration interval, and couples this with scale-cascaded mask bootstrapping (Fu et al., 3 Jun 2025). The paper reports that disabling densification in vanilla 3DGS can significantly reduce artifacts, but at the cost of oversmoothing, so its delayed-growth design aims to preserve later detail recovery without early transient fitting.

"AD-GS" adopts an alternating regime. High densification aggressively adds Gaussians under photometric supervision, while low densification uses an elevated opacity pruning threshold $\epsilon_L > \epsilon_\alpha$, a stricter gradient threshold $\tau_L > \tau_{\text{pos}}$, pseudo-view consistency, and edge-aware depth smoothness (Patle et al., 13 Sep 2025). The method is explicitly motivated by the claim that continuous aggressive densification in sparse-view training creates geometric noise and floaters, whereas alternating growth and cleanup is self-correcting.

"TIDI-GS" generalizes pruning into a conservative evidence-based procedure. A Gaussian becomes a floater candidate only if it satisfies weak-evidence conditions involving multi-view visibility, opacity, learned importance, and an exponential moving average of the position-gradient norm; candidate removal is then further filtered by detail-preserving guards based on non-DC spherical harmonic energy, local texture or color variation, and thinness or anisotropy cues (Yang et al., 14 Jan 2026). Rather than equating low opacity with artifact status, this framework treats floaters as unsupported geometry with low evidence and high spatial isolation.

"MVGSR" uses multi-view evidence to reset transmittance and suppress clutter-induced floaters. Its MV-Prune mechanism defines a multi-view contribution measure and resets transmittance when the contribution exceeds a pruning threshold, with the stated goal of letting gradients flow again in cluttered regions (Hou et al., 11 Mar 2025). The paper reports that MV-Prune reduces storage by 30.19% on average while keeping 99.4% of the original PSNR and only a 0.01 decrease in Chamfer Distance, and also states that it compresses by 60% with comparable rendering quality.

Not all representation-control methods are purely subtractive. "Low-Frequency First" argues that the primary source of floaters is the under-optimized Gaussian, not merely the unpruned Gaussian, and therefore proposes selective expansion through the Low-Frequency-Come-First algorithm (Wang et al., 4 Aug 2025). Under this view, some floaters are better addressed by temporarily enlarging problematic Gaussians so that they learn low-frequency structure before being specialized again.

4. Geometry priors and multi-view consistency constraints

A second major direction introduces explicit geometric supervision so that Gaussians are penalized for being photometrically useful but geometrically unsupported. In few-shot reconstruction, "Depth-Regularized Optimization for 3D Gaussian Splatting in Few-Shot Images" aligns dense monocular depth to sparse COLMAP depth via an affine transform and optimizes a differentiable depth loss,

$$\mathcal{L}_{\text{depth}} = \left\| D - D^{\ast}_{\text{dense}} \right\|_1,$$

supplemented by a smoothness term and early stopping (Chung et al., 2023). On NeRF-LLFF, the reported mean PSNR improves from 12.25/13.75/15.26/16.17 for vanilla 3DGS to 15.94/17.17/18.15/18.74 for 2/3/4/5-view settings.

"SAGS" injects geometry through a local-global graph representation and predicts Gaussian centers as displacements from the original COLMAP points, $$\boldsymbol{\mu}_i = \mathbf{p}_i + \Delta \mathbf{p}_i$$, rather than as unconstrained positions (Ververas et al., 2024). The paper attributes its reduction of floating artifacts and irregular distortions to this structure-aware inductive bias, and also reports up to $24\times$ storage reduction in the lightweight SAGS-Lite variant.

"FeatureGS" changes the objective rather than the parameterization. It adds eigenvalue-derived losses based on planarity, omnivariance, or eigenentropy, with the explicit intent of making Gaussians and their neighborhoods more surface-like and less randomly dispersed (Jäger et al., 29 Jan 2025). In the matched-PSNR setting, the paper reports about a 30% improvement in geometric accuracy, about a 90% reduction in floaters, and about a 90% reduction in the number of Gaussians.

"StableGS" uses bidirectional cross-view depth consistency to expose floaters as inter-view geometric contradictions, and supplements this with a dual-opacity model so that translucent objects are not suppressed together with artifacts (Wang et al., 24 Mar 2025). It further incorporates DUSt3R depth estimation both as a prior loss and as a source of dense initialization for weakly textured regions.

"TriaGS" enforces global geometric consensus through differentiable triangulation-guided geometric consistency. Starting from a rendered 3D point $X_r$, it reprojects that point into a bundle of neighboring views, re-triangulates a consensus point $$\mathbf{C} = \sum_{i=1}^{N} \alpha_i \prod_{j=1}^{i-1}(1-\alpha_j)\mathbf{c}_i,$$0 via differentiable SVD, and penalizes disagreement with a Geman–McClure loss,

$$\mathbf{C} = \sum_{i=1}^{N} \alpha_i \prod_{j=1}^{i-1}(1-\alpha_j)\mathbf{c}_i,$$1

The method uses $$\mathbf{C} = \sum_{i=1}^{N} \alpha_i \prod_{j=1}^{i-1}(1-\alpha_j)\mathbf{c}_i,$$2 neighboring views in the headline experiments, activates TGGC after a 15,000-iteration warm-up in a 30,000-iteration training run, and reports a mean Chamfer Distance of $$\mathbf{C} = \sum_{i=1}^{N} \alpha_i \prod_{j=1}^{i-1}(1-\alpha_j)\mathbf{c}_i,$$3 mm on DTU (Tran et al., 6 Dec 2025).

Underwater reconstruction motivates related but domain-specific geometric constraints. "OceanSplat" adds trinocular view consistency, a synthetic epipolar depth prior from triangulation, a depth residual loss, and a depth-aware alpha adjustment that is active only in early training (Kweon et al., 8 Jan 2026). These components are designed to prevent the scattering medium from being absorbed into the Gaussians as false geometry.

5. Visibility, projection, and appearance-model sources of artifact formation

Not all floating-like failures originate from geometric under-regularization. Several papers argue that the rendering and appearance model itself can create unstable visibility or view-dependent structures that either are floaters or closely resemble them.

"Sort-free Gaussian Splatting via Weighted Sum Rendering" diagnoses standard 3DGS as order-dependent because it uses alpha compositing and therefore requires depth sorting (Hou et al., 2024). The paper does not frame its main problem as floating artifacts specifically, but it explicitly argues that sorting causes temporal popping and other view-instability artifacts, and that its commutative weighted-sum renderer eliminates popping while achieving on average $$\mathbf{C} = \sum_{i=1}^{N} \alpha_i \prod_{j=1}^{i-1}(1-\alpha_j)\mathbf{c}_i,$$4 faster rendering on a Snapdragon 8 Gen 3 GPU. The stated practical implication is indirect but relevant: if floating-like effects stem from unstable compositing, removing sorting addresses part of the mechanism.

"Gaussian Blending" makes a related critique at the pixel level. It argues that scalar alpha blending saturates transmittance across the whole pixel even when a foreground splat covers only part of that pixel, producing erosion-induced blurring, dilation-induced staircase artifacts, incorrect occlusion, and, in mipmap-based settings, unseen-resolution “floaters” (Koo et al., 19 Nov 2025). By treating alpha and transmittance as spatially varying distributions, the method reports more than $$\mathbf{C} = \sum_{i=1}^{N} \alpha_i \prod_{j=1}^{i-1}(1-\alpha_j)\mathbf{c}_i,$$5 lower transmittance error on average than previous methods and maintains real-time rendering speed with no additional memory cost.

"VRSplat" is more direct about floaters. It attributes VR floaters to projection-based distortions arising from the local affine approximation used to project each 3D Gaussian, with error increasing away from the image center and becoming especially visible under large field of view and stereo rendering (Tu et al., 15 May 2025). Its system combines Mini-Splatting, StopThePop, and Optimal Projection, fine-tunes the compact model for 5K iterations without densification, reports 72+ FPS, and validates the result through a controlled user study with 25 participants.

"RTR-GS" identifies yet another source: spherical harmonic overfitting to high-frequency reflective appearance (Zhou et al., 10 Jul 2025). In this account, floaters are small, surface-detached blobs or discontinuities created when a single SH-based forward radiance model tries to explain both geometry and sharp specular behavior. The method separates low-frequency radiance transfer from high-frequency deferred reflections; in the ablation without radiance transfer, novel-view PSNR drops from 33.99 to 32.15, supporting the claim that stronger low-frequency constraints reduce floating artifacts and improve normals.

6. Challenging regimes: distractors, transients, medium effects, and large-scale restoration

Floating artifacts become especially prominent when the scene violates the static, clean, clear-medium assumptions built into vanilla 3DGS. In distractor-heavy data, "Robust 3D Gaussian Splatting for Novel View Synthesis in Presence of Distractors" argues that moving people, household items, shadows, or other transient content get reconstructed as view-dependent effects or floating artifacts unless they are explicitly ignored during optimization (Ungermann et al., 2024). The method learns residual-based distractor masks, refines them with logistic regression and Segment Anything, and applies the loss only on unmasked pixels; on distractor-polluted scenes it reports PSNR 28.39, SSIM 0.8875, and LPIPS 0.1321 versus 26.53, 0.87253, and 0.1568 for standard Gaussian Splatting, with less than 1% runtime overhead from SAM.

"MVGSR" and "RobustSplat" specialize this idea for surface reconstruction and transient-free synthesis, respectively. "MVGSR" describes distractors as artifacts clustered in front of the camera lens or as viewpoint-dependent color representations attached to an object's surface, and uses DINOv2-based feature disagreement, SAM refinement, multi-view consistency loss, and MV-Prune to suppress them (Hou et al., 11 Mar 2025). "RobustSplat" argues that densification itself models transient disturbances unless it is delayed, and reports best mean scores of 23.22 PSNR / 0.818 SSIM / 0.149 LPIPS on NeRF On-the-go and 29.36 / 0.895 / 0.135 on RobustNeRF (Fu et al., 3 Jun 2025).

Underwater reconstruction introduces a different ambiguity: medium effects can be encoded as geometry. "OceanSplat" states that medium intensity is often absorbed into the 3D Gaussians, leading to many floating artifacts, and combats this with trinocular consistency and depth-aware opacity control (Kweon et al., 8 Jan 2026). "UW-3DGS" formalizes the same issue as a physics plus uncertainty problem and adds Physics-Aware Uncertainty Pruning (PAUP), which scores each Gaussian using view instability and underwater image-formation inconsistency (Xing et al., 8 Aug 2025). The paper reports a Big Float Gauss Ratio of 1.3% for the full model versus 8.2% without PAUP, and repeatedly claims about a 65% reduction in floating artifacts.

At the opposite end of the pipeline, "ArtifactWorld" treats floaters as targets for post hoc restoration rather than only for in-training prevention (Wang et al., 14 Apr 2026). It introduces a phenomenological taxonomy of 4 domains and 9 artifact types, constructs 107.5K paired video clips, predicts artifact heatmaps inside a video diffusion backbone, and uses closed-loop generative reconstruction to distill repaired views back into 3DGS. This suggests that, for very sparse inputs, some floating artifacts are now being treated as a restoration problem as well as a reconstruction problem.

7. Empirical patterns, trade-offs, and recurrent misconceptions

A recurrent misconception is that stronger image guidance alone is sufficient. "HumanGaussian" explicitly separates cleaner gradients from actual artifact elimination: annealed negative prompt guidance addresses over-saturation, but the floating artifacts themselves are eliminated in the late Gaussian size-based prune-only phase rather than by the diffusion model (Liu et al., 2023). A related misconception is that low-opacity pruning alone solves the problem. "TIDI-GS" rejects this by showing that legitimate thin structures and specular details can also have low opacity, which is why it introduces detail-preserving guards (Yang et al., 14 Jan 2026).

The literature also shows that anti-floater interventions often create measurable trade-offs. In the fixed-iteration setting, "FeatureGS" achieves substantial geometric gains and floater reduction at the cost of about a 3.3 dB PSNR drop relative to baseline 3DGS (Jäger et al., 29 Jan 2025). "Sort-free Gaussian Splatting via Weighted Sum Rendering" improves stability and speed but explicitly describes the renderer as a generalized, less physically constrained approximation (Hou et al., 2024). "StableGS" notes that the standard stream used for depth consistency can reduce depth precision slightly, while the dual-opacity stream is used to recover detail and translucency (Wang et al., 24 Mar 2025). "TIDI-GS" likewise keeps its monocular depth loss intentionally soft because monocular depth can be noisy or unreliable on reflective boundaries, transparent regions, and object edges (Yang et al., 14 Jan 2026).

Across papers, the most consistent empirical pattern is that floating artifacts diminish when the model is forced to satisfy evidence beyond raw RGB reprojection. That evidence may take the form of aligned monocular depth, DUSt3R priors, local-global scene graphs, eigenvalue-based geometric regularizers, triangulated multi-view consensus, distractor masks, delayed densification, physics-aware opacity control, or improved compositing and projection (Chung et al., 2023, Ververas et al., 2024, Wang et al., 24 Mar 2025, Tran et al., 6 Dec 2025, Kweon et al., 8 Jan 2026). A plausible implication is that “floating artifacts in 3D Gaussian Splatting” is less a single bug than a recurring symptom of mismatches among supervision, visibility, densification, and scene assumptions.