StableGS: Floater-Free 3D Gaussian Splatting

• StableGS is a 3D Gaussian Splatting framework that addresses floater artifacts by mitigating opacity-color coupling and insufficient geometric constraints.
• It employs a cross-view depth consistency loss and a dual-opacity Gaussian model to suppress spurious Gaussians while preserving translucent materials.
• The method integrates DUSt3R depth priors with dense initialization to improve geometric reconstruction and rendering quality on standard benchmarks.

Searching arXiv for the StableGS paper and closely related 3D Gaussian Splatting works. Searching for "StableGS 3D Gaussian Splatting". StableGS is a framework for 3D Gaussian Splatting (3DGS) that targets a specific optimization pathology of novel view synthesis: floater artifacts, namely small, translucent Gaussians that drift into empty space and persist during training. Introduced in 2025, the method augments the standard 3DGS pipeline with cross-view depth consistency constraints, a dual-opacity Gaussian formulation, and DUSt3R-based depth priors with dense-reconstruction initialization, with the stated goal of eliminating floaters while preserving translucent materials and improving weakly textured geometry (Wang et al., 24 Mar 2025). In the lineage of 3DGS-based radiance-field rendering (Kerbl et al., 2023), StableGS shifts supervision from purely photometric fitting toward a joint photometric-geometric regime.

1. Problem formulation and diagnosis of floater artifacts

StableGS is motivated by the observation that vanilla 3DGS often “locks in” floaters during training (Wang et al., 24 Mar 2025). In the standard 3DGS rendering equation,

$$\sigma_k(x)=\alpha_k\exp\!\Bigl(-\tfrac12(x-\mu_k)^T\Sigma_k^{-1}(x-\mu_k)\Bigr),$$

the opacity $\alpha_k$ and color $c_k$ are jointly optimized. StableGS identifies two intertwined causes of failure.

The first is opacity-color coupling. The reported mechanism is that clusters of floaters can produce blended colors $c_f$ that approximately match the background, so $\partial L/\partial \alpha_k$ and $\partial L/\partial c_k$ vanish. When this occurs, the optimizer cannot “turn off” the spurious Gaussians, and training stalls in local minima (Wang et al., 24 Mar 2025). The second is lack of geometric supervision: purely photometric losses, including $L1$ or SSIM on rendered RGB, do not distinguish true foreground structure from floaters (Wang et al., 24 Mar 2025).

This diagnosis is central to the method’s identity. StableGS does not treat floaters primarily as a rasterization artifact or a post hoc pruning problem; rather, it treats them as an optimization degeneracy induced by coupled appearance-geometry fitting under insufficient geometric constraints. This suggests a reinterpretation of 3DGS instability as a failure of identifiability under photometric supervision alone.

2. Cross-view depth consistency as floater suppression

The first core component is a cross-view depth consistency loss. For two frames $i$ and $j$, StableGS renders depth maps $\bar D_i$ and $\alpha_k$0 alongside color, then uses known camera extrinsics to warp one depth map into the other view, denoted $\alpha_k$1 (Wang et al., 24 Mar 2025). The consistency term is

$\alpha_k$2

Here $\alpha_k$3 is a set of view-pairs chosen to satisfy covisibility, angular, and baseline constraints (Wang et al., 24 Mar 2025). The loss is added to the photometric objective with coefficient $\alpha_k$4, and the reported empirical effect is that it drives $\alpha_k$5 for floaters (Wang et al., 24 Mar 2025).

The mechanism is geometric: a floating Gaussian that projects plausibly in one view but occupies inconsistent depth in another incurs a direct penalty. Because the constraint is mutual and cross-view, it attacks the spatial inconsistency that photometric fitting alone cannot detect. StableGS therefore turns multi-view agreement into an explicit regularizer on Gaussian occupancy rather than relying on RGB residuals to indirectly recover geometry.

3. Dual-opacity Gaussian model and translucent materials

A potential problem with aggressive depth consistency is that it could suppress legitimate translucent objects, especially glass. StableGS addresses this with a dual-opacity Gaussian model in which each Gaussian $\alpha_k$6 carries two opacities, $\alpha_k$7 and an auxiliary $\alpha_k$8 (Wang et al., 24 Mar 2025). The densities are defined as

$\alpha_k$9

Rendering in “standard” mode uses $c_k$0 to produce $c_k$1, while “dual-opacity” mode uses $c_k$2 to produce $c_k$3 (Wang et al., 24 Mar 2025). Over two neighboring frames $c_k$4, the combined color loss is

$c_k$5

The key relation is $c_k$6. StableGS uses this asymmetry so that floaters, which produce anomalously small depths, are eliminated in $c_k$7 under depth consistency, while $c_k$8 can restore true translucency in $c_k$9 and $c_f$0 (Wang et al., 24 Mar 2025). A common misconception is therefore avoided: the method is not a blanket suppression of low-opacity structure. Its stated design is to decouple geometric opacity from material appearance specifically to avoid destroying translucent objects when enforcing depth consistency.

4. DUSt3R depth priors and dense initialization

StableGS states that weakly textured regions, such as white walls, still challenge pure splatting (Wang et al., 24 Mar 2025). Its third component injects external geometric priors through DUSt3R (Wang et al., 2023). The pipeline runs pairwise depth estimation to produce point maps $c_f$1, then solves a global scale optimization: $c_f$2 This yields consistent depth priors $c_f$3 and confidences $c_f$4 (Wang et al., 24 Mar 2025).

The corresponding depth-prior loss is

$c_f$5

weighted by $c_f$6 (Wang et al., 24 Mar 2025).

StableGS further applies a learnable correction

$c_f$7

to DUSt3R depths, fuses them via TSDF to produce an initial dense mesh, and scatters Gaussians for robust initialization (Wang et al., 24 Mar 2025). This part of the framework is significant because it does not merely regularize the final geometry; it changes the starting point of optimization through dense initialization. A plausible implication is that the method reduces both poor local minima and slow geometric convergence in low-texture areas where photometric gradients vanish.

5. Optimization pipeline and objective

The overall optimization pipeline begins with COLMAP for coarse poses and sparse points (Wang et al., 24 Mar 2025). For each pair $c_f$8, StableGS computes DUSt3R outputs $c_f$9, solves the global scale problem, refines depths via $\partial L/\partial \alpha_k$0, and fuses a TSDF to obtain the initial Gaussian set $\partial L/\partial \alpha_k$1 (Wang et al., 24 Mar 2025).

Training then proceeds for $\partial L/\partial \alpha_k$2 to $\partial L/\partial \alpha_k$3, with the example value $\partial L/\partial \alpha_k$4 k iterations. At each iteration, the method renders $\partial L/\partial \alpha_k$5, $\partial L/\partial \alpha_k$6, $\partial L/\partial \alpha_k$7, and $\partial L/\partial \alpha_k$8 for minibatch frames, and optimizes

$\partial L/\partial \alpha_k$9

In the provided LaTeX form, this is written as

$\partial L/\partial c_k$0

Backpropagation updates the Gaussian parameters

$\partial L/\partial c_k$1

(Wang et al., 24 Mar 2025). This unified objective makes the method structurally different from approaches that rely only on RGB reconstruction plus densification heuristics. StableGS explicitly combines appearance fitting, multi-view geometric consensus, and external depth priors within a single optimization loop.

6. Empirical behavior, interpretation, and scope

StableGS is reported to outperform existing state-of-the-art methods across standard benchmarks including MipNeRF360, Blender, and Tanks & Temples (Wang et al., 24 Mar 2025). The detailed example provided is the four indoor MipNeRF360 scenes:

Method	PSNR / SSIM / LPIPS	Setting
3DGS (baseline)	30.46 / 0.9170 / 0.1975	four indoor MipNeRF360 scenes
Mip-Splatting	30.97 / 0.9242 / 0.1870	four indoor MipNeRF360 scenes
RadeGS	30.51 / 0.9213 / 0.1887	four indoor MipNeRF360 scenes
StableGS	31.11 / 0.9222 / 0.1892	four indoor MipNeRF360 scenes
StableGS + prior	31.19 / 0.9245 / 0.1838	four indoor MipNeRF360 scenes
StableGS + prior + DenseInit	31.12 / 0.9246 / 0.1829	four indoor MipNeRF360 scenes

The qualitative claims are equally specific. StableGS renders crisp fine details, with “no grass-blurring,” eliminates sky-region floaters, and accurately reproduces glass and specular materials via its dual-opacity mode (Wang et al., 24 Mar 2025). In weak-texture corners such as blank walls, the depth priors anchor geometry where pure photometric gradients vanish (Wang et al., 24 Mar 2025).

Within the broader Gaussian-splatting literature, StableGS addresses a static-scene training instability—floater artifacts rooted in opacity-color coupling and weak geometry supervision. By contrast, dynamic-scene methods such as SharpTimeGS focus on temporal visibility, motion drift, and densification balance in 4D Gaussian representations rather than on floater suppression in static 3DGS (Liao et al., 3 Feb 2026). The term “StableGS” is also used in unrelated numerical contexts, including stochastic optimization and ghost stabilisation for the Material Point Method (Leplat et al., 2022, Coombs, 2022). In 3D vision, however, StableGS denotes the floater-free 3D Gaussian Splatting framework centered on cross-view depth consistency, dual-opacity disentanglement, and DUSt3R-guided initialization (Wang et al., 24 Mar 2025).