RobustGS: Enhancing 3D Gaussian Splatting

• RobustGS is a robustness-oriented variant of 3D Gaussian Splatting that mitigates degradation in reconstruction by integrating feature enhancement and adaptive smoothing.
• It employs a Generalized Degradation Learner and a semantic-aware multi-view state-space module to maintain cross-view consistency and accurate geometric correspondence.
• In SLAM applications, the CB-KNN method adaptively smooths local Gaussian projections to stabilize pose optimization without altering the core 3D representation.

RobustGS denotes robustness-oriented variants of 3D Gaussian Splatting (3DGS) that are designed to preserve reliable reconstruction, tracking, or localization when the assumptions of clean imagery, perfectly estimated Gaussian parameters, or artifact-free rendering break down. In the current literature, the label is used most explicitly for a plug-and-play module that strengthens feedforward 3DGS under low-quality imaging conditions, and it is also used in a broader SLAM sense for a 3DGS-based tracking strategy that improves robustness by regularizing rasterization with adaptive kernel smoothing rather than by pursuing maximal rendering sharpness alone (Wu et al., 5 Aug 2025, Zhang et al., 28 Nov 2025).

1. Scope and technical setting

RobustGS arises within the broader transition of 3DGS from a rendering primitive to a general scene representation for reconstruction, SLAM, and localization. Standard 3DGS methods are highly effective when views are clean, scenes are static, and Gaussian parameters are well estimated, but they are vulnerable to degradation, outliers, transient content, blur, floaters, and pose-optimization instability. RobustGS methods address these failure modes by intervening upstream in feature extraction, during rasterization, or inside the optimization loop itself (Wu et al., 5 Aug 2025, Zhang et al., 28 Nov 2025).

Variant	Problem setting	Core mechanism
RobustGS (Wu et al., 5 Aug 2025)	Feedforward 3DGS under low-quality conditions	Generalized Degradation Learner and semantic-aware Multi-View State-Space Enhancement Module
RobustGS via CB-KNN (Zhang et al., 28 Nov 2025)	3DGS-based SLAM	Adaptive local smoothing of the K-nearest Gaussians during rendering for pose optimization

This usage sits within a wider robust-3DGS research landscape. Large-scale UAV localization has introduced scale-aware pose initialization and Laplacian-based reliability masking for artifact-aware refinement, while transient-free reconstruction has motivated delayed densification and coarse-to-fine mask bootstrapping (Zhang et al., 7 Apr 2026, Fu et al., 3 Jun 2025). A plausible implication is that “robustness” in 3DGS has become a systems property rather than a single module-level objective: different pipelines secure robustness at different points of failure.

2. RobustGS for feedforward 3D Gaussian Splatting

The feedforward variant of RobustGS addresses the observation that existing feedforward 3DGS approaches typically assume that input multi-view images are clean and high-quality. Under noise, low light, or rain, these methods produce inaccurate geometry and degraded 3D reconstruction. RobustGS is introduced as a general and efficient multi-view feature enhancement module that can be seamlessly integrated into existing pretrained pipelines in a plug-and-play manner, without retraining the base model (Wu et al., 5 Aug 2025).

Its design is organized around two principal components. The first is the Generalized Degradation Learner (GenDeg), which extracts generic representations and distributions of multiple degradations from multi-view inputs. The second is the Multi-View State-Space Enhancement Module (MV-SSEM), which restores corrupted features before 3D Gaussian generation. MV-SSEM includes a semantic-aware state-space model: it first uses degradation representations to enhance corrupted inputs in feature space, then aggregates semantically similar information across different views to extract fine-grained cross-view correspondences. This combination is intended to avoid the cross-view inconsistency introduced by naive single-image restoration and to preserve the geometry-critical correspondence structure required by feedforward 3DGS.

The degradation learner is supervised by reconstruction, contrastive, and classification terms. In the notation given for RobustGS, the losses are

$$\mathcal{L}_{\text{rec}} = \lambda \| \hat{I}_{\mathrm{deg}} - I_{\mathrm{deg}} \|_1 + \mathcal{L}_{\text{perc}}(\hat{I}_{\mathrm{deg}}, I_{\mathrm{deg}})$$

$$\mathcal{L}_{\text{con}} = -\log \frac{\exp(\langle z_i, z_j \rangle/\tau)}{\sum_{k \ne i} \exp(\langle z_i, z_k \rangle/\tau)}$$

$$\mathcal{L}_{\text{cls}} = \operatorname{CrossEntropy}\!\left(f_{\text{cls}}(z_{\text{deg}}), y_{\text{deg}}\right)$$

with the final objective

$$\mathcal{L} = \lambda_1 \mathcal{L}_{\text{rec}} + \lambda_2 \mathcal{L}_{\text{con}} + \lambda_3 \mathcal{L}_{\text{cls}}.$$

The state-space update is given as

$$h_k = \overline{A} h_{k-1} + \overline{B} x_k,\qquad y_k = C h_k + D x_k,$$

and degradation-aware modulation adjusts the SSM parameters through

$$W_{\text{mod}} = W \odot \phi_W(z_{\text{deg}}),\quad W \in \{B, C, \Delta\}.$$

Semantic grouping is performed by routing tokens to learned prompts via Gumbel-Softmax,

$$\mathbf{w}_i = \mathrm{GumbelSoftmax}(\mathrm{sim}(\mathbf{e}_i, \mathcal{P})),$$

followed by semantic modulation of the output projection,

$$C_i^{\text{mod}} = C + \sum_{k=1}^K w_{ik} \cdot \mathbf{p}_k.$$

A central design claim of this variant is that enhancement should occur in multi-view feature space rather than as independent image restoration before or after reconstruction. This reframes robustness from an image-enhancement problem into a representation-consistency problem.

3. Training protocol, ablations, and reported behavior under degradations

The reported training configuration uses DIV2K for GenDeg with synthetic degradations including dark, fog, rain, snow, high-contrast, and impulse noise, while the 3DGS pipeline is evaluated on RealEstate10K using the same degradation procedures (Wu et al., 5 Aug 2025). RobustGS is integrated into existing feedforward 3DGS models such as PixelSplat and MVSplat without retraining the base model, and it is compared with image restoration baselines including PromptIR and AdaIR in three modes: restoration before 3D reconstruction, retraining under the same degradations, and restoration after 3D reconstruction.

The reported empirical pattern is consistent across degradation types. RobustGS is described as consistently outperforming prior baselines across all degradations and on average; for the PixelSplat backbone, the summary reports an average PSNR improvement of approximately $2$ dB, and the same trend is stated to hold for MVSplat. The method is also reported to be more efficient than strong image restoration alternatives, with lower parameter count, less computation, and faster runtime. Qualitatively, RobustGS is reported to yield reconstructions that are more faithful and artifact-free under dark, fog, rain, and related corruptions, whereas competing baselines suffer from geometric and texture artifacts due to cross-view inconsistency.

The ablation results emphasize that the three GenDeg losses contribute non-redundant gains; degradation embedding injection into the SSM is more effective than simple concatenation; semantic prompt pooling performs best at $64$ clusters and $$\mathcal{L}_{\text{con}} = -\log \frac{\exp(\langle z_i, z_j \rangle/\tau)}{\sum_{k \ne i} \exp(\langle z_i, z_k \rangle/\tau)}$$0-dimensional prompts; and multi-view enhancement outperforms single-view enhancement. The semantic-guided reordering mechanism is also identified as essential. These findings position RobustGS as a modular enhancement layer rather than a new feedforward backbone.

4. RobustGS for 3DGS-based SLAM via adaptive kernel smoothing

A distinct usage of RobustGS appears in 3DGS-based SLAM, where the central claim is that rendering quality is not the primary determinant of tracking accuracy. The method proposed in "Robust 3DGS-based SLAM via Adaptive Kernel Smoothing" argues instead that the rasterization process must be made resilient to imperfections in 3DGS parameters. Its practical mechanism is Corrective Blurry KNN (CB-KNN), which adaptively modifies the RGB values and locations of the K-nearest neighboring Gaussians within a local region, introducing a controlled blur that regularizes the pose-optimization stage while leaving the map itself unchanged (Zhang et al., 28 Nov 2025).

For each pixel $$\mathcal{L}_{\text{con}} = -\log \frac{\exp(\langle z_i, z_j \rangle/\tau)}{\sum_{k \ne i} \exp(\langle z_i, z_k \rangle/\tau)}$$1, the method selects the set $$\mathcal{L}_{\text{con}} = -\log \frac{\exp(\langle z_i, z_j \rangle/\tau)}{\sum_{k \ne i} \exp(\langle z_i, z_k \rangle/\tau)}$$2 of K-nearest Gaussians based on 2D projections. It computes the 2D centroid

$$\mathcal{L}_{\text{con}} = -\log \frac{\exp(\langle z_i, z_j \rangle/\tau)}{\sum_{k \ne i} \exp(\langle z_i, z_k \rangle/\tau)}$$3

then shifts each Gaussian projection toward that centroid:

$$\mathcal{L}_{\text{con}} = -\log \frac{\exp(\langle z_i, z_j \rangle/\tau)}{\sum_{k \ne i} \exp(\langle z_i, z_k \rangle/\tau)}$$4

where $$\mathcal{L}_{\text{con}} = -\log \frac{\exp(\langle z_i, z_j \rangle/\tau)}{\sum_{k \ne i} \exp(\langle z_i, z_k \rangle/\tau)}$$5 and $$\mathcal{L}_{\text{con}} = -\log \frac{\exp(\langle z_i, z_j \rangle/\tau)}{\sum_{k \ne i} \exp(\langle z_i, z_k \rangle/\tau)}$$6 avoids division by zero. Color is also smoothed through a weighted local average,

$$\mathcal{L}_{\text{con}} = -\log \frac{\exp(\langle z_i, z_j \rangle/\tau)}{\sum_{k \ne i} \exp(\langle z_i, z_k \rangle/\tau)}$$7

Rendering then uses the corrected Gaussians $$\mathcal{L}_{\text{con}} = -\log \frac{\exp(\langle z_i, z_j \rangle/\tau)}{\sum_{k \ne i} \exp(\langle z_i, z_k \rangle/\tau)}$$8 for color and depth:

$$\mathcal{L}_{\text{con}} = -\log \frac{\exp(\langle z_i, z_j \rangle/\tau)}{\sum_{k \ne i} \exp(\langle z_i, z_k \rangle/\tau)}$$9

$$\mathcal{L}_{\text{cls}} = \operatorname{CrossEntropy}\!\left(f_{\text{cls}}(z_{\text{deg}}), y_{\text{deg}}\right)$$0

The neighborhood size is adaptive:

$$\mathcal{L}_{\text{cls}} = \operatorname{CrossEntropy}\!\left(f_{\text{cls}}(z_{\text{deg}}), y_{\text{deg}}\right)$$1

with dependence on local Gaussian density $$\mathcal{L}_{\text{cls}} = \operatorname{CrossEntropy}\!\left(f_{\text{cls}}(z_{\text{deg}}), y_{\text{deg}}\right)$$2 and scene motion complexity $$\mathcal{L}_{\text{cls}} = \operatorname{CrossEntropy}\!\left(f_{\text{cls}}(z_{\text{deg}}), y_{\text{deg}}\right)$$3. Crucially, all corrections are per-frame or per-keyframe and do not change the actual 3DGS parameters of the map. This separation is the defining property of the method: robustness is injected into the rendering used for tracking, not into the stored reconstruction.

Within the SLAM pipeline, CB-KNN is applied only during keyframes for computational efficiency and maximal effect. Non-keyframes use standard rendering with constant-velocity initialization. During keyframe pose optimization, only pixels with high coverage $$\mathcal{L}_{\text{cls}} = \operatorname{CrossEntropy}\!\left(f_{\text{cls}}(z_{\text{deg}}), y_{\text{deg}}\right)$$4 contribute to the loss:

$$\mathcal{L}_{\text{cls}} = \operatorname{CrossEntropy}\!\left(f_{\text{cls}}(z_{\text{deg}}), y_{\text{deg}}\right)$$5

This formulation explicitly privileges stable rasterization over uncompromising sharpness during camera tracking.

5. Reported SLAM robustness and reconstruction performance

The mechanism-level interpretation of CB-KNN is stated in four parts: suppression of outliers, filling discontinuities, RGB noise regularization, and adaptive smoothing that increases in sparse, noisy, or fast-moving regions while decreasing in dense and detail-rich areas (Zhang et al., 28 Nov 2025). Because the underlying 3DGS parameters remain unaltered, the reported claim is that scene reconstruction quality is not sacrificed for tracking robustness.

On Replica, TUM-RGBD, and ScanNet, the method is reported to yield smoother and more accurate trajectories, closely matching ground truth and exhibiting less drift or loss. The summary gives a concrete Replica result in which average ATE RMSE is reduced from $$\mathcal{L}_{\text{cls}} = \operatorname{CrossEntropy}\!\left(f_{\text{cls}}(z_{\text{deg}}), y_{\text{deg}}\right)$$6 cm for SplaTAM to $$\mathcal{L}_{\text{cls}} = \operatorname{CrossEntropy}\!\left(f_{\text{cls}}(z_{\text{deg}}), y_{\text{deg}}\right)$$7 cm for the proposed method, with similar improvements on TUM-RGBD and ScanNet. Relative Pose Error is reported to outperform previous methods on all benchmarks. Mapping quality, measured by PSNR, SSIM, LPIPS, and depth $$\mathcal{L}_{\text{cls}} = \operatorname{CrossEntropy}\!\left(f_{\text{cls}}(z_{\text{deg}}), y_{\text{deg}}\right)$$8, is stated to match or slightly improve over SplaTAM, and the method is reported to introduce no observable blur in training or novel-view rendering despite intentional smoothing during pose optimization.

Metric	SplaTAM	CB-KNN
ATE RMSE (Replica, cm)	0.39	0.32
PSNR (Replica, dB)	34.11	34.17
Keyframe time (ms)	3.05	3.42

The efficiency claim is correspondingly narrow: CB-KNN on keyframes adds negligible overhead, with keyframe optimization reported at approximately $$\mathcal{L}_{\text{cls}} = \operatorname{CrossEntropy}\!\left(f_{\text{cls}}(z_{\text{deg}}), y_{\text{deg}}\right)$$9 ms/frame. The method is also described as enabling faster tracking and mapping iterations than prior art because optimization is more stable. A common misconception addressed by this work is that sharper rendering necessarily yields better tracking; the reported evidence instead supports a deliberate trade-off in which controlled blur can improve pose estimation.

6. Relation to adjacent robust 3DGS methods

RobustGS is part of a larger family of 3DGS methods that isolate specific failure modes and design robustness where those failures originate. In large-scale UAV localization, LSGS-Loc introduces a scale-aware pose initialization strategy that combines scene-agnostic relative pose estimation with explicit 3DGS scale constraints, together with a Laplacian-based reliability mask that restricts photometric refinement to high-quality regions (Zhang et al., 7 Apr 2026). This addresses two issues that also motivate RobustGS in SLAM: unstable initialization and sensitivity to unreliable rendered regions.

For dynamic scenes, RobustSplat identifies densification itself as a source of transient artifacts. Its delayed Gaussian growth strategy postpones splitting and cloning so that the static scene is optimized before mask reliability becomes critical, while scale-cascaded mask bootstrapping uses lower-resolution feature similarity supervision first and later shifts to high-resolution supervision for precise transient masks (Fu et al., 3 Jun 2025). This is conceptually aligned with RobustGS in SLAM: both treat early, unstable optimization as the point where brittle decisions propagate into persistent failure.

A further orthogonal direction is robustness to compression and copyright-preserving deployment. CompMarkGS embeds watermarks in anchor features, injects quantization noise through a quantization distortion layer, and uses frequency-aware anchor growing to preserve watermark integrity and rendering quality under quantization-based compression (In et al., 17 Mar 2025). Although this is not a reconstruction- or tracking-robustness method, it shows that robustness in 3DGS also extends to deployment-time distortions and model manipulation.

Taken together, these works indicate that RobustGS is best understood not as a single canonical algorithm but as a research orientation within 3DGS. Across feedforward reconstruction, SLAM, localization, transient suppression, and compressed deployment, the shared principle is to decouple nominal rendering fidelity from downstream stability whenever the two objectives conflict.