Recent advances in 3D Gaussian Splatting (3DGS) have enabled impressive real-time photorealistic rendering. However, conventional training pipelines inherently assume full multi-view consistency among input images, which makes them sensitive to distractors that violate this assumption and cause visual artifacts. In this work, we revisit an underexplored aspect of 3DGS: its inherent ability to suppress inconsistent signals. Building on this insight, we propose PDF-GS (Progressive Distractor Filtering for Robust 3D Gaussian Splatting), a framework that amplifies this self-filtering property through a progressive multi-phase optimization. The progressive filtering phases gradually remove distractors by exploiting discrepancy cues, while the following reconstruction phase restores fine-grained, view-consistent details from the purified Gaussian representation. Through this iterative refinement, PDF-GS achieves robust, high-fidelity, and distractor-free reconstructions, consistently outperforming baselines across diverse datasets and challenging real-world conditions. Moreover, our approach is lightweight and easily adaptable to existing 3DGS frameworks, requiring no architectural changes or additional inference overhead, leading to a new state-of-the-art performance. The code is publicly available at https://github.com/kangrnin/PDF-GS.
Figures & tables
Figure 2: Self-filtering behavior of vanilla 3DGS. After standard 3DGS training, distractor objects visible in the train view (a) are removed in the rendered result (b), indicating that inconsistent regions tend to either disappear or become blurred in the reconstruction ( Crab2 scene in RobustNeRF dataset [ 21 ] ).
Figure 3: Conceptual illustration of PDF-GS. Our method progressively filters out transient and view-inconsistent distractors through iterative refinement. During the filtering phases, regions exhibiting multi-view inconsistencies are identified by the discrepancies between rendered views and training images. These inconsistent regions are then masked out, while stable view-consistent regions exhibiting small discrepancy are preserved. This progressive filtering process yields a purified 3D Gaussian representation, but often lacks fine-grained details. In the final reconstruction phase, we aim to further restore the fine-grained details that are not fully optimized during the filtering phases. Accordingly, we re-introduce the standard 3DGS loss which leads to both high-fidelity and distractor-free scene reconstruction. The output of the reconstruction phase serves as the final result of our framework.
Figure 4: Effectiveness of the structure-oriented objective (Eq. 4 ). The model trained with standard 3DGS loss (b) overfits transient distractors and exhibits color bleeding, while using structure-oriented objective (c) preserves structural consistency and better suppresses distractors ( Patio scene in NeRF On-the-go [ 18 ] ).
Figure 5: Analysis on the number of filtering phases. Quantitative results show a gradual increase in reconstruction scores as the number of filtering phases increases, with performance saturating around three phases (NeRF On-the-go dataset [ 18 ] ).
Figure 6: Qualitative results of PDF-GS (Ours) and baseline State-of-the-Art methods. Our method generates noticeably fewer distractor-induced artifacts and more accurate reconstruction of static objects and backgrounds than previous methods.
Setting
PSNR
SSIM
LPIPS
Filtering Phase Components
Struct. Loss + Sparse Color †
23.32
0.82
0.15
Struct. Loss only
23.06
0.82
0.15
Sparse Color only
23.20
0.82
0.15
Neither
23.02
0.81
0.15
Re-initialization Scheme
Table 3: Ablation studies on the NeRF On-the-go dataset [ 18 ] . We ablate filtering phase components, re-initialization strategies, and threshold scheduling. † indicates our default setting.
Figure 7: Effect of re-initialization between filtering phases. (a) Without re-initialization. Accumulated errors propagate across filtering iterations and appear as persistent artifacts. (b) With re-initialization (Ours). Such error buildup is avoided, resulting in a cleaner and more stable reconstruction. ( Corner scene in NeRF On-the-go dataset [ 18 ] ).
Figure 8: Progressive mask refinement across phases. The dotted outlines in the images of the leftmost column denote the distractor region, and the red overlay indicates the masked pixels. As training progresses through phases, the masks become increasingly refined and better aligned with true distractors. For example, in the third row ( fountain scene from the NeRF On-the-go dataset [ 18 ] ), background cars that were not masked in earlier phases are correctly identified as distractors in later phases.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Figure A1 : Phase-wise evolution of reconstruction quality. Both datasets show a gradual increase in reconstruction quality as the model advances through the phases of our progressive filtering framework, with improvements accumulating across successive phases.
Method
PSNR
SSIM
LPIPS
RobustSplat ∗ [ 4 ] (DINOv2)
29.24
0.89
0.13
+ Progressive Filtering (DINOv2)
29.32
0.90
0.14
RobustSplat ∗ [ 4 ] (DINOv3)
29.14
0.89
0.13
+ Progressive Filtering (DINOv3)
29.43
0.89
0.15
Appendix
Table A1 : Evaluation of applying our progressive filtering procedure to a single-pass baseline (RobustSplat [ 4 ] ) on the RobustNeRF dataset [ 21 ] . We report results using both DINOv2 [ 15 ] and DINOv3 [ 22 ] backbones, and ∗ indicates our reproduced results.
Masking Method
PSNR
SSIM
LPIPS
Discrepancy-based (Ours)
23.32
0.82
0.15
Learned predictor (RobustSplat)
23.39
0.82
0.14
Appendix
Table A2 : Comparison of different masking strategies within our framework on the NeRF On-the-go dataset [ 18 ] .
Method
Android
Crab2
Statue
Yoda
RobustSplat [ 4 ]
21.6 min
24.5 min
28.7 min
24.2 min
Ours
25.2 min
23.5 min
28.1 min
23.9 min
Appendix
Table A3 : Wall clock training time comparison on the RobustNeRF [ 21 ] dataset.
Method
PSNR
SSIM
LPIPS
DINOv3 †
29.65
0.90
0.13
DINOv2
29.42
0.89
0.14
SSIM
28.88
0.89
0.13
PSNR
28.71
0.89
0.14
Appendix
Table A4: Quantitative results with different masking metrics. † indicates our default setting. Our method with DINOv3 yields the best performance. Even when using low-level metrics such as PSNR and SSIM, our method outperforms vanilla 3DGS.
In-the-wild 3D Gaussian Splatting remains challenging due to transient distractors and illumination-induced cross-view appearance inconsistencies. Existing methods mainly rely on image-level masking to suppress unreliable supervision, but masking alone cannot fully eliminate residual occlusions or resolve illumination-induced inconsistencies, both of which can introduce conflicting cross-view gradients. These unresolved conflicts may destabilize Gaussian optimization and lead to visible reconstruction artifacts. We propose a conflict-aware 3DGS framework that addresses this problem from both image-space supervision and gradient-level optimization. Semantic Consistency-Guided Masking learns pixel-wise consistency scores to adaptively refine prior masks and suppress unreliable supervision before gradient formation. A dual-view Conflict-Aware Gradient Harmonization strategy further reconciles view-specific gradients by mutually rotating them into an orthogonal configuration, reducing negative directional interference across views. We also introduce conflict-aware densification and pruning to stabilize Gaussian growth and remove persistently conflicting primitives. Extensive experiments on standard in-the-wild benchmarks demonstrate that our method achieves state-of-the-art rendering quality under complex transient distractors and cross-view inconsistencies.
Yulei Kang, Tianze Zhu, Jian-Fang Hu +2
Sun Yat-sen University · Northeastern University · Guangdong Province Key Laboratory of Information Security Technology, China +1
3D Gaussian Splatting (3DGS) has demonstrated remarkable performance in novel view synthesis and 3D scene reconstruction, but its quality often degrades in real-world environments due to transient distractors, such as moving objects and varying shadows. Existing methods commonly introduce semantic priors from pre-trained vision models either to group pixels into coherent regions or to define perceptual error metrics. However, semantic grouping is often misaligned with the binary static/transient distinction, while perceptual features can be fragile under appearance perturbations introduced during 3DGS optimization. We propose 3DGS-HPC, a framework that addresses these issues by combining two complementary principles: a patch-wise classification strategy that leverages local spatial consistency for robust region-level decisions, and a hybrid classification metric that adaptively integrates photometric and perceptual cues for more reliable separation. Extensive experiments demonstrate the superiority and robustness of our method in mitigating distractors to improve 3DGS-based novel view synthesis. Our project page is https://cnhaox.github.io/3DGS-HPC/ .
Jiahao Chen, Yipeng Qin, Ganlong Zhao +3
Sun Yat-sen University, Guangzhou, China · Texas A&M University, College Station, USA · Cardiff University, Wales, UK +2
A pretrained 3D Gaussian Splatting (3DGS) reconstruction can reproduce its capture views well while retaining off-object structure that becomes visible under reframing. We present CleanGS, a post-hoc cleanup method that removes such structure without retraining the underlying representation. Given calibrated views and binary target masks, CleanGS scores candidate three-view subsets using mask reliability, camera baseline, viewing-direction diversity, and complementary sparse 3D support. Rather than collapsing these objectives a priori, it computes the exact Pareto front over subsets of the reduced candidate pool and selects a balanced on-dominated subset using weighted Chebyshev scalarization. The selected views then drive rasterization-aware two-of-three voting to retain a subset of the original Gaussians. We evaluate the exported asset directly and introduce a fixed-pose widened-field protocol that exposes structure outside the original framing. Across 18 scenes, CleanGS achieves 19.03 dB standard and 17.83 dB widened-field PSNR, compared with 16.44/16.29 dB for FlashSplat and 15.28/15.10 dB for GaussianCut with the same three masks. CleanGS leads five of six photometric metrics and remains within 0.32 dB of the uncleaned reconstruction in standard PSNR. Its filtering-and-export stage takes a median of 20 seconds and 1.9 GB peak memory, compared with 3.1 minutes and 41.6 GB for FlashSplat and 4.2 minutes and 34.9 GB for GaussianCut. Code is available at https://github.com/smlab-niser/clean-gs/.
Jyotirmaya Shivottam, Om Kumar, Subhankar Mishra
School of Computer Sciences National Institute of Science Education and Research Bhubaneswar, India