CleanGS: Pareto-Guided Post-Hoc Cleanup of 3D Gaussian Splatting Scenes

Jyotirmaya Shivottam; Om Kumar; Subhankar Mishra

cs.CVJan 1, 2026

CleanGS: Pareto-Guided Post-Hoc Cleanup of 3D Gaussian Splatting Scenes

Authors: Jyotirmaya Shivottam, Om Kumar, Subhankar Mishra

Organizations: School of Computer Sciences National Institute of Science Education and Research Bhubaneswar, India

Abstract

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/.

Figures & tables

Figure 1. Overview of CleanGS . Candidate masks are scored for reliability, and three-view subsets for camera baseline, viewing-angle diversity, and complementary masked 3D support. We compute the exact Pareto front over the reduced candidate pool, select one balanced subset from it, then apply rasterization-aware two-of-three voting. Retained Gaussians are exported unchanged as a subset of the input scene. Five stage cards. Stage one is a pretrained Gaussian scene with calibrated cameras. Stage two scores each view for reliability, baseline diversity, support diversity and viewing-angle diversity. Stage three plots all subsets in objective space, highlights the Pareto front and marks the selected subset, with the three selected masks below. Stage four projects Gaussians into the selected views and marks consistent ones as kept and inconsistent ones as discarded. Stage five shows the cleaned asset rendered as splats, as covariance ellipses and as points.

Panel of Figure 2 — Figure 2. Ordinary framing hides retained clutter. Top: captured view. Bottom: $6\times$ wider field of view at the same pose. The uncleaned asset reveals substantial off-object structure, while FlashSplat and GaussianCut remove the monument’s plinth. CleanGS preserves the plinth while remaining clean after widening. GaussianCut † denotes the all-mask reference. Captured and widened renderings of a statue. Widening reveals off-object structure in the uncleaned asset; CleanGS preserves the plinth removed by the baselines.

Panel of Figure 3 — Figure 3. Selected three-view subsets used for Gaussian filtering $S^{\star}$ for Family , tractor , and table , with masks outlined. The selected views balance mask reliability with geometric and support diversity. Three selected masked views for each of the three scenes.

	Masks used	Standard view			Widened field of view			Leakage ( $1\times$ ) $\downarrow$	Retained $\downarrow$
Method	for cleanup	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Uncleaned 3DGS	0	19.35	0.620	0.194	12.84	0.300	0.327	0.983	100.0%
FlashSplat ( Shen et al., 2024 )	selected 3	16.44	0.557	0.257	16.29	0.411	0.205	0.141	14.9%
GaussianCut ( Jain et al., 2024 )	selected 3	15.28	0.533	0.279	15.10	0.381	0.227	0.089	17.2%
GaussianCut † ( Jain et al., 2024 )	all capture	17.27	0.572	0.234	16.71	0.421	0.194	0.055	20.2%
CleanGS (ours)	selected 3	19.03	0.643	0.188	17.83	0.501	0.167	0.206	22.8%

Table 1. Retained-asset comparison on 18 scenes. Standard: captured framing; Widened:

6\times

fixed-pose widened field of view. All methods use the same scored frames and evaluation masks; FlashSplat, GaussianCut, and CleanGS use the same three selected masks, and GaussianCut † uses all capture masks. Best photometric result is in bold ; best cleanup statistic is underlined .

Panel of Figure 4 — Figure 4. Retained target appearance. Captured and $6\times$ widened renderings use the same camera pose, so widening reveals content already retained in the exported asset. FlashSplat introduces visible target artifacts, while GaussianCut often produces tighter silhouettes at the cost of darker target regions. CleanGS better preserves details such as the plinth, inscription, lettering, and fountain basin. Captured and widened renderings of four scenes comparing the uncleaned scene, FlashSplat, matched three-mask GaussianCut, and CleanGS.

Method	Time $\downarrow$	Range	Peak memory $\downarrow$
FlashSplat ( Shen et al., 2024 )	3.1 min	0.7–6.2 min	41.6 GB
GaussianCut ( Jain et al., 2024 )	4.2 min	1.0–14.4 min	34.9 GB
CleanGS (ours)	20 s	12–58 s	1.9 GB

Table 2. Filtering and export cost per scene over 18 scenes at the matched three-mask budget. Times and peak memory exclude view selection.

Panel of Figure 5 — Figure 5. Residual target-adjacent structure. Nearby surfaces in super markets and picnic remain supported across selected views. Their widened leakage is $0.221$ and $0.099$ , against a benchmark mean of $0.066$ . GaussianCut † removes these regions with all-mask supervision; on picnic , it is $1.49\,\textrm{dB}$ better in widened PSNR. Two rows showing ground truth, GaussianCut widened, and CleanGS at both fields of view, for the picnic and super markets scenes, where CleanGS retains a large attached surface.

Appendix figures & tables27 assets

Supplementary material from the paper’s appendix.

Appendix

Measure	Selected	Random	$p$
Pairwise $\phi$ , co-visible Gaussians (mean)	0.61	0.51	0.32
$\Pr[\nu\geq 2]$ / independence prediction (mean)	3.3	2.6	0.09
Supported by exactly one view (fraction)	0.27	0.45	0.001

Appendix

Table 3. Empirical dependence of per-view support. Support indicators

X_{v}(g)

of the three filtering views over 18 scenes. Random subsets use seeds 0–2, averaged within scene; two table runs with an unregistered camera are excluded.

p

is a two-sided Wilcoxon signed-rank test.

Scene	Cap 60 (main)	Cap 120	Uncapped
century	19.94 dB, 0.391	18.98 ( $-0.96$ ), 0.351	18.57 ( $-1.37$ ), 0.347
red brick build	16.34 dB, 0.196	16.82 ( $+0.48$ ), 0.201	17.63 ( $+1.29$ ), 0.207

Appendix

Table 4. Effect of the 60-view pool cap. Standard (captured-framing) PSNR and leakage on the two scenes with more than 60 eligible candidates;

\Delta

is relative to the main cap of 60.

Stage	Median	Range
Selection, explicit front (main)	21.0 s	2.5–101.9 s
mask scoring, COLMAP, view data	$\sim$ 20 s	2.4–50 s
explicit Pareto front	$<0.1$ s	up to 51 s
Selection, direct argmin	17.3 s	2.5–50.3 s
Filtering and export	18.7 s	10.2–57.1 s
Total, explicit front	41 s	12.8–128 s

Appendix

Table 5. End-to-end cost of CleanGS over 18 scenes on one A100-80GB GPU, run sequentially. Selection runs on the CPU.

Membership, vote	Leak (std)	PSNR (std)	Leak ( $6\times$ )	PSNR ( $6\times$ )	Retained
Center, 2-of-3 (main)	0.206	19.03	0.066	17.83	22.8%
$\tau=0.5$ , 2-of-3	0.194	19.03	0.064	17.85	22.9%
$\tau=0.75$ , 2-of-3	0.169	18.56	0.052	17.70	22.5%
$\tau=0.9$ , 2-of-3	0.148	17.96	0.044	17.32	22.1%
$\tau=0.9$ , 3-of-3	0.085	15.45	0.024	15.26	17.7%
$\tau=0.9$ , 3-of-4	0.109	16.15	0.026	15.72	19.3%

Appendix

Table 6. Footprint-aware mask membership with the same three selected masks over 18 scenes. Std: captured framing;

6\times

: widened framing.

p

compares each row with center lookup (two-sided Wilcoxon).

Dataset	Scene	Frames	Uncleaned			FlashSplat			GaussianCut †			CleanGS
			PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
DL3DV	car_set2	43	18.61	0.704	0.225	17.51	0.681	0.249	17.85	0.687	0.241	18.39	0.715	0.204
	courthouses	52	19.37	0.665	0.239	18.90	0.648	0.252	18.97	0.650	0.249	22.89	0.703	0.202
	fountain	44	20.63	0.602	0.096	14.69	0.487	0.215	18.19	0.566	0.143	19.47	0.632	0.108
	gardens	44	18.79	0.551	0.153	15.84	0.505	0.208	17.45	0.537	0.170	18.64	0.579	0.143
	red_brick_build	39	20.05	0.566	0.237	16.45	0.514	0.293	17.13	0.529	0.277	16.34	0.541	0.278

Appendix

Table 7. Per-scene standard-view photometrics. Results over the same 18 scenes as the main comparison.

Dataset	Scene	Frames	Uncleaned			FlashSplat			GaussianCut †			CleanGS
			PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
DL3DV	car_set2	43	15.31	0.445	0.313	17.50	0.508	0.211	17.78	0.508	0.196	18.45	0.577	0.169
	courthouses	52	12.22	0.324	0.492	17.65	0.480	0.238	17.59	0.480	0.237	20.78	0.568	0.184
	fountain	44	12.64	0.192	0.214	16.01	0.430	0.091	18.60	0.489	0.061	20.14	0.604	0.048
	gardens	44	12.32	0.159	0.219	16.38	0.414	0.096	16.77	0.428	0.089	17.22	0.484	0.088
	red_brick_build	39	19.54	0.353	0.295	16.55	0.315	0.324	17.16	0.331	0.313	16.50	0.337	0.312

Appendix

Table 8. Per-scene widened-field photometrics. Results over the same 18 scenes at

6\times

widened framing.

		Leakage $\downarrow$			Retained $\downarrow$
Dataset	Scene	FS	GCut †	Ours	FS	GCut †	Ours
DL3DV	car_set2	0.023	0.012	0.037	25.8%	32.9%	32.6%
	courthouses	0.023	0.012	0.037	7.2%	8.2%	7.6%
	fountain	0.046	0.000	0.047	1.1%	1.5%	1.8%
	gardens	0.006	0.000	0.082	2.5%	3.7%	5.8%
	red_brick_build	0.019	0.026	0.016	29.2%	41.9%	29.3%

Appendix

Table 9. Per-scene cleanup diagnostics. Widened-field silhouette leakage and retained Gaussian fraction.

	Mean		Median
Method	Std	$6\times$	Std	$6\times$
Uncleaned 3DGS	19.35	12.84	19.43	12.71
FlashSplat	16.44	16.29	16.56	16.47
GaussianCut	15.28	15.10	16.40	16.02
GaussianCut †	17.27	16.71	17.71	17.30
CleanGS	19.03	17.83	19.31	18.33

Appendix

Table 10. Mean and median masked PSNR over 18 scenes. Std: captured framing;

6\times

: widened framing.

Method	$1\times$	$2\times$	$4\times$	$6\times$
Uncleaned 3DGS	19.35	18.72	14.72	12.84
FlashSplat	16.44	16.65	16.59	16.29
GaussianCut †	17.27	17.43	17.15	16.71
CleanGS	19.03	19.31	18.73	17.83

Appendix

Table 11. PSNR across widening factors. Masked PSNR at captured framing and at

2\times

4\times

, and

6\times

widening over the same 18 exported assets. CleanGS ranks first at each tested widened setting.

Method	$1\times$	$2\times$	$4\times$	$6\times$
Uncleaned 3DGS	0.983	0.896	0.813	0.803
FlashSplat	0.141	0.085	0.049	0.042
GaussianCut	0.089	0.047	0.022	0.016
GaussianCut †	0.055	0.038	0.015	0.009
CleanGS	0.206	0.123	0.075	0.066

Appendix

Table 12. Leakage across widening factors. Mean unthresholded silhouette leakage over the same 18 exported assets at captured framing (

1\times

) and at

2\times

4\times

, and

6\times

widening.

	Pareto	Random	$\Delta$	Wins	$p$
PSNR $\uparrow$	19.52	17.96	$+1.56$	15/17	0.0005
SSIM $\uparrow$	0.657	0.620	$+0.037$	15/17	0.0004
LPIPS $\downarrow$	0.166	0.202	$-0.036$	15/17	0.0011
PSNR (6 $\times$ ) $\uparrow$	18.39	17.39	$+1.00$	14/17	0.0150
SSIM (6 $\times$ ) $\uparrow$	0.514	0.481	$+0.033$	14/17	0.0038
LPIPS (6 $\times$ ) $\downarrow$	0.141	0.170	$-0.029$	13/17	0.0129

Appendix

Table 13. Pareto-selected views against random views at the same budget. Both arms use the same vote and evaluator and differ only in the three input views. Five seeds are averaged within each scene, with paired results reported over the 17 scenes with defined outputs for both arms. “Wins” counts scenes favoring selection, and

p

is a two-sided Wilcoxon signed-rank test. Std: captured framing;

6\times

: widened framing.

Filter	Masks	PSNR (std)	PSNR ( $6\times$ )	Leak	Retained	$p$
GaussianCut	selected	15.28	15.10	0.089	17.2%	0.20
	random	15.92	15.72	0.085	18.2%	0.20
FlashSplat	selected	16.44	16.29	0.141	14.9%	0.15
	random	16.91	16.56	0.149	16.0%	0.15
CleanGS	selected	19.52	18.39	0.197	21.7%	0.017
	random	18.72	18.02	0.158	19.3%	0.017

Appendix

Table 14. Selected against random masks for each filter. Each filter receives either the three selected masks or the seed-0 random three-view set. Std: captured framing;

6\times

: widened framing.

p

is a two-sided Wilcoxon test on standard PSNR. The CleanGS rows cover 17 scenes because the random set empties super markets .

Selector	PSNR (std)	PSNR ( $6\times$ )	Leak (std)	Leak ( $6\times$ )	Retained
Pareto (main)	19.03	17.83	0.206	0.066	22.8%
Weighted sum	19.00	17.71	0.193	0.060	21.3%
Greedy	18.96	17.72	0.194	0.060	21.5%
Top-3 reliability	19.07	17.70	0.224	0.077	23.1%
Reliability + farthest point	19.02	18.15	0.166	0.032	20.6%

Appendix

Table 15. Pareto selection against simpler selectors on 18 scenes, with the same reduced pool, two-of-three vote, and evaluator. Std: captured framing;

6\times

: widened framing.

Figure 6. Leakage–PSNR frontier over 18 scenes at captured (left) and $6\times$ widened (right) framing, with leakage on a log scale. Labels give the vote ( $k$ -of- $n$ ) and the footprint threshold $\tau$ . The FlashSplat sweep varies the removal slackness $s$ on a different three-view subset; the hollow square is FlashSplat at $s=0$ with our selected masks. Random-view points use the same filters and average seeds 0–2. Two plots of masked PSNR against silhouette leakage on a log scale, with curves for CleanGS with selected views, CleanGS with random views, and the FlashSplat slackness sweep, and single points for FlashSplat with the selected masks and for GaussianCut.

$\sigma$	Changed	PSNR (std)	PSNR ( $6\times$ )	Leak (std)	Leak ( $6\times$ )
0.05	5/18	19.14	17.76	0.206	0.073
0.10	1/18	19.05	17.84	0.207	0.066
0.15 (main)	–	19.03	17.83	0.206	0.066
0.20	2/18	19.05	17.70	0.205	0.072
0.30	3/18	19.04	17.96	0.206	0.065

Appendix

Table 16. Sensitivity to the occupancy width

\sigma

with minimum aggregation over 18 scenes; all other settings are fixed. “Changed” counts the scenes whose three selected views differ from those at the main

\sigma=0.15

; the other scenes keep the main result. Std: captured framing;

6\times

: widened framing.

Weight zeroed	$\Delta$ PSNR	$\Delta$ PSNR 6×	Median $\Delta_{6\times}$	Leakage $\downarrow$
Main	–	–	–	0.204
without $A$	-0.48	-0.07	+0.01	0.163
without $B$	+0.13	+0.09	+0.00	0.204
without $O$	-0.08	-0.16	-0.05	0.218
without $T$	-0.05	-0.22	-0.07	0.211

Appendix

Table 17. Sensitivity to zeroing individual scalarization weights. Each row sets one Chebyshev weight to zero and renormalizes the remaining weights while retaining the four-objective Pareto front, thereby measuring the influence of each weight. Mean aggregation. Std: captured framing;

6\times

: widened framing.

Weights	$n$	$\Delta$ PSNR	$\Delta$ PSNR 6×	Med. $\Delta_{6\times}$	Leak $\downarrow$	Wins
Main	17	–	–	–	0.204	–
equal, $0.25$ each	17	$+0.02$	$-0.03$	$+0.00$	0.192	5/17
$B$ heavy, $0.40$	17	$+0.14$	$+0.17$	$+0.08$	0.188	9/17
$O$ heavy, $0.40$	17	$-0.02$	$+0.07$	$+0.00$	0.188	5/17
$T$ heavy, $0.40$	17	$+0.13$	$+0.06$	$+0.00$	0.191	7/17
$O,T$ heavy, $0.35$	17	$+0.09$	$+0.48$	$+0.00$	0.180	8/17

Appendix

Table 18. Chebyshev weight sensitivity. The main weights are

(A,B,O,T)=(0.45,0.30,0.15,0.10)

. Each variant changes the relative emphasis placed on the four objectives. Mean aggregation. Std: captured framing;

6\times

: widened framing.

Reliability score	$n$	$\Delta$ PSNR	$\Delta$ PSNR 6×	Med. $\Delta_{6\times}$	Leak $\downarrow$	Wins
Main	17	–	–	–	0.204	–
target $0.05$	17	$-1.69$	$-1.43$	$-0.18$	0.164	5/17
target $0.75$	17	$-0.18$	$-0.20$	$+0.00$	0.215	4/17
linear, target $1.0$	17	$-0.27$	$-0.26$	$+0.00$	0.215	4/17
linear decay	17	$-0.01$	$+0.39$	$+0.00$	0.183	7/17
step decay	17	$+0.19$	$+0.56$	$+0.18$	0.155	9/17

Appendix

Table 19. Per-view reliability score. The main configuration uses a Gaussian-shaped occupancy score with

r_{0}{=}0.30

and

\sigma{=}0.15

; each variant changes the occupancy target or decay profile while leaving the remaining pipeline fixed. Mean aggregation. Std: captured framing;

6\times

: widened framing.

Configuration	$n$	$\Delta$ PSNR	$\Delta$ PSNR 6×	Med. $\Delta_{6\times}$	Leak $\downarrow$	Wins
Main	17	–	–	–	0.204	–
$k{=}2$ , 2-of-2	17	$-2.28$	$-1.52$	$-0.18$	0.115	6/17
$k{=}4$ , 2-of-4	15	$+0.52$	$+0.21$	$+0.04$	0.220	10/15
$k{=}5$ , 2-of-5	13	$+0.57$	$-0.18$	$+0.01$	0.250	9/13
$k{=}4$ , 3-of-4	17	$-1.19$	$-0.66$	$-0.03$	0.144	8/17
$k{=}5$ , 3-of-5	13	$+0.10$	$+0.54$	$+0.13$	0.147	10/13

Appendix

Table 20. Selected-view budget and vote threshold. Each arm is paired against the main

k{=}3

, 2-of-3 configuration on its completed scene set. “Median” is the median per-scene change in widened PSNR, and “Wins” counts scenes where the arm performs better. Exact Pareto construction is all-pairs dominance over

\binom{n}{k}

subsets, which is

3.4\times 10^{4}

k{=}3

but

4.9\times 10^{5}

k{=}4

and

5.5\times 10^{6}

k{=}5

for a capped pool of 60. Selection does not depend on the vote threshold, so the two

k{=}4

rows differ in

n

for run-level resource reasons alone and are not directly comparable with each other. Mean aggregation. Std: captured framing;

6\times

: widened framing.

Configuration	PSNR (std)	PSNR ( $6\times$ )	Leak (std)	Leak ( $6\times$ )	Retained
$k{=}3$ , 1-of-3	20.16	17.47	0.345	0.182	29.1%
$k{=}3$ , 2-of-3 (main)	19.03	17.83	0.206	0.066	22.8%
$k{=}3$ , 3-of-3	16.44	15.88	0.120	0.035	18.6%
$k{=}2$ , 2-of-2	16.87	16.25	0.132	0.036	19.3%
$k{=}4$ , 3-of-4	16.96	16.19	0.155	0.040	20.1%
$k{=}5$ , 3-of-5	19.03	18.05	0.180	0.048	22.4%

Appendix

Table 21. Selected-view budget and vote threshold with minimum aggregation on all 18 scenes. Std: captured framing;

6\times

: widened framing.

Endpoint	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	$\downarrow$
	Standard			Widened			Leak
Vote only	19.86	0.658	0.170	18.25	0.516	0.158	0.221
+ clustering	16.02	0.478	0.281	14.97	0.370	0.303	0.115

Appendix

Table 22. Largest-component geometric refinement. Macro averages over 18 scenes. Both endpoints use the same eroded evaluation masks, isolating the effect of the clustering stage; their absolute values are therefore not directly comparable with the Table 1 endpoint, which is evaluated on the full masks. Std: captured framing;

6\times

: widened framing.

Scene	vote	+cl.	$\Delta$	vote	+cl.	Cl.	top
	Std PSNR			Leakage		Clustering
table	22.29	2.72	-19.57	0.226	0.012	112	16%
tractor	20.56	8.97	-11.58	0.223	0.002	92	9%
redcar	19.33	8.58	-10.75	0.272	0.004	39	13%
Horse	22.32	11.69	-10.63	0.070	0.005	22	47%
car_set2	18.56	8.50	-10.06	0.372	0.004	35	16%

Appendix

Table 23. Per-scene effect of largest-component clustering. Scenes are ordered by standard-PSNR change. “Cl.” is the number of HDBSCAN clusters in the voted set, and “top” is the fraction of voted Gaussians in the largest cluster. Values use the same eroded evaluation masks as Table 22 . Std: captured framing;

6\times

: widened framing.

	Configuration	Std PSNR	Std LPIPS	Zoom PSNR	Leak
A	no erosion, eval on eroded masks	19.86	0.170	18.25	0.221
B	no erosion, eval on full masks	19.03	0.188	17.83	0.206
C	4 px erosion, eval on full masks	18.37	0.201	17.63	0.196

Appendix

Table 24. Effect of mask erosion. Rows B and C isolate 4 px supervision-mask erosion using the same full evaluation masks. Row A shows the corresponding no-erosion result when evaluation masks are also eroded. Std: captured framing;

6\times

: widened framing.

Panel of Figure 7 — Figure 7. Target-adjacent radiance and silhouette leakage. The roses scene at captured framing (top) and $6\times$ widened framing (bottom). GaussianCut removes the illuminated tabletop beneath the bouquet and darkens the glass vase, while CleanGS preserves the vase, water, and nearby highlight. CleanGS is $1.61$ dB better in widened PSNR on the displayed frame, although the retained tabletop region lies outside the binary target mask and therefore contributes to leakage. Ground truth, uncleaned, GaussianCut and CleanGS renderings of a rose bouquet at ordinary and widened fields of view.

Panel of Figure 8 — Figure 8. Effect of additional GaussianCut supervision. GaussianCut is shown with the same three selected masks used by CleanGS and with all available capture masks. Denser supervision removes additional residual structure and produces tighter silhouettes across these examples. The all-mask setting includes masks at scored cameras and corresponds to the daggered GaussianCut reference in the main text. Four rows comparing GaussianCut given three masks against GaussianCut given every capture mask, at both fields of view.

Panel of Figure 9 — Figure 9. Qualitative results across DL3DV and NeRFBusters scenes. All eight DL3DV scenes and the NeRFBusters car scene are shown at captured and $6\times$ widened framing. Widening preserves camera pose and output resolution while reducing focal length, exposing additional content retained in each exported asset. GaussianCut is shown using its all-mask $\text{GaussianCut}^{\dagger}$ configuration. A nine-row grid of renderings for nine scenes, with four columns at captured framing and four at widened field of view.

Panel of Figure 10 — Figure 10. Qualitative results across NeRFBusters and Tanks and Temples scenes. Six additional NeRFBusters scenes and all three Tanks and Temples scenes are shown at captured and $6\times$ widened framing. The widened views preserve camera pose and output resolution while exposing Gaussian content retained outside the original framing. GaussianCut is shown using its all-mask $\text{GaussianCut}^{\dagger}$ configuration. A nine-row grid of renderings for nine scenes, with four columns at captured framing and four at widened field of view.

3D Gaussian Splatting

3D Gaussian Splatting represents scenes as optimizable 3D Gaussian primitives for fast, photorealistic rendering from new viewpoints.

See all 1,019 papers

Sep 25, 2026

ClearGS: Reliability-Aware Gaussian Splatting from Handheld Videos

We present ClearGS for 3D Gaussian Splatting (3DGS) from handheld videos with uneven viewpoint coverage and mixed frame quality. Rather than selecting frames with binary decisions, ClearGS uses Reliability-aware View Allocation (RVA) to assign graded raw-supervision weights based on appearance reliability, degradation risk, and geometric utility, while weakly reactivating useful suppressed frames to maintain trajectory coverage. Since weighting cannot restore details lost to blur or distortion, ClearGS further introduces Render-Guided In-Video Restoration (RIVR). The current 3DGS render provides a pose-aligned structural candidate, a frozen no-reference restoration expert restores the corresponding raw video observation without any clean reference image, and no-reference perceptual scores select among the render, restored observation, and high-frequency fused candidate. ClearGS then applies Full-Trajectory Repair Consolidation to revisit accepted repairs and preserve details introduced early. On GS2E and GSOTM, ClearGS achieves state-of-the-art overall performance, with consistent CLIP-IQA and MUSIQ gains and LPIPS reductions in most degradation settings, without paired sharp supervision or matched clean references.

Xuanzhi Liu, Xinyi Wu, Hang Pan +4

Shenzhen University of Advanced Technology · HONOR · University of New South Wales +1

3D Gaussian Splatting 3D Reconstruction Multi-View 3D Reconstruction

Ask PDF

Apr 14, 2026

PDF-GS: Progressive Distractor Filtering for Robust 3D Gaussian Splatting

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.

Kangmin Seo, MinKyu Lee, Tae-Young Kim +3

Sungkyunkwan University

3D Gaussian Splatting 3D Reconstruction Multi-View 3D Reconstruction

Ask PDF

Jun 26, 2026

StructSplat: Generalizable 3D Gaussian Splatting from Uncalibrated Sparse Views

We present StructSplat, a feed-forward and generalizable 3D Gaussian reconstruction framework that operates directly on uncalibrated images without requiring camera parameters. Existing methods either rely on per-scene optimization or assume known camera poses, and often entangle geometry and appearance within a unified backbone, limiting reconstruction fidelity and generalization. Our key idea is to adopt a structured representation that organizes geometry, semantic, and texture cues with explicit roles in the reconstruction process. Specifically, we introduce a pixel-aligned feature injection mechanism to enable accurate texture modeling from 2D observations, incorporate semantic-aware priors to improve global consistency, and design a camera alignment strategy to prevent information leakage and improve generalization. Experiments show that our method significantly outperforms prior approaches on challenging benchmarks. On DL3DV, our method achieves 28.045 PSNR, surpassing AnySplat (22.377) by +5.67 dB. In cross-dataset evaluation, our method achieves +1.94 dB over AnySplat on ACID and +1.72 dB on RealEstate10K. Project page: https://structsplat.github.io Code: https://github.com/J-C-Zhao/StructSplat

Jia-Chen Zhao, Beiqi Chen, Xinyang Chen +2

Harbin Institute of Technology (Shenzhen) · Great Bay University · Guangzhou CloudButterfly Technology Co., Ltd.

3D Gaussian Splatting 3D Reconstruction Sparse-View 3D Reconstruction Feed-Forward 3D Reconstruction

Ask PDF

Abstract

3D Gaussian Splatting

Explore similar work

ClearGS: Reliability-Aware Gaussian Splatting from Handheld Videos

PDF-GS: Progressive Distractor Filtering for Robust 3D Gaussian Splatting

StructSplat: Generalizable 3D Gaussian Splatting from Uncalibrated Sparse Views