Efficient Confidence-Aware Densification for 3DGS

Manuel Hofer; Thomas Köhler; Felix Windisch; Bernhard Kerbl; Markus Steinberger; Lukas Radl

cs.CVOct 9, 2026

Efficient Confidence-Aware Densification for 3DGS

Authors: Manuel Hofer, Thomas Köhler, Felix Windisch, Bernhard Kerbl, Markus Steinberger, Lukas Radl

Organizations: Graz University of Technology, Austria · University of Copenhagen, Denmark · Huawei Technologies, Austria

Abstract

3DGS-MCMC has replaced the gradient-driven density control of 3D Gaussian Splatting with opacity-based sampling and relocation, and is now the de facto standard for prior-free densification under a fixed primitive budget. Because opacity stays high where the scene is already well reconstructed, this sampling sends most new primitives to the foreground and leaves the background under-reconstructed. At low budgets, prior-based methods such as ConeGS and EDGS outperform it, at the cost of a pre-trained radiance field or dense image matches and a separate initialization stage. We argue for a comeback of error-driven densification within the fixed-budget MCMC framework. Our confidence-aware densification keeps the fixed budget and gradual growth of 3DGS-MCMC, but places primitives by the image error instead of the opacity. New primitives are drawn mostly where screen-space gradients are large. We favor primitives with a low learned confidence, which marks regions the reconstruction still explains poorly, and relocate the primitives that contribute least to the rendered images to these places. Averaged over 24 scenes, our method outperforms 3DGS-MCMC in PSNR, SSIM and LPIPS at every budget from 100k to 2M primitives. At 100k, without a prior or an initialization stage, it recovers 0.84dB of the 1.05dB PSNR by which ConeGS leads 3DGS-MCMC, and from 500k on, it has the highest PSNR and SSIM of all fixed-budget methods we compare with, including ConeGS and EDGS. While inherently prior-free, our densification can also exploit priors: from the ConeGS or EDGS initialization, it matches ConeGS in PSNR and SSIM at 100k and is best in all three metrics at 500k.

Figures & tables

Panel of Figure 1 — Figure 1: Biased sampling in 3DGS-MCMC. We render the positions of all source primitives drawn during densification on the Mip-NeRF 360 Garden scene as a point cloud, for our method (left) and 3DGS-MCMC (right). 3DGS-MCMC draws its sources in clusters around the central object and leaves large parts of the background with few sources, while ours cover the background evenly.

Panel of Figure 2 — Figure 2: 3DGS-MCMC Foreground Bias : Number of primitives blended into each pixel for the Mip-NeRF 360 Garden scene (brighter means more) after training with budgets of $\mathrm{100k}$ (top) and $\mathrm{500k}$ (bottom). In 3DGS-MCMC, most primitives end up on the foreground objects; ours are spread more evenly over foreground and background. The labels give the mean PSNR of each model over the test views of this scene.

Panel of Figure 3 — Figure 3: Convergence of Confidence : Rendered confidence maps [ 29 ] of Bonsai (top) and Counter (bottom) at iterations 500, 5k and 30k, next to the ground truth (GT); red marks confidences below 1 and blue above. At iteration 500, the maps are nearly white because all confidences start at 1. We therefore start by sampling by absolute gradients alone and blend in the confidence over the first 5000 iterations (Section 4.1 ).

Panel of Figure 4 — Figure 4: Sampling Distributions : Top: ground truth and a rendering of our trained Bicycle model. Bottom: our confidence-weighted absolute gradients (Equation ( 5 )), the SSIM term and the opacity-proportional sampling of 3DGS-MCMC in image space (brighter is more likely). Opacity-proportional sampling misses the high-error regions in the red box, which both our distributions favor.

Figure 5: Relocation Schedule : Cumulative number of relocated primitives during training at each budget, for our relocation every

400

iterations (left) and for the opacity-threshold relocation of 3DGS-MCMC (right), measured for the Bonsai scene [ 2 ] . We relocate more primitives than 3DGS-MCMC at

\mathrm{100k}

and

\mathrm{500k}

, but fewer at

\mathrm{1M}

and

\mathrm{2M}

Figure 6: Quantitative Evaluation : Average PSNR, SSIM and LPIPS over the 24 scenes of four datasets [ 2 , 16 , 18 , 13 ] at four primitive budgets. Our method and its prior-free baseline 3DGS-MCMC [ 15 ] start from SfM initialization.

		$\Delta$ Ours $-$ 3DGS-MCMC
Metric	Region	$\mathrm{100k}$	$\mathrm{500k}$	$\mathrm{1M}$	$\mathrm{2M}$
PSNR $\uparrow$	Background	$+0.69$	$+1.20$	$+1.03$	$+0.94$
PSNR $\uparrow$	Foreground	$+0.41$	$+0.23$	$+0.15$	$-0.03$
SSIM $\uparrow$	Background	$+0.050$	$+0.070$	$+0.061$	$+0.046$
SSIM $\uparrow$	Foreground	$+0.021$	$+0.019$	$+0.013$	$+0.007$
LPIPS $\downarrow$	Background	$-0.044$	$-0.083$	$-0.083$	$-0.072$

Table 1: Foreground vs. background performance. Mean metric differences between our method and 3DGS-MCMC across outdoor Mip-NeRF 360 scenes (SfM initialization), split by image region. Background is defined as the farthest 25% of test pixels via COLMAP-aligned [ 31 , 32 ] Depth Anything V2 [ 38 ] . Positive PSNR and SSIM and negative LPIPS differences favor our method. Gains are largest in the distant background.

		Mip-NeRF 360			Tanks & Temples			Deep Blending			DL3DV
Budget	Method	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS
$\mathrm{100k}$	Ours	26.32	0.748	0.336	23.47	0.815	0.251	29.55	0.899	0.283	26.74	0.842	0.244
	ConeGS	26.38	0.750	0.326	23.36	0.811	0.240	29.81	0.896	0.286	27.11	0.850	0.233
	EDGS	25.75	0.741	0.333	22.69	0.802	0.254	28.99	0.890	0.295	26.10	0.829	0.256
	Taming 3DGS	25.57	0.715	0.367	22.73	0.785	0.286	28.94	0.884	0.310	26.18	0.814	0.278
	3DGS-MCMC	25.75	0.726	0.354	22.24	0.782	0.282	28.86	0.887	0.308	25.72	0.811	0.279

Table 2: Per-Dataset Results : Means over the scenes of each dataset for the fixed-budget methods at budgets of

\mathrm{100k}

and

\mathrm{500k}

, with SfM initialization for our method. The values are the dataset averages of supplementary Tables S.1 and S.2. In each block, red, orange and yellow mark the best, second and third value of each column; equal values share a color.

Panel of Figure 7 — Figure 7: Qualitative Comparison at $\mathrm{500k}$ Primitives : Test views of seven scenes from Mip-NeRF 360 (Garden, Bicycle, Room, Counter), Tanks & Temples (Train), Deep Blending (Playroom) and DL3DV (Car). The columns show the ground truth, our method and 3DGS-MCMC [ 15 ] , both with SfM initialization, and the prior-based methods ConeGS [ 1 ] and EDGS [ 17 ] , all at a budget of $\mathrm{500k}$ . The label under each image is the per-scene test PSNR of the method, averaged over all test views of the scene (supplementary Table S.2); it is not the PSNR of the view shown. Each inset magnifies the marked background region by $3\times$ , illustrating how much background detail each method retains at this low budget.

Method	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	#G
	Mip-NeRF 360 Outdoor			Mip-NeRF 360 Indoor			T&T & DB			DL3DV
Ours	25.04	0.754	0.189	31.92	0.933	0.165	27.33	0.888	0.179	29.13	0.906	0.155
3DGS-MCMC	24.94	0.744	0.220	31.61	0.931	0.177	26.41	0.882	0.213	28.90	0.901	0.168	1.94M
3DGS	24.71	0.730	0.236	30.78	0.923	0.188	26.69	0.880	0.203	27.85	0.884	0.187
Ours	25.07	0.754	0.189	31.81	0.932	0.168	27.25	0.887	0.182	28.84	0.898	0.168
3DGS-MCMC	24.93	0.743	0.221	31.45	0.929	0.181	26.63	0.881	0.214	28.54	0.894	0.180	1.55M

Table 3: Equal-Budget Comparison : In each block, we compare our method and 3DGS-MCMC with one variable-budget method at its per-scene primitive count: 3DGS, 3DGS + GOF, DashGaussian [ 14 , 41 , 3 ] and the simplification methods Mini-Splatting2 and GaussianSpa [ 7 , 43 ] . #G is the mean primitive budget of each block. Columns average the Mip-NeRF 360 outdoor and indoor scenes, the Tanks & Temples (T&T) and Deep Blending (DB) scenes, and the DL3DV scenes. Red shading marks the best value in each block.

	$\mathrm{100k}$			$\mathrm{500k}$
Configuration	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS
Ours (full)	26.32	0.748	0.336	27.84	0.819	0.231
w/ MCMC noise	26.17	0.741	0.341	27.62	0.808	0.244
w/o color initialization	26.22	0.744	0.341	27.68	0.816	0.236
w/o position sampling	25.87	0.741	0.342	26.96	0.788	0.262
w/ MCMC sampling	26.14	0.744	0.338	27.74	0.814	0.239

Table 4: Ablation Study : Means over the 9 Mip-NeRF 360 scenes [ 2 ] at budgets of

\mathrm{100k}

and

\mathrm{500k}

. In each row of the middle blocks, we remove, replace or add one component of our method, whose default is

\beta_{\text{conf}}=0.075

. In w/o position sampling , we remove all positional randomness. The bottom block shows 3DGS-MCMC with and without its continuous position noise, and with either our relocation ( w/ our relocation ) or our sampling distribution ( w/ our sampling ) in place of its opacity rule. Our full configuration is on par or best in every column. Shading runs from red for the best to yellow for the worst value in each column.

Distribution	w/ conf.	w/ 20% SSIM	PSNR	SSIM	LPIPS
Abs. gradients	✓	✓	27.84	0.819	0.231
	✓	-	27.70	0.818	0.231
	-	✓	27.74	0.816	0.230
	-	-	27.61	0.816	0.230
SSIM	✓	-	27.72	0.818	0.231
SSIM	-	-	27.70	0.812	0.241

Table 5: Sampling Ablation Study : Distributions for drawing source primitives on the 9 Mip-NeRF 360 scenes [ 2 ] at a budget of

\mathrm{500k}

. Each block starts from absolute gradients, the SSIM term or the confidence alone. With the first check mark, we weight this distribution by the confidence factor, and with the second, we draw 20% of the source primitives from the SSIM term. The first row is our full distribution. Shading as in Table 4 .

fixed-budget (@500k)
Method	Mem. [GB]	Init. [min]	Train [min]
EDGS	7.3	1.8	7.2
ConeGS	4.9	2.2	11.8
Ours	4.6	–	2.8
variable-budget
Mini-Splatting2 $\dagger$	8.2	–	1.9

Table 6: Memory and Runtime during optimization, as means over the 9 Mip-NeRF 360 scenes on a single NVIDIA RTX 5090. Peak memory (Mem.) includes initialization (Init.), i.e. the dense matching of EDGS and the iNGP pre-training of ConeGS. Train is the optimization time w/o initialization. Bold marks the best value in each block.

Panel of Figure 8 — Figure 8: Effect of the SSIM Term : Our method on Garden and Bicycle with and without drawing 20% of the source primitives from the SSIM term (Section 4.1 ). Including the SSIM term leads to more detailed background reconstruction.

		100k			500k
Method	Init.	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS
Ours	SfM	26.54	0.809	0.282	28.05	0.863	0.202
Ours	EDGS	26.72	0.813	0.277	28.20	0.867	0.195
EDGS	EDGS	25.93	0.799	0.288	27.51	0.855	0.202
Ours	iNGP	26.79	0.813	0.278	28.22	0.866	0.197
ConeGS	iNGP	26.75	0.813	0.273	27.87	0.859	0.199

Table 7: Matched Initialization : Our method started from the dense initialization of EDGS [ 17 ] or ConeGS [ 1 ] , compared to the respective prior-based method and to our method with SfM initialization; means over all 24 scenes. Bold marks the better value of each pair of rows with shared initialization.

Figure 9: Growth Beyond the Budget : PSNR, SSIM and LPIPS of our method when it may grow to

\lambda_{\text{grow}}

times the budget during training, as means over the 13 Mip-NeRF 360, Tanks & Temples and Deep Blending scenes of Table 3 . In each row, the budget is the per-scene primitive count of the simplification method shown dashed, Mini-Splatting2 (top) or GaussianSpa (bottom), so

\lambda_{\text{grow}}=1

repeats the corresponding block of Table 3 . At

\lambda_{\text{grow}}

of 1.5 or 2, our LPIPS is on par with both methods.

		Ours			ConeGS			EDGS			Taming 3DGS			3DGS-MCMC
	Scene	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS
Mip-NeRF360	Bicycle	23.82	0.636	0.396	24.08	0.655	0.378	23.71	0.636	0.392	23.30	0.588	0.436	23.34	0.603	0.421
	Flowers	20.48	0.503	0.471	20.59	0.509	0.465	20.12	0.496	0.467	19.98	0.462	0.507	20.15	0.474	0.491
	Garden	24.89	0.703	0.357	24.68	0.709	0.336	24.45	0.705	0.341	24.20	0.665	0.383	24.37	0.682	0.366
	Stump	25.68	0.706	0.354	25.79	0.709	0.349	25.13	0.673	0.377	24.19	0.626	0.423	24.78	0.657	0.394
	Treehill	22.60	0.577	0.474	22.79	0.569	0.476	22.16	0.591	0.449	22.28	0.560	0.494	22.28	0.563	0.486

Table S.1 : Per-scene results at a budget of

\mathrm{100k}

primitives, SfM initialization.

		Ours			ConeGS			EDGS			Taming 3DGS			3DGS-MCMC
	Scene	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS
Mip-NeRF360	Bicycle	25.19	0.753	0.261	25.26	0.758	0.242	24.97	0.752	0.251	24.53	0.685	0.339	24.74	0.715	0.307
	Flowers	21.91	0.610	0.360	21.46	0.604	0.347	21.08	0.598	0.351	21.14	0.560	0.406	21.41	0.570	0.393
	Garden	27.13	0.842	0.171	26.86	0.837	0.157	26.45	0.831	0.164	26.54	0.812	0.206	26.39	0.816	0.195
	Stump	27.15	0.797	0.220	26.96	0.794	0.214	26.52	0.769	0.238	26.07	0.731	0.299	26.53	0.754	0.273
	Treehill	23.09	0.644	0.360	22.79	0.608	0.372	22.57	0.653	0.319	22.98	0.615	0.406	22.80	0.622	0.394

Table S.2 : Per-scene results at a budget of

\mathrm{500k}

primitives, SfM initialization.

		Ours			ConeGS			EDGS			Taming 3DGS			3DGS-MCMC
	Scene	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS
Mip-NeRF360	Bicycle	25.44	0.778	0.214	25.45	0.779	0.196	25.27	0.778	0.201	24.96	0.722	0.291	25.10	0.749	0.259
	Flowers	22.07	0.633	0.313	21.52	0.624	0.303	21.32	0.624	0.306	21.47	0.586	0.373	21.75	0.597	0.358
	Garden	27.68	0.867	0.120	27.45	0.861	0.115	27.08	0.856	0.121	27.19	0.843	0.153	26.86	0.846	0.147
	Stump	27.30	0.808	0.189	27.23	0.802	0.185	26.79	0.784	0.204	26.36	0.753	0.262	26.78	0.776	0.235
	Treehill	22.96	0.654	0.317	22.47	0.604	0.334	22.61	0.659	0.282	22.94	0.629	0.372	22.88	0.639	0.352

Table S.3 : Per-scene results at a budget of

\mathrm{1M}

primitives, SfM initialization.

		Ours			ConeGS			EDGS			Taming 3DGS			3DGS-MCMC
	Scene	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS
Mip-NeRF360	Bicycle	25.61	0.791	0.181	25.49	0.783	0.174	25.41	0.791	0.170	25.21	0.750	0.246	25.27	0.768	0.224
	Flowers	21.92	0.642	0.274	21.41	0.628	0.276	21.48	0.639	0.272	21.65	0.602	0.349	21.91	0.614	0.328
	Garden	27.98	0.879	0.096	27.64	0.870	0.098	27.42	0.869	0.100	27.56	0.860	0.122	27.44	0.864	0.118
	Stump	27.24	0.808	0.174	27.01	0.795	0.178	26.91	0.790	0.185	26.55	0.766	0.231	26.99	0.789	0.208
	Treehill	22.78	0.653	0.283	22.21	0.592	0.302	22.52	0.656	0.254	23.09	0.639	0.339	22.81	0.647	0.316

Table S.4 : Per-scene results at a budget of

\mathrm{2M}

primitives, SfM initialization.

		Ours			3DGS-MCMC			Ours + 3DGS-MCMC noise
	Scene	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS
$\mathrm{100k}$	Bicycle	23.14	0.587	0.423	21.28	0.458	0.537	20.87	0.422	0.592
	Flowers	19.93	0.468	0.490	18.18	0.363	0.568	19.67	0.434	0.519
	Garden	24.43	0.687	0.351	23.29	0.601	0.415	23.52	0.629	0.388
	Stump	23.19	0.580	0.438	19.91	0.388	0.578	19.44	0.371	0.643
	Treehill	21.80	0.538	0.501	20.49	0.452	0.575	21.89	0.512	0.532

Table S.5 : Per-scene results on Mip-NeRF 360 with random initialization, at a budget of

\mathrm{100k}

primitives (top) and at the primitive counts of the original 3DGS. In each row, red, orange and yellow mark the best, second and third method; equal values share a color.

		Ours (EDGS init.)			EDGS
	Scene	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS
Mip-NeRF360	Bicycle	23.85	0.637	0.397	23.71	0.636	0.392
	Flowers	20.61	0.508	0.471	20.12	0.496	0.467
	Garden	24.81	0.712	0.344	24.45	0.705	0.341
	Stump	25.78	0.709	0.353	25.13	0.673	0.377
	Treehill	22.69	0.583	0.468	22.16	0.591	0.449

Table S.6 : Per-scene results at a budget of

\mathrm{100k}

primitives; our method initialized from the EDGS point cloud.

		Ours (EDGS init.)			EDGS
	Scene	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS
Mip-NeRF360	Bicycle	25.16	0.759	0.256	24.97	0.752	0.251
	Flowers	21.82	0.613	0.353	21.08	0.598	0.351
	Garden	27.18	0.847	0.162	26.45	0.831	0.164
	Stump	27.41	0.806	0.214	26.52	0.769	0.238
	Treehill	23.04	0.648	0.346	22.57	0.653	0.319

Table S.7 : Per-scene results at a budget of

\mathrm{500k}

primitives; our method initialized from the EDGS point cloud.

		Ours (iNGP init.)			ConeGS
	Scene	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS
Mip-NeRF360	Bicycle	23.97	0.637	0.399	24.08	0.655	0.378
	Flowers	20.68	0.511	0.471	20.59	0.509	0.465
	Garden	24.85	0.711	0.346	24.68	0.709	0.336
	Stump	25.87	0.712	0.351	25.79	0.709	0.349
	Treehill	22.76	0.572	0.479	22.79	0.569	0.476

Table S.8 : Per-scene results at a budget of

\mathrm{100k}

primitives; our method initialized from the ConeGS (iNGP) point cloud. Both methods train for 30k iterations.

		Ours (iNGP init.)			ConeGS
	Scene	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS
Mip-NeRF360	Bicycle	25.30	0.760	0.260	25.26	0.758	0.242
	Flowers	21.97	0.622	0.349	21.46	0.604	0.347
	Garden	27.13	0.843	0.163	26.86	0.837	0.157
	Stump	27.43	0.808	0.212	26.96	0.794	0.214
	Treehill	23.22	0.639	0.367	22.79	0.608	0.372

Table S.9 : Per-scene results at a budget of

\mathrm{500k}

primitives; our method initialized from the ConeGS (iNGP) point cloud. Both methods train for 30k iterations.

	$\mathrm{100k}$			$\mathrm{500k}$			$\mathrm{1M}$			$\mathrm{2M}$
Share	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS
5%	26.24	0.747	0.333	27.72	0.814	0.231	27.99	0.826	0.203	28.09	0.831	0.186
7.5%	26.28	0.748	0.334	27.78	0.817	0.230	28.04	0.829	0.201	28.14	0.834	0.183
10% (ours)	26.32	0.748	0.336	27.84	0.819	0.231	28.10	0.831	0.202	28.21	0.835	0.183
12.5%	26.31	0.748	0.337	27.85	0.820	0.231	28.13	0.832	0.202	28.24	0.838	0.182
15%	26.30	0.746	0.339	27.86	0.820	0.233	28.15	0.833	0.202	28.27	0.839	0.182

Table S.10: Relocation Share : Our method with different shares of primitives relocated per relocation step before the share falls off from iteration 10k (Section 4.2 ; 10% in all other experiments), as means over the 9 Mip-NeRF 360 scenes [ 2 ] . Bold marks the best value in each column.

		$\Delta$ Ours $-$ 3DGS-MCMC
Metric	Region	$\mathrm{100k}$	$\mathrm{500k}$	$\mathrm{1M}$	$\mathrm{2M}$
PSNR $\uparrow$	Background	$+1.34$	$+1.38$	$+1.20$	$+1.04$
PSNR $\uparrow$	Foreground	$+0.45$	$+0.36$	$+0.25$	$+0.18$
SSIM $\uparrow$	Background	$+0.015$	$+0.012$	$+0.009$	$+0.007$
SSIM $\uparrow$	Foreground	$+0.010$	$+0.006$	$+0.004$	$+0.002$
LPIPS $\downarrow$	Background	$-0.024$	$-0.025$	$-0.022$	$-0.018$

Table S.11 : Background vs. Foreground, Indoor : the same comparison as Table 1 on the indoor Mip-NeRF 360 scenes. The background of each scene is defined as the farthest

25\%

of all its test pixels by Depth Anything V2 [ 38 ] depth of the ground-truth test images, aligned per image to the observed COLMAP [ 31 , 32 ] points. We also exclude a

\pm 5\%

quantile band around the cut and eroding both regions by

3

pixels. We report the differences of the means over room, counter, kitchen and bonsai.

Panel of Figure 28 — Figure S.1 : Comparison with variable-budget methods. For each scene, the top row shows the baseline and the bottom row ours trained with the primitive budget of that baseline.

		$\mathrm{100k}$			$\mathrm{500k}$			$\mathrm{1M}$			$\mathrm{2M}$
Dataset	Method	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS
MipNeRF 360	Ours	26.32	0.747	0.336	27.84	0.819	0.231	28.10	0.831	0.202	28.21	0.835	0.183
MipNeRF 360	Revising	24.24	0.691	0.379	26.57	0.768	0.273	26.93	0.788	0.238	27.13	0.800	0.213
Tanks&Temples	Ours	23.47	0.815	0.251	24.30	0.860	0.164	24.62	0.868	0.142	24.59	0.871	0.129
Tanks&Temples	Revising	20.84	0.760	0.298	23.06	0.832	0.193	23.47	0.844	0.168	23.78	0.852	0.153
Deep Blending	Ours	29.55	0.899	0.283	30.06	0.912	0.241	30.03	0.911	0.230	30.16	0.911	0.223

Table S.12: Comparison with Revising Densification : Our method against Revising Densification at Gaussian budgets of 100k, 500k, 1M and 2M, as means over the 9 Mip-NeRF 360 scenes, the 2 Tanks&Temples scenes (Train, Truck) and the 2 Deep Blending scenes (Playroom, DrJohnson). Bold marks the best value in each column pair.

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

Oct 19, 2026

Counting Contributions, Not Views: Occlusion-Aware Densification for 3D Gaussian Splatting

3D Gaussian Splatting grows a dense scene from a sparse initialization through adaptive density control (ADC), driven by per-Gaussian statistics gathered during rendering. These statistics pass a frustum-based visibility filter blind to occlusion: fully occluded Gaussians still count as observed, diluting the averaged gradient signal and suppressing densification precisely where geometry is occluded. We correct this with an occlusion-aware filter that restricts densification statistics to views in which a Gaussian actually contributes, and replace the criterion's raw view-count normalizer with an average-contribution one. Integrated into a StopThePop-based framework, the two together improve every image-quality metric over the occlusion-blind baseline on occlusion-heavy scenes (up to 6% lower LPIPS), at less than half the Gaussians a view-count normalizer requires. Under the longer schedules these large scenes require, the count stays bounded where most criteria diverge, and on smaller benchmarks our method stays competitive at a fraction of the competitors' count.

Marcel Rogge, René Schuster, Didier Stricker

University Kaiserslautern-Landau – RPTU · German Research Center for Artificial Intelligence – DFKI

3D Gaussian Splatting 3DGS Training Adaptive Gaussian Densification

Ask PDF

Feb 23, 2026

RAP: Fast Feedforward Rendering-Free Attribute-Guided Primitive Importance Score Prediction for Efficient 3D Gaussian Splatting Processing

3D Gaussian Splatting (3DGS) has emerged as a leading technology for high-quality 3D scene reconstruction. However, the iterative refinement and densification process leads to the generation of a large number of primitives, each contributing to the reconstruction to a substantially different extent. Estimating primitive importance is thus crucial, both for removing redundancy during reconstruction and for enabling efficient compression and transmission. Existing methods typically rely on rendering-based analyses, where each primitive is evaluated through its contribution across multiple camera viewpoints. However, such methods are sensitive to the number and selection of views, rely on specialized differentiable rasterizers, and have long calculation times that grow linearly with view count, making them difficult to integrate as plug-and-play modules and limiting scalability and generalization. To address these issues, we propose RAP, a fast feedforward rendering-free attribute-guided method for efficient importance score prediction in 3DGS. RAP infers primitive significance directly from intrinsic Gaussian attributes and local neighborhood statistics, avoiding rendering-based or visibility-dependent computations. A compact MLP predicts per-primitive importance scores using rendering loss, pruning-aware loss, and significance distribution regularization. After training on a small set of scenes, RAP generalizes effectively to unseen data and can be seamlessly integrated into reconstruction, compression, and transmission pipelines. Our code is publicly available at https://github.com/yyyykf/RAP.

Kaifa Yang, Qi Yang, Yiling Xu +1

Shanghai Jiao Tong University · University of Missouri–Kansas City

3D Gaussian Splatting 3D Gaussian Pruning 3DGS Compression

Ask PDF

Jun 8, 2026

REFINE: Super-efficient 3D Gaussian Splatting Pruning via Rendering-Free Primitive Importance

Existing pruning methods for 3D Gaussian splatting (3DGS) suffer from either severe quality degradation or prohibitive computational overhead. In this paper, we propose REFINE, a highly accelerated 3DGS pruning framework centered on a novel rendering-free primitive importance metric. Our approach leverages an analytically approximated, rendering-aware Hessian field to quantify the expected perceptual error induced by the removal of individual primitives. By modeling the joint modulation of visibility, projection geometry and the content adaptive hyperparameter, we entirely bypass costly forward rendering passes and derive an anisotropic perceptual weight field that serves as a high-fidelity proxy for primitive importance. Extensive experiments across multiple benchmark datasets demonstrate that REFINE maintains highly competitive rendering quality while achieving a

3,000\times

reduction in pruning-related computational complexity, translating to a practical

\sim 20\times

speedup in device latency compared to state-of-the-art pruning methods.

Zhang Chen, Shuai Wan, Mengting Yu +2

School of Electronics and Information, Northwestern Polytechnical University · School of Telecommunication Engineering, Xidian University · Department of Computer Science, City University of Hong Kong

3D Gaussian Splatting 3D Gaussian Pruning 3DGS Acceleration

Ask PDF

Abstract

3D Gaussian Splatting

Explore similar work

Counting Contributions, Not Views: Occlusion-Aware Densification for 3D Gaussian Splatting

RAP: Fast Feedforward Rendering-Free Attribute-Guided Primitive Importance Score Prediction for Efficient 3D Gaussian Splatting Processing

REFINE: Super-efficient 3D Gaussian Splatting Pruning via Rendering-Free Primitive Importance