Unified Loss-Aware Density Control for 3D Gaussian Splatting

Tae-Young Kim; Jiwoo Chung; Jae-Pil Heo

cs.CVNov 1, 2026

Unified Loss-Aware Density Control for 3D Gaussian Splatting

Authors: Tae-Young Kim, Jiwoo Chung, Jae-Pil Heo

Organizations: Sungkyunkwan University

Abstract

The efficiency and fidelity of 3D Gaussian Splatting critically depend on adaptive density control, which grows and suppresses Gaussian primitives throughout optimization. However, the standard pipeline still relies on heuristic view-space gradients, scale thresholds, and global opacity resets, rather than explicitly assessing the loss-reducing effect of each split, clone, or opacity update. This mismatch can allocate primitives to already-sufficient regions, miss under-reconstructed structures, and destabilize training through blanket opacity changes. In this work, we recast adaptive density control as a collection of local loss-reduction decisions. We derive efficient Taylor-based per-Gaussian surrogates that predict the objective change caused by three primitive-level operations: splitting, cloning, and opacity decay. This yields a unified loss-aware framework with three components. First, we extend second-order splitting to a normalized joint position-scale space, enabling children to adapt not only their locations but also their anisotropic extents. Second, we formulate duplicate cloning as a surrogate minimization problem, selecting only Gaussians whose duplicated contribution is predicted to improve reconstruction and assigning each child a principled opacity. Third, we replace global opacity reset with selective soft opacity decay that weakens harmful primitives while preserving useful ones. Across standard 3DGS benchmarks, our method produces more targeted primitive allocation and achieves improved rendering quality with comparable or smaller Gaussian budgets compared to baselines.

Figures & tables

Figure 1: Intuition of our loss-aware adaptive density control (ADC). We visualize the effect of each density-control operation from the same intermediate state at the 3,000 ${}^{\text{th}}$ training iteration. (a) Splitting: we compare the position-only split of [ 32 ] with our position-scale split, visualizing splitting Gaussians as ellipsoids. While the position-only split produces misplaced offspring around the thin bicycle handle, our position-scale split places offspring more appropriately and forms the handle structure more faithfully. (b) Cloning: we compare standard 3DGS and ours by individually cloning candidate Gaussians and measuring the resulting loss change; green / red indicate candidates whose duplication decreases / increases the loss. Our cloning selects mostly loss-reducing candidates and recovers missed structures, whereas heuristic cloning introduces loss-increasing candidates. (c) Opacity decay: we compare the opacity reset of standard 3DGS with our selective opacity decay. Opacity reset makes all Gaussian primitives nearly transparent, leading to unstable training. In contrast, our method weakens only Gaussians predicted to hinder optimization. (d) Component comparison: standard 3DGS relies on heuristics. SteepGS improves splitting with a position-only loss-aware rule but keeps cloning and opacity update heuristic. Our method applies loss-aware decisions to position-scale splitting, cloning, and opacity decay.

Figure 2: Overview of our framework. (a) During training, we accumulate per-Gaussian surrogate statistics, including the joint splitting matrix and surrogate terms for cloning and opacity decay. At each densification interval, these surrogates determine whether each Gaussian is split, cloned, or assigned opacity decay. (b) For each Gaussian, we sequentially evaluate splitting, cloning, and opacity decay using accumulated loss-aware surrogate statistics. Joint splitting computes a loss-reducing direction in the joint position-scale space and initializes two children with corresponding position and log-scale perturbations. Cloning minimizes a local Taylor-based surrogate with respect to the duplicate opacity and appends a duplicate child when it is predicted to decrease the loss. Opacity decay similarly estimates a loss-reducing opacity decay strength and selectively decreases the opacity of Gaussians whose current contributions are predicted to be harmful.

Dataset

Mip-NeRF360

Tanks&Temples

Deep Blending

Method

PSNR

\uparrow

SSIM

\uparrow

LPIPS

\downarrow

Num

PSNR

\uparrow

SSIM

\uparrow

LPIPS

\downarrow

Num

PSNR

\uparrow

SSIM

\uparrow

LPIPS

\downarrow

Num

3DGS [ 12 ]

27.35

0.814

0.217

3.4

23.39

0.841

0.183

1.8

29.52

0.903

0.242

3.0

AbsGS [ 37 ]

27.47_{\pm 0.04}

0.819_{\pm 0.00}

\pagecolor{thirdcolor}{0.191}_{\pm 0.00}

3.1

23.67_{\pm 0.06}

0.852_{\pm 0.00}

\pagecolor{thirdcolor}{0.163}_{\pm 0.00}

1.3

\pagecolor{thirdcolor}{29.62}_{\pm 0.04}

0.901_{\pm 0.00}

\pagecolor{thirdcolor}{0.236}_{\pm 0.00}

1.9

PixelGS [ 41 ]

27.58_{\pm 0.01}

\pagecolor{secondcolor}{0.824}_{\pm 0.00}

\pagecolor{secondcolor}{0.190}_{\pm 0.00}

5.5

23.78_{\pm 0.04}

\pagecolor{bestcolor}{0.857}_{\pm 0.00}

\pagecolor{bestcolor}{0.150}_{\pm 0.00}

4.5

28.93_{\pm 0.03}

0.897_{\pm 0.00}

0.248_{\pm 0.00}

4.6

DC4GS [ 11 ]

\pagecolor{thirdcolor}{27.59}_{\pm 0.02}

\pagecolor{bestcolor}{0.826}_{\pm 0.00}

\pagecolor{bestcolor}{0.189}_{\pm 0.00}

2.6

\pagecolor{thirdcolor}{23.94}_{\pm 0.04}

\pagecolor{secondcolor}{0.857}_{\pm 0.00}

\pagecolor{secondcolor}{0.161}_{\pm 0.00}

1.1

\pagecolor{secondcolor}{29.65}_{\pm 0.03}

\pagecolor{bestcolor}{0.906}_{\pm 0.00}

\pagecolor{bestcolor}{0.235}_{\pm 0.00}

1.5

GDAGS [ 42 ]

27.50_{\pm 0.01}

\pagecolor{thirdcolor}{0.822}_{\pm 0.00}

0.192_{\pm 0.00}

3.0

23.66_{\pm 0.06}

0.852_{\pm 0.00}

0.167_{\pm 0.00}

1.4

29.61_{\pm 0.05}

0.904_{\pm 0.00}

\pagecolor{secondcolor}{0.236}_{\pm 0.00}

2.0

Table 1: Quantitative comparison on standard 3DGS benchmarks. Num denotes average Gaussians (M); ∗ follows the 3DGS+residual-splitting setting of [ 23 ] to isolate the effect of the proposed method.

Figure 3: Qualitative comparison on standard 3DGS benchmarks. Red boxes highlight zoom-in regions where our method better preserves fine details and object boundaries.

Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Num	Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Num
(a): Baseline [ 32 ]	$27.05$	$0.795$	$0.248$	$2.1$	(c): (b) + Sec 4.3	$27.37$	$0.813$	$0.229$	$\mathbf{1.6}$
(b): (a) + Sec 4.2	$27.10$	$0.810$	$0.230$	$2.0$	(d): (c) + Sec 4.4	$\mathbf{27.72}$	$\mathbf{0.818}$	$\mathbf{0.219}$	$1.9$

Table 2: Ablation study of the proposed components on Mip-NeRF360 [ 2 ] .

Dataset

Mip-NeRF360

Tanks&Temples

Deep Blending

Method

PSNR

\uparrow

SSIM

\uparrow

LPIPS

\downarrow

Num

PSNR

\uparrow

SSIM

\uparrow

LPIPS

\downarrow

Num

PSNR

\uparrow

SSIM

\uparrow

LPIPS

\downarrow

Num

Position-only [ 32 ]

27.36

0.798

0.239

2.0

23.94

0.847

0.178

1.1

29.47

0.896

0.255

1.1

Joint position-scale (Ours)

\mathbf{27.72}

\mathbf{0.818}

\mathbf{0.219}

1.9

\mathbf{24.20}

\mathbf{0.853}

\mathbf{0.174}

1.0

\mathbf{29.79}

\mathbf{0.906}

\mathbf{0.245}

1.3

Table 3: Direct comparison of position-only and joint position-scale splitting.

Method	Training time (min) $\downarrow$	Peak memory (GB) $\downarrow$	FPS $\uparrow$
3DGS [ 12 ]	$\mathbf{32.5}$	$\mathbf{9.91}$	$157$
Ours	$47.5$	$11.32$	$\mathbf{232}$

Table 4: Comparison of training cost and rendering speed (FPS) on a single NVIDIA A6000 GPU.

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Mip-NeRF360				Tanks&Temples				Deep Blending
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Num	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Num	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Num
Ours ( $\tau_{\mathrm{split}}=-10^{-7}$ )	$27.68$	$0.815$	$0.225$	$1.6$	$24.11$	$0.852$	$0.176$	$0.8$	$29.61$	$0.905$	$0.245$	$1.2$
Ours ( $\tau_{\mathrm{split}}=-10^{-8}$ )	$27.72$	$0.818$	$0.219$	$1.9$	$24.20$	$0.853$	$0.174$	$1.0$	$29.79$	$0.906$	$0.245$	$1.3$
Ours ( $\tau_{\mathrm{split}}=-10^{-9}$ )	$27.72$	$0.818$	$0.219$	$2.0$	$24.26$	$0.853$	$0.174$	$1.1$	$29.72$	$0.906$	$0.245$	$1.3$

Appendix

Table 5: Experiment on the splitting hyperparameter. We set

\tau_{\mathrm{split}}=-10^{-8}

for default configuration.

Dataset	Mip-NeRF360				Tanks&Temples				Deep Blending
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Num	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Num	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Num
Ours ( $\tau_{\mathrm{clone}}=10^{-5}$ )	$27.70$	$0.818$	$0.218$	$1.9$	$24.21$	$0.854$	$0.174$	$1.0$	$29.79$	$0.905$	$0.245$	$1.3$
Ours ( $\tau_{\mathrm{clone}}=10^{-6}$ )	$27.72$	$0.818$	$0.219$	$1.9$	$24.20$	$0.853$	$0.174$	$1.0$	$29.79$	$0.906$	$0.245$	$1.3$
Ours ( $\tau_{\mathrm{clone}}=10^{-7}$ )	$27.71$	$0.818$	$0.219$	$1.9$	$24.10$	$0.853$	$0.174$	$1.0$	$29.80$	$0.906$	$0.245$	$1.4$

Appendix

Table 6: Experiment on the clone hyperparameter. We set

\tau_{\mathrm{clone}}=10^{-6}

for default configuration.

Dataset	Mip-NeRF360				Tanks&Temples				Deep Blending
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Num	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Num	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Num
Ours ( $\tau_{\mathrm{decay}}=10^{-3}$ )	$27.71$	$0.818$	$0.219$	$1.6$	$24.00$	$0.850$	$0.184$	$0.8$	$29.51$	$0.904$	$0.250$	$1.0$
Ours ( $\tau_{\mathrm{decay}}=10^{-4}$ )	$27.72$	$0.818$	$0.219$	$1.9$	$24.20$	$0.853$	$0.174$	$1.0$	$29.79$	$0.906$	$0.245$	$1.3$
Ours ( $\tau_{\mathrm{decay}}=10^{-5}$ )	$27.74$	$0.818$	$0.218$	$2.0$	$24.20$	$0.854$	$0.170$	$1.2$	$29.61$	$0.904$	$0.245$	$1.5$

Appendix

Table 7: Experiment on the decay hyperparameter. We set

\tau_{\mathrm{decay}}=10^{-4}

for default configuration.

Priority order	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Num
C $\rightarrow$ D $\rightarrow$ S	$26.92$	$0.854$	$0.201$	$1.9$
C $\rightarrow$ S $\rightarrow$ D	$26.84$	$0.854$	$0.201$	$1.9$
D $\rightarrow$ C $\rightarrow$ S	$26.69$	$0.853$	$0.204$	$1.9$
D $\rightarrow$ S $\rightarrow$ C	$26.86$	$0.854$	$0.200$	$2.1$
S $\rightarrow$ D $\rightarrow$ C	$26.76$	$0.853$	$0.201$	$2.1$
S $\rightarrow$ C $\rightarrow$ D (Ours)	$\mathbf{26.93}$	$\mathbf{0.854}$	$\mathbf{0.200}$	$2.0$

Appendix

Table 8: Effect of operation priority, averaged over Bicycle, Truck, and Dr. Johnson. S, C, and D denote splitting, cloning, and opacity decay, respectively.

Arbitration strategy	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Num
Adaptive loss-scale arbitration	$26.87$	$0.853$	$0.203$	$2.0$
Fixed S $\rightarrow$ C $\rightarrow$ D (Ours)	$\mathbf{26.93}$	$\mathbf{0.854}$	$\mathbf{0.200}$	$2.0$

Appendix

Table 9: Comparison of adaptive arbitration and fixed priority on the same three scenes as in Tab. 8 .

Mip-NeRF360

Tanks&Temples

Deep Blending

bicycle

flowers

garden

stump

treehill

room

counter

kitchen

bonsai

train

truck

drjohnson

playroom

3DGS

25.19

21.51

27.31

26.58

22.31

31.38

28.93

30.84

32.05

21.79

24.99

28.98

30.07

AbsGS

25.27_{\pm 0.02}

21.26_{\pm 0.03}

27.57_{\pm 0.02}

26.59_{\pm 0.10}

21.93_{\pm 0.04}

31.60_{\pm 0.12}

29.12_{\pm 0.06}

31.63_{\pm 0.25}

32.26_{\pm 0.16}

21.68_{\pm 0.12}

25.65_{\pm 0.04}

29.16_{\pm 0.04}

30.08_{\pm 0.06}

PixelGS

25.30_{\pm 0.03}

21.56_{\pm 0.02}

27.48_{\pm 0.04}

26.82_{\pm 0.04}

22.26_{\pm 0.03}

31.40_{\pm 0.05}

29.16_{\pm 0.02}

31.74_{\pm 0.07}

32.45_{\pm 0.07}

22.02_{\pm 0.07}

25.53_{\pm 0.01}

28.07_{\pm 0.03}

29.79_{\pm 0.04}

DC4GS

25.48_{\pm 0.04}

21.55_{\pm 0.02}

27.62_{\pm 0.14}

26.89_{\pm 0.01}

22.13_{\pm 0.02}

31.78_{\pm 0.07}

28.99_{\pm 0.09}

31.54_{\pm 0.20}

32.30_{\pm 0.01}

22.07_{\pm 0.07}

25.82_{\pm 0.02}

29.17_{\pm 0.02}

30.12_{\pm 0.05}

GDAGS

25.34_{\pm 0.03}

21.25_{\pm 0.08}

27.58_{\pm 0.03}

26.78_{\pm 0.05}

21.98_{\pm 0.03}

31.54_{\pm 0.06}

29.07_{\pm 0.03}

31.76_{\pm 0.08}

32.20_{\pm 0.10}

21.67_{\pm 0.11}

25.64_{\pm 0.02}

29.22_{\pm 0.04}

30.00_{\pm 0.13}

Appendix

Table 10: PSNR scores for Mip-NeRF360, Tanks&Temples, and Deep Blending scenes.

Mip-NeRF360

Tanks&Temples

Deep Blending

bicycle

flowers

garden

stump

treehill

room

counter

kitchen

bonsai

train

truck

drjohnson

playroom

3DGS

0.764

0.605

0.866

0.771

0.632

0.918

0.907

0.925

0.940

0.808

0.876

0.899

0.907

AbsGS

0.782_{\pm 0.00}

0.621_{\pm 0.00}

0.872_{\pm 0.00}

0.777_{\pm 0.00}

0.614_{\pm 0.00}

0.925_{\pm 0.00}

0.912_{\pm 0.00}

0.928_{\pm 0.00}

0.945_{\pm 0.00}

0.817_{\pm 0.00}

0.887_{\pm 0.00}

0.896_{\pm 0.00}

0.906_{\pm 0.00}

PixelGS

0.779_{\pm 0.00}

0.636_{\pm 0.00}

0.871_{\pm 0.00}

0.785_{\pm 0.00}

0.634_{\pm 0.00}

0.922_{\pm 0.00}

0.915_{\pm 0.00}

0.931_{\pm 0.00}

0.946_{\pm 0.00}

0.826_{\pm 0.00}

0.887_{\pm 0.00}

0.888_{\pm 0.00}

0.905_{\pm 0.00}

DC4GS

0.792_{\pm 0.00}

0.635_{\pm 0.00}

0.874_{\pm 0.00}

0.793_{\pm 0.00}

0.631_{\pm 0.00}

0.925_{\pm 0.00}

0.911_{\pm 0.00}

0.929_{\pm 0.00}

0.945_{\pm 0.00}

0.825_{\pm 0.00}

0.889_{\pm 0.00}

0.902_{\pm 0.00}

0.910_{\pm 0.00}

GDAGS

0.786_{\pm 0.00}

0.626_{\pm 0.00}

0.872_{\pm 0.00}

0.785_{\pm 0.00}

0.622_{\pm 0.00}

0.924_{\pm 0.00}

0.911_{\pm 0.00}

0.930_{\pm 0.00}

0.944_{\pm 0.00}

0.818_{\pm 0.00}

0.886_{\pm 0.00}

0.900_{\pm 0.00}

0.907_{\pm 0.00}

Appendix

Table 11: SSIM scores for Mip-NeRF360, Tanks&Temples, and Deep Blending scenes.

Mip-NeRF360

Tanks&Temples

Deep Blending

bicycle

flowers

garden

stump

treehill

room

counter

kitchen

bonsai

train

truck

drjohnson

playroom

3DGS

0.212

0.336

0.108

0.216

0.324

0.220

0.201

0.128

0.205

0.214

0.151

0.245

0.240

AbsGS

0.172_{\pm 0.00}

0.271_{\pm 0.00}

0.099_{\pm 0.00}

0.197_{\pm 0.00}

0.281_{\pm 0.00}

0.200_{\pm 0.00}

0.188_{\pm 0.00}

0.122_{\pm 0.00}

0.190_{\pm 0.00}

0.193_{\pm 0.00}

0.132_{\pm 0.00}

0.241_{\pm 0.00}

0.232_{\pm 0.00}

PixelGS

0.180_{\pm 0.00}

0.261_{\pm 0.00}

0.099_{\pm 0.00}

0.188_{\pm 0.00}

0.277_{\pm 0.00}

0.210_{\pm 0.00}

0.183_{\pm 0.00}

0.119_{\pm 0.00}

0.191_{\pm 0.00}

0.179_{\pm 0.00}

0.120_{\pm 0.00}

0.255_{\pm 0.00}

0.241_{\pm 0.00}

DC4GS

0.166_{\pm 0.00}

0.269_{\pm 0.00}

0.099_{\pm 0.00}

0.188_{\pm 0.00}

0.272_{\pm 0.00}

0.202_{\pm 0.00}

0.190_{\pm 0.00}

0.122_{\pm 0.00}

0.191_{\pm 0.00}

0.189_{\pm 0.00}

0.132_{\pm 0.00}

0.238_{\pm 0.00}

0.233_{\pm 0.00}

GDAGS

0.171_{\pm 0.00}

0.274_{\pm 0.00}

0.099_{\pm 0.00}

0.193_{\pm 0.00}

0.279_{\pm 0.00}

0.204_{\pm 0.00}

0.192_{\pm 0.00}

0.122_{\pm 0.00}

0.192_{\pm 0.00}

0.197_{\pm 0.00}

0.137_{\pm 0.00}

0.239_{\pm 0.00}

0.233_{\pm 0.00}

Appendix

Table 12: LPIPS scores for Mip-NeRF360, Tanks&Temples, and Deep Blending scenes.

Dataset	Tanks&Temples				Deep Blending
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Num	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Num
(a): Baseline [ 32 ]	$23.48$	$0.839$	$0.193$	$1.26$	$29.59$	$0.904$	$0.251$	$1.57$
(b): (a) + Sec. 4.2	$23.53$	$0.842$	$0.191$	$1.27$	$29.65$	$0.907$	$0.248$	$1.64$
(c): (b) + Sec. 4.3	$23.82$	$0.845$	$0.189$	$\mathbf{0.98}$	$29.70$	$\mathbf{0.907}$	$0.245$	$\mathbf{1.31}$
(d): (c) + Sec. 4.4	$\mathbf{24.20}$	$\mathbf{0.853}$	$\mathbf{0.174}$	$1.04$	$\mathbf{29.79}$	$0.906$	$\mathbf{0.245}$	$1.32$

Appendix

Table 13: Ablation study of the proposed components on Tanks&Temples and Deep Blending. Variants (a)–(d) follow Tab. 2 . Num denotes the average number of Gaussians in millions.

Dataset	Mip-NeRF360 (7 scenes)				Tanks&Temples				Deep Blending
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Num	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Num	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Num
3DGS [ 12 ]	$28.90$	$0.863$	$0.196$	$1.8$	$23.62$	$0.842$	$0.186$	$1.0$	$29.48$	$0.901$	$0.254$	$1.3$
3DGS-MCMC [ 14 ]	$29.28$	$\mathbf{0.885}$	$\mathbf{0.168}$	$1.8$	$24.04$	$0.852$	$0.176$	$1.0$	$29.53$	$0.897$	$0.266$	$1.3$
Ours	$\mathbf{29.30}$	$0.874$	$0.184$	$1.8$	$\mathbf{24.20}$	$\mathbf{0.853}$	$\mathbf{0.174}$	$1.0$	$\mathbf{29.79}$	$\mathbf{0.906}$	$\mathbf{0.245}$	$1.3$

Appendix

Table 14: Matched-budget comparison with 3DGS-MCMC. Mip-NeRF360 results are averaged over seven publicly available scenes, unlike the nine-scene results in Tab. 15 . Num denotes the average number of Gaussians in millions.

Dataset	Mip-NeRF360 (9 scenes)				Tanks&Temples				Deep Blending
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Num	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Num	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Num
3DGS [ 12 ]	$27.32$	$0.804$	$0.233$	$1.9$	$23.62$	$0.842$	$0.186$	$1.0$	$29.48$	$0.901$	$0.254$	$1.3$
DC4GS [ 11 ]	$27.52$	$0.816$	$\mathbf{0.201}$	$1.9$	$23.95$	$0.852$	$\mathbf{0.168}$	$1.0$	$29.65$	$0.901$	$\mathbf{0.244}$	$1.3$
Ours	$\mathbf{27.72}$	$\mathbf{0.818}$	$0.219$	$1.9$	$\mathbf{24.20}$	$\mathbf{0.853}$	$0.174$	$1.0$	$\mathbf{29.79}$	$\mathbf{0.906}$	$0.245$	$1.3$

Appendix

Table 15: Comparison with DC4GS under comparable Gaussian budgets.

Figure 6: Convergence comparison with 3DGS on Bicycle, Truck, and Dr. Johnson. Columns show PSNR, LPIPS, and the number of Gaussians.

Figure 7: Patch-level comparison with DC4GS. The examples illustrate strengths in reconstructing object structures and object-background boundaries, as well as limitations in preserving fine details in repeated high-frequency textures.

3D Gaussian Splatting

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

See all 1,051 papers

May 1, 2026

Beyond Heuristics: Learnable Density Control for 3D Gaussian Splatting

While 3D Gaussian Splatting (3DGS) has demonstrated impressive real-time rendering performance, its efficacy remains constrained by a reliance on heuristic density control. Despite numerous refinements to these handcrafted rules, such methods inherently lack the flexibility to adapt to diverse scenes with complex geometries. In this paper, we propose a paradigm shift for density control from rigid heuristics to fully learnable policies. Specifically, we introduce LeGS, a framework that reformulates density control as a parameterized policy network optimized via Reinforcement Learning (RL). Central to our approach is the tailored effective reward function grounded in sensitivity analysis, which precisely quantifies the marginal contribution of individual Gaussians to reconstruction quality. To maintain computational tractability, we derive a closed-form solution that reduces the complexity of reward calculation from

O(N^2)

O(N)

. Extensive experiments on the Mip-NeRF 360, Tanks & Temples, and Deep Blending datasets demonstrate that LeGS significantly outperforms state-of-the-art methods, striking a superior balance between reconstruction quality and efficiency. The code will be released at https://github.com/AaronNZH/LeGS

Zhenhua Ning, Xin Li, Jun Yu +3

Pengcheng Laboratory, Shenzhen · Harbin Institute of Technology, Shenzhen

3D Gaussian Splatting 3D Reconstruction Adaptive Gaussian Densification

Ask PDF

Aug 23, 2026

Fast and Compact 3D Gaussian Splatting with Polarized Opacity Prior

3D Gaussian Splatting (3DGS) achieves state-of-the-art rendering quality at real-time speeds but suffers from "model bloat" - a large number of redundant, low-opacity Gaussians that inflate memory usage and training costs. This inefficiency stems from the standard "densify-then-prune" paradigm, which expands the model aggressively before relying on pruning to achieve compactness. To mitigate this problem, we present an efficient training framework that builds an intrinsically compact representation, replacing the conventional densify-then-prune cycle. Our method leverages a synergistic design: an L2 reconstruction loss to provide error-proportional gradients that stabilize optimization, and a novel Polarized Opacity Prior (POP) to actively manage the Gaussian population. POP steers informative primitives toward full opacity and uninformative ones toward transparency, enabling natural pruning and accelerating rendering through Early Ray Termination. Experiments on three public datasets demonstrate that our approach consistently achieves accelerated 3DGS training with significantly fewer Gaussians while maintaining comparable visual reconstruction quality. These results show that the proposed framework provides a simple and effective path toward fast and inherently compact 3DGS training.

Zi-Ming Wang, Kai-Wen Duan, Kowei Huang +2

National Tsing Hua University, Taiwan · National Institute of Informatics, Japan

3D Gaussian Splatting 3DGS Acceleration 3DGS Compression 3D Gaussian Pruning

Ask PDF

Oct 29, 2026

ACD-GS: Asymmetric Curvature-aware Densification for 3D Gaussian Splatting

3D Gaussian Splatting (3DGS) has emerged as a powerful paradigm for real-time photorealistic novel view synthesis. However, its training-time densification remains largely heuristic, often resulting in redundant Gaussian proliferation and inaccurate geometric reconstruction. Existing methods predominantly rely on first-order screen-space gradients to trigger cloning and splitting, which fail to distinguish true geometric discontinuities from high-frequency textures or noisy residuals. To address this limitation, we propose ACD-GS, an Asymmetric Curvature-aware Densification framework for compact and geometry-aware 3D Gaussian optimization. Our key insight is that second-order depth responses provide a more reliable structural prior for density control, enabling asymmetric regulation of cloning and splitting to better capture geometric discontinuities. Building on this, we introduce a multi-view photometric consistency constraint to suppress transient single-view errors, together with a capacity-aware adaptive regulation strategy that balances densification and pruning during training. Unlike post-hoc compression or quantization-based approaches, ACD-GS directly prevents redundant Gaussian generation during training, yielding inherently compact representations. Extensive experiments demonstrate that our method reduces nearly 60% of Gaussian count and storage while maintaining competitive rendering fidelity, consistently outperforming existing methods across multiple datasets. Our code is available in https://waveviewer.github.io/ACD-GS/.

Houqiang Zhong, Tianchi Zhu, Qiang Hu +3

School of Information Science and Electronic Engineering, Shanghai Jiao Tong University, Shanghai, China · SJTU Paris Elite Institute of Technology, Shanghai Jiao Tong University, Shanghai, China · Cooperative Medianet Innovation Center, Shanghai Jiao Tong University, Shanghai, China

3D Gaussian Splatting 3DGS Training Adaptive Gaussian Densification 3D Gaussian Pruning Multi-View 3D Reconstruction

Ask PDF

Abstract

3D Gaussian Splatting

Explore similar work

Beyond Heuristics: Learnable Density Control for 3D Gaussian Splatting

Fast and Compact 3D Gaussian Splatting with Polarized Opacity Prior

ACD-GS: Asymmetric Curvature-aware Densification for 3D Gaussian Splatting