Speeding Up the Learning of 3D Gaussians with Much Shorter Gaussian Lists

Jiaqi Liu; Zhizhong Han

Figure 2: Our method achieves the fastest training time while maintaining comparable reconstruction quality.

Figure 3: Overview of our method and effects. (a) The Gaussian list of 3DGS. (b) Distribution of Gaussian list length showing our method produces significantly shorter lists. (c) Gaussian list reduction after applying scale reset. (d, e) Scale and opacity distributions comparing 3DGS and 3DGS with scale reset, showing scale reset produces smaller Gaussians with higher opacities. (f) Gaussian list reduction after applying entropy regularization. (g, h) Scale and opacity distributions comparing 3DGS and 3DGS with entropy regularization, demonstrating entropy constraint produces smaller Gaussians and more polarized opacities. “3DGS” results are produced with LiteGS.

Figure 4: Scale reset achieves shorter Gaussian lists (limited x-range for clarity).

Figure 5: A toy example demonstrating how entropy regularization polarizes Gaussian weights along a ray.

Method	Iters	Mip-NeRF 360					Deep Blending					Tanks and Temples
Method	Iters	$N_{\text{G}}$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Time(s) $\downarrow$	$N_{\text{G}}$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Time(s) $\downarrow$	$N_{\text{G}}$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Time(s) $\downarrow$
AdR-Gaussian	30K	1.3M	26.92	0.792	0.257	504.86	1.4M	29.72	0.905	0.250	515.78	0.9M	23.49	0.835	0.202	321.08
EDGS	5K	3.5M	26.46	0.817	0.205	318.06	2.3M	29.08	0.898	0.260	214.76	2.2M	22.23	0.831	0.184	219.06
Mini-Splatting2-Dense	18K	3.6M	27.56	0.827	0.184	220.22	4.2M	29.78	0.908	0.222	162.38	2.6M	23.43	0.852	0.152	162.24
3DGS	30K	3.3M	27.55	0.819	0.209	919.51	2.8M	29.74	0.907	0.237	963.66	1.8M	23.77	0.854	0.166	560.52
Taming-3DGS			27.85	0.823	0.208	402.54		30.07	0.910	0.234	322.06		24.14	0.855	0.169	243.25

Table 1: Quantitative comparison across three datasets. We follow Taming-3DGS’s target Gaussian count setup (3.3M, 2.8M, 1.8M for the three datasets respectively) for fair comparison. For methods whose Gaussian number cannot be controlled, we report their default counts.

Method	Iters	Mip-NeRF 360					Deep Blending					Tanks and Temples
Method	Iters	$N_{\text{G}}$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Time(s) $\downarrow$	$N_{\text{G}}$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Time(s) $\downarrow$	$N_{\text{G}}$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Time(s) $\downarrow$
Mini-Splatting2	18K	0.6M	27.38	0.821	0.215	92.68	0.6M	30.11	0.912	0.239	78.87	0.4M	23.11	0.841	0.186	72.45
DashGaussian			26.89	0.777	0.284	65.01		29.93	0.901	0.270	48.11		23.42	0.823	0.229	57.94
LiteGS			26.75	0.765	0.299	66.33		29.75	0.903	0.275	49.16		23.07	0.816	0.234	49.74
Ours	18K	0.6M	26.49	0.762	0.303	43.87	0.6M	29.43	0.898	0.278	37.76	0.4M	22.96	0.819	0.226	44.81

Table 2: Quantitative comparison with 18K iterations and fewer Gaussians, following Mini-Splatting2’s setup. The performance gap compared to Tab. 1 stems from LiteGS backbone limitations under constrained settings, not our acceleration strategy.

Figure 6: Per-iteration training time breakdown.

Figure 7: Qualitative comparison of rendered results. Each row shows results from 3DGS, Taming 3DGS, DashGaussian, LiteGS, and ours with PSNR and training time. More comparisons are detailed in Fig. 16 , Fig. 17 , Fig. 18 and Fig. 19 of the supplementary material.

Figure 8: Qualitative comparison of rendering results, comparing our method with DashGaussian. Each row shows results at certain training times with PSNR and SSIM metrics displayed. More comparisons are detailed in Fig. 20 of the supplementary material.

Method	Iters	$N_{\text{G}}$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Time(s) $\downarrow$
FastGS	30K	0.4M	27.43	0.797	0.261	88.55
Ours	30K	0.4M	26.50	0.754	0.314	66.32

Table 3: Quantitative comparison with FastGS on Mip-NeRF 360 under FastGS’s default setting (0.4M Gaussians).

Method	Mip-NeRF 360
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Time(s) $\downarrow$
L	27.75	0.822	0.208	191.17
L+R	27.33	0.815	0.212	147.33
L+E	27.35	0.815	0.218	162.53
L+R+E	27.14	0.812	0.215	141.28
L+D	27.85	0.822	0.213	134.99

Table 4: Ablation study analyzing effects of individual modules.

Method	Mip-NeRF 360
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Time(s) $\downarrow$
L+D+R(0.2)+E(0.01)	27.35	0.811	0.222	102.86
L+D+R(0.2)+E(0.02)	27.14	0.808	0.226	99.10
L+D+R(0.3)+E(0.015)	27.31	0.811	0.223	104.39
L+D+R(0.1)+E(0.015)	27.14	0.807	0.227	98.34
L+D+R(0.2)+E(0.015)	27.28	0.810	0.224	99.58

Table 5: Ablation study on parameter selection.

Method	Mip-NeRF 360
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Time(s) $\downarrow$
L+D+V(1.0)+E(0.015)	27.17	0.808	0.230	107.91
L+D+R(0.2)+E(0.015)	27.28	0.810	0.224	99.58

Table 6: Comparison of volume regularization versus scale reset.

Method	Mip-NeRF 360
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Time(s) $\downarrow$
L+D+R(0.2)+O(3.0)	27.29	0.807	0.231	106.16
L+D+R(0.2)+E(0.015)	27.28	0.810	0.224	99.58

Table 7: Comparison of opacity versus entropy constraint.

Method	Mip-NeRF 360
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Time(s) $\downarrow$
3DGS (16 $\times$ 16)	27.55	0.819	0.209	919.51
3DGS (8 $\times$ 8)	27.56	0.819	0.208	1004.30

Table 8: Ablation study of tile size impact on time and quality.

Figure 9: Training time versus iterations. Plot for compact models are shown in Supplementary Fig. 13 .

Figure 10: PSNR convergence across iterations. Plot for compact models is shown in Supplementary Fig. 14 .

Figure 11: Comparison of Gaussian list length distributions across individual and combined modules.

Figure 12: Gaussian count heatmap comparison across modules. Colors represent per-tile Gaussian counts. See color bar in Fig. 1 .

Method	FPS $\uparrow$
3DGS	140.91
Taming	219.60
LiteGS	233.76
Ours	343.92

Table 9: Average testing FPS on the Mip-NeRF 360 dataset.

Figure 13: Training time versus iterations for full and compact models, corresponding to Fig. 9 .

Figure 14: PSNR convergence across iterations for full and compact models, corresponding to Fig. 10 .

Figure 15: PSNR and time curves across iterations showing effects of individual modules. See Sec. 5.4 for more details.

Figure 16: Qualitative comparison of rendered results on scene Stump. Our method achieves comparable rendering quality (PSNR: 27.33) while being the fastest, with training time of only 87.62 seconds.

Figure 17: Qualitative comparison of rendered results on scene Counter. Our method achieves comparable rendering quality (PSNR: 30.38) while being the fastest, with training time of only 86.66 seconds.

Figure 18: Qualitative comparison of rendered results on scene Bicycle. Our method achieves comparable rendering quality (PSNR: 21.29) while being the fastest, with training time of only 115.37 seconds.

Figure 19: Qualitative comparison of rendered results on scene Room. Our method achieves comparable rendering quality (PSNR: 34.23) while being the fastest, with training time of only 79.42 seconds.

Figure 20: Qualitative comparison of rendering quality across two scenes from Mip-NeRF 360, comparing our method with DashGaussian. Each row shows results at training times of 15s, 30s, 50s, 70s, and 90s. PSNR and SSIM metrics are displayed in each image.

Figure 21: Gaussian count heatmap comparison across scenes in Mip-NeRF 360 dataset, corresponding to Fig. 1 . Each row shows ground truth and heatmaps for 3DGS, Taming-3DGS, LiteGS, and our method. Colors represent per-tile Gaussian counts, where purple indicates low counts and yellow indicates high counts. Our method consistently achieves the lowest Gaussian counts (darker colors).

Figure 22: Gaussian count heatmap comparison across scenes in Mip-NeRF 360 dataset for ablation study, corresponding to Fig. 12 . Each row shows ground truth and heatmaps for 3DGS, 3DGS+Scale Reset, 3DGS+Entropy Constraint, and 3DGS+Reset+Entropy. Colors represent per-tile Gaussian counts, where purple indicates low counts and yellow indicates high counts.

Figure 23: Per-pixel entropy heatmap comparison between LiteGS and our method across scenes in the Mip-NeRF 360 dataset. Darker colors correspond to lower entropy.

Figure 24: Distribution comparison per scene across all methods on length, scale, and opacity. Results labeled as “3DGS” are obtained using LiteGS. See Sec. 5.4 for more details.

Figure 25: Visualization of Gaussians on selected scenes from the Mip-NeRF 360 dataset. Compared to 3DGS, our method produces smaller and more compact Gaussians.

Speeding Up the Learning of 3D Gaussians with Much Shorter Gaussian Lists

Abstract

Figures & tables

3DGS Training

EffGS: Efficient and High-Fidelity Gaussian Splatting

Fast and Compact 3D Gaussian Splatting with Polarized Opacity Prior

TemporalGS: Training-Free Plug-and-Play Acceleration for 3D Gaussian Splatting Rendering via Temporal Priors

Abstract

3DGS Training

Explore similar work

EffGS: Efficient and High-Fidelity Gaussian Splatting

Fast and Compact 3D Gaussian Splatting with Polarized Opacity Prior

TemporalGS: Training-Free Plug-and-Play Acceleration for 3D Gaussian Splatting Rendering via Temporal Priors