Fast and Compact 3D Gaussian Splatting with Polarized Opacity Prior

Zi-Ming Wang; Kai-Wen Duan; Kowei Huang; Akihiro Sugimoto; Shang-Hong Lai

cs.CVAug 23, 2026

Fast and Compact 3D Gaussian Splatting with Polarized Opacity Prior

Authors: Zi-Ming Wang, Kai-Wen Duan, Kowei Huang, Akihiro Sugimoto, Shang-Hong Lai

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

Abstract

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.

Figures & tables

Figure 1: Average PSNR versus training time on Tanks and Temples dataset [ 15 ] . Our method achieves the fastest training speed among state-of-the-art approaches while maintaining competitive reconstruction quality. The size of each marker indicates the number of Gaussians used in the corresponding model.

Figure 2: Opacity statistics on the Kitchen scene [ 1 ] . Approximately 19.9% of Gaussians have opacity $>0.9$ , while 33.7% remain at opacity $<0.1$ , indicating a large number of redundant primitives.

Figure 3: Artifact demonstration on the Kitchen scene when forcing $o_{i}\to 1$ for all Gaussians: (a) Ground Truth, (b) Rendered image, (c) Error map $|Render-GT|$ , (d) $\mathcal{L}_{1}$ loss gradient w.r.t. pixel color, (e) SSIM loss gradient w.r.t. pixel color, (f) $\mathcal{L}_{2}$ loss gradient w.r.t. pixel color.

Panel of Figure 4 — Figure 4: Effectiveness of our proposed method on Mip-NeRF 360 [ 1 ] . (a) $\mathcal{L}_{2}$ gradients (green) adapt to the error magnitude, whereas $\mathcal{L}_{1}$ gradients (red) remain constant. (b) POP significantly reduces the average Gaussians per pixel, enabling Early Ray Termination compared to the 3DGS baseline.

Dataset	Mip-NeRF 360 [ 1 ]						Deep Blending [ 12 ]
Method	Num $\downarrow$	Time $\downarrow$	Ckpt $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Num $\downarrow$	Time $\downarrow$	Ckpt $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
3DGS [ 14 ]	2,509k	1,304	595	27.250	0.811	0.226	2,484k	1,197	588	29.847	0.907	0.089
Gsplat [ 29 ]	3,168k	886	713	27.661	0.824	0.166	3,042k	789	685	29.573	0.904	0.169
PUP-3DGS [ 11 ]	312k	1,216	74	26.772	0.795	0.258	282k	1,205	67	29.471	0.903	0.103
Mini-Splatting [ 7 ]	495k	896	117	27.605	0.833	0.202	348k	834	82	29.987	0.908	0.253
Taming-3DGS [ 19 ]	346k	329	82	27.147	0.772	0.294	294k	282	69	29.886	0.905	0.270

Table 1: Quantitative evaluation on Mip-NeRF 360 and Deep Blending . Training time and checkpoint size (Ckpt) are reported in seconds and MB , respectively. Best, second-best, and third-best results are highlighted in red , orange , and yellow , respectively.

Dataset	Tanks & Temples
Method	Num $\downarrow$	Time (s) $\downarrow$	Ckpt (MB) $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
3DGS [ 14 ]	1,573k	683	372	23.808	0.853	0.091
Gsplat [ 29 ]	1,840k	543	414	23.508	0.845	0.127
PUP-3DGS [ 11 ]	183k	745	43	22.672	0.804	0.161
PUP-3DGS ∗ [ 11 ]	275k	766	65	23.049	0.818	0.222
Mini-Splatting [ 7 ]	201k	591	48	23.212	0.835	0.202

Table 2: Quantitative evaluation on Tanks & Temples .

\ast

indicates that the method is evaluated under a reduced compression strength for fair comparison.

Panel of Figure 5 — Figure 5: Qualitative comparison on various scenes. Our method (top) achieves higher reconstruction fidelity with fewer artifacts compared to 3DGS (middle). Red boxes highlight improved fine details (e.g., Truck, Bonsai) and smoother surface reconstruction (e.g., Train, Playroom).

Panel of Figure 6 — Figure 6: Ablation on loss functions. (a) Impact of different supervisions on Gaussian count. (b) Trade-off between PSNR and Gaussian count for $\mathcal{L}_{1}$ with varying scaling factors (circles) and the $\mathcal{L}_{2}$ loss (star).

Modules			Mip-NeRF 360					Tanks & Temples					Deep Blending
L2	Blur	POP	Num $\downarrow$	Time $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Num $\downarrow$	Time $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Num $\downarrow$	Time $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
✓	✗	✗	210k	271	26.57	0.725	0.339	221k	271	23.53	0.809	0.191	193k	259	28.51	0.870	0.252
✓	✓	✗	583k	343	27.22	0.771	0.255	557k	267	23.71	0.821	0.161	297k	275	28.56	0.873	0.243
✓	✗	✓	140k	264	26.28	0.715	0.354	157k	176	23.46	0.798	0.209	94k	254	28.43	0.867	0.272
✓	✓	✓	268k	279	27.01	0.760	0.277	269k	209	23.62	0.811	0.181	118k	257	28.63	0.869	0.266

Table 3: Ablation on our components . “Blur” denotes our Blur-based Densification module. Our full method is highlighted in gray . Bold numbers indicate the best performance for each metric.

Dataset	Mip-NeRF 360					Tanks & Temples					Deep Blending
Blur Iter.	Num $\downarrow$	Time $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Num $\downarrow$	Time $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Num $\downarrow$	Time $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
3k–4k	175k	269	26.54	0.736	0.321	185k	186	23.45	0.804	0.196	100k	257	28.21	0.863	0.271
3k–7k	268k	279	27.01	0.760	0.277	269k	209	23.62	0.811	0.181	118k	257	28.63	0.869	0.266
3k–11k	394k	326	27.13	0.768	0.260	412k	238	23.55	0.814	0.172	188k	264	28.53	0.869	0.261
3k–15k	562k	358	27.17	0.773	0.250	647k	275	23.55	0.816	0.165	331k	284	28.26	0.867	0.262

Table 4: Impact of blur-based densification schedules. We highlight our balanced 3k–7k setting in gray . Bold numbers indicate the best performance for each metric.

Dataset	Mip-NeRF 360					Tanks & Temples					Deep Blending
Method	Num $\downarrow$	Time $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Num $\downarrow$	Time $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	Num $\downarrow$	Time $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Reset	548k	328	26.82	0.780	0.260	432k	247	22.91	0.795	0.201	234k	270	28.51	0.871	0.258
POP	268k	279	27.01	0.760	0.277	269k	209	23.62	0.811	0.181	118k	257	28.63	0.869	0.266

Table 5: Ablation study of our Polarized Opacity Prior (POP) against the standard Opacity Reset . POP allows for a more compact representation and efficient training while maintaining competitive visual quality across all datasets.

Figure 7: Average number of Gaussians during training on the Mip-NeRF 360 [ 1 ] dataset. Note that PUP-3DGS [ 11 ] performs post-training compression after 30,000 iterations.

Explore similar work

MMGS: 10 $\times$ Compressed 3DGS through Optimal Transport Aggregation based on Multi-view Ranking
May 19, 2026Beizhen Zhao, Sicheng Yu, Ziran Yin +23DGS Compression 3D Gaussian Splatting
The Hong Kong University of Science and Technology (Guangzhou)
Proxy-GS: Unified Occlusion Priors for Training and Inference in Structured 3D Gaussian Splatting
Sep 29, 2025Yuanyuan Gao, Yuning Gong, Yifei Liu +63D Gaussian Splatting 3DGS Acceleration
Shanghai Artificial Intelligence Laboratory · Northwestern Polytechnical University · Sichuan University +2
REFINE: Super-efficient 3D Gaussian Splatting Pruning via Rendering-Free Primitive Importance
Jun 8, 2026Zhang Chen, Shuai Wan, Mengting Yu +23D Gaussian Splatting 3D Gaussian Pruning
School of Electronics and Information, Northwestern Polytechnical University · School of Telecommunication Engineering, Xidian University · Department of Computer Science, City University of Hong Kong

May 19, 2026J/K move · Enter open · S save

MMGS: 10 $\times$ Compressed 3DGS through Optimal Transport Aggregation based on Multi-view Ranking

Beizhen Zhao, Sicheng Yu, Ziran Yin, Dongxu Shen+1

The Hong Kong University of Science and Technology (Guangzhou)

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

Open paper AskPDF arXiv

While 3D Gaussian Splatting (3DGS) has revolutionized 3D reconstruction, it suffers from significant overhead due to massive redundant primitives. Existing compression methods typically rely on local sampling or fixed pruning thresholds, which often struggle to balance redundancy reduction with high-fidelity rendering. To address this, we propose a novel framework that formulates Gaussian optimization as a global geometric distribution matching problem. Specifically, our approach integrates three components: (1) we introduce a multi-view 3D Gaussian contribution ranking mechanism that filters primitives using geometric consistency instead of local heuristics; (2) we propose a global Optimal Transport (OT)-based aggregation algorithm that merges redundant primitives while preserving the underlying geometry; and (3) we design an OT-based densification operator that maintains the Gaussian's distributional properties for stable optimization. Our approach achieves state-of-the-art rendering quality with only \textbf{10

\%

} primitives and \textbf{10

\times

} accelerated training speeds compared to vanilla 3DGS.

Abstract

Explore similar work

MMGS: 10×\times× Compressed 3DGS through Optimal Transport Aggregation based on Multi-view Ranking

Proxy-GS: Unified Occlusion Priors for Training and Inference in Structured 3D Gaussian Splatting

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

MMGS: 10 $\times$ Compressed 3DGS through Optimal Transport Aggregation based on Multi-view Ranking