Gaussian Stippling: Efficient Sorting-Free 3D Gaussian Rendering through Hybrid Sampling and Spatiotemporal Reconstruction

Zijian Huang; Suiliang Mai; Chuankun Zheng; Yuan Meng; Yuchi Huo

cs.GRSep 29, 2026

Gaussian Stippling: Efficient Sorting-Free 3D Gaussian Rendering through Hybrid Sampling and Spatiotemporal Reconstruction

Authors: Zijian Huang, Suiliang Mai, Chuankun Zheng, Yuan Meng, Yuchi Huo

Organizations: State Key Laboratory of CAD&CG, Zhejiang University, Hangzhou, China

Abstract

Conventional 3D Gaussian Splatting (3DGS) requires depth sorting and ordered alpha blending to correctly render overlapping Gaussian primitives. Stochastic transparency enables sorting-free rendering by replacing fractional alpha contributions with discrete stochastic visibility samples, but produces substantial spatial and temporal noise at low sample counts. We refer to this conversion from continuous Gaussian splats to discrete visibility samples as \textit{Gaussian Stippling}. Based on this, we present an efficient order-independent rendering and reconstruction framework that operates directly on unmodified 3DGS assets. Our method adaptively integrates primitive-based and fragment-based stippling, leveraging their complementary strengths across different rendering regimes to significantly improve rendering throughput. To recover high-quality images from sparse stochastic samples, we further introduce a lightweight Gaussian-aware spatiotemporal reconstruction network. By exploiting the Gaussian attributes retained by each stipple, the network aggregates structured stochastic clues across both space and time, effectively suppressing stippling noise. Experiments show that our hybrid Gaussian stippling method, coupled with a spatiotemporal reconstruction network trained on diverse scenes, generalizes to unseen scenes and enables interactive, temporally stable, and visually plausible rendering on mobile devices without retraining or preprocessing. With scene-specific training and appropriately scaled sampling and network capacity, our method further outperforms the baselines in visual quality, offering a high-fidelity configuration for quality-prioritized applications.

Figures & tables

Figure 1: Sorting-free stochastic rendering achieves high throughput but produces substantial spatial noise (top). We augment its observations with Gaussian attributes and reconstruct them using a lightweight spatiotemporal network.

Figure 2: Overview of our proposed Gaussian stippling and reconstruction pipeline. In Stage 1, given a 3DGS asset containing

N

Gaussian primitives, a cost-aware routing strategy assigns

N_{F}

primitives to the fragment-based stream and the remaining primitives to the primitive-based stream, which generates

N_{P}

stochastic sample points. A shared depth test then selects accepted stipples together with their Gaussian-local descriptors to generate the current-frame observation. In Stage 2, the reconstruction network aggregates the current observation and forward-reprojected historical observations to produce the reconstructed image.

Figure 3: Time-cost ratio of two streams for Layer $L=1$ . The red solid line is the fitted boundary and the gray dotted line is the measured boundary.

Figure 4: Per-view latency. Ours has the lowest latency among the compared methods on the evaluated views and reduces the viewpoint-dependent variation observed with Gaussian Point Splatting.

Method	Mip-NeRF360				Tanks&Temples				Deep Blending
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	FPS $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	FPS $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	FPS $\uparrow$
StochasticSplats	17.29	0.2427	0.8942	664.95	13.30	0.1754	0.9961	843.08	21.61	0.3562	0.8063	556.03
Gaussian Point Splatting	17.28	0.2413	0.8947	620.56	13.30	0.1749	0.9961	440.99	21.60	0.3559	0.8065	438.95
Ours (1-spp)	17.28	0.2407	0.8947	880.07	13.29	0.1738	0.9965	1001.78	21.61	0.3559	0.8065	1024.15
3DGS	28.90	0.8708	0.1856	122.18	23.39	0.8421	0.1837	113.19	29.52	0.9038	0.2459	111.18
MobileGS †	28.20	0.8570	0.2103	234.97	- ∗	- ∗	- ∗	- ∗	30.07	0.9109	0.2486	292.26

Table 1: Quality against ground truth at the training resolution; throughput at 1080p. The first three rows report renderer-only FPS; Ours-S/L rows include rendering and reconstruction. Best/second-best PSNR and FPS within each block are bold/underlined; ties share rank. ∗ MobileGS optimization produced NaNs on Tanks&Temples. † MobileGS FPS includes its view-dependent MLP (TensorRT) and rasterization; its released timing code omits the MLP.

Figure 5: Qualitative results of 1-spp Ours-S . A compact reconstruction network suppresses stochastic noise using Gaussian-aware observations from the current and historical frames.

Figure 6: Qualitative results of 16-spp Ours-L , illustrating image reconstruction with a higher observation budget and a larger network.

Variant	Bonsai	Counter	Train	Avg.
Full descriptor	30.749	28.263	21.450	26.820
RGB only	30.376	28.093	21.117	26.529
w/o alpha	30.732	28.253	21.439	26.808
w/o conic	30.653	28.197	21.373	26.741
w/o opacity	30.758	28.249	21.421	26.809
w/o $d_{M}^{2}$	30.710	28.263	21.453	26.809

Table 2: Ablation of temporal history and Gaussian-local descriptors using Ours-S . We evaluate variants that remove all historical samples, keep only a single history frame, or omit specific descriptor channels from the full input. All values are PSNR evaluated against the standard 3DGS target.

Method	Flow.	House	Pavil.	Plant	Stat.	Avg.
3DGS	22.73	23.10	21.25	22.76	21.62	22.29
Ours (1-spp)	13.47	13.45	12.67	13.92	13.27	13.36
SS (4-spp)	17.80	17.86	16.86	18.23	17.39	17.63
GPS (4-spp)	17.83	17.87	16.86	18.22	17.51	17.66
Ours (1-spp + S)	21.74	22.35	21.01	21.60	20.94	21.53

Table 3: PSNR on held-out DL3DV views with respect to photographic ground truth. Avg. is the arithmetic mean over the five scenes. The best and second-best values are shown in bold and underlined, respectively. SS (4-spp), GPS (4-spp), and Ours (1-spp + S) run at 353.32, 282.91, and 353.43 FPS, respectively.

Method	Truck	Room	DrJohnson	Avg.
3DGS-GL	5.46	12.8	3.2	7.15
SS	24.8	65.8	69.4	53.3
GPS	41.3	52.8	53.0	49.03
SortfreeGS	37.2	50.3	27.2	38.2
MobileGS	–	32.6	28.1	30.4
Ours(1-spp)	78.8	112.7	86.3	92.6

Table 4: Mobile 540p throughput (FPS) on vivo x300pro with MediaTek tianji9500, using one scene per benchmark. Our renderer uses Vulkan, and Ours-S runs in FP16 on the NPU. Avg. is the arithmetic mean over available scene results. The best and second-best values are shown in bold and underlined, respectively.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Method	Truck	Room	DrJohnson	Avg.
3DGS-GL	5.46	12.8	3.2	7.15
SS	24.8	65.8	69.4	53.3
GPS	41.3	52.8	53.0	49.03
SortfreeGS	32.94	22.03	31.95	28.97
MobileGS	–	26.98	25.61	26.29
Ours (1-spp)	78.8	112.7	86.3	92.6

Appendix

Table A1: Mobile 540p throughput (FPS) on vivo x300pro with MediaTek tianji9500, using one scene per benchmark. Our renderer uses Vulkan, and Ours-S runs in FP16 on the NPU. Avg. is the arithmetic mean over available scene results. The best and second-best values are shown in bold and underlined, respectively.

Figure A1: Crossover boundaries ( $t_{\mathrm{primitive}}=t_{\mathrm{fragment}}$ ) for varying layer counts $L\in\{1,\dots,8\}$ plotted over the opacity-area plane. The almost perfectly overlapping curves indicate that depth complexity has a negligible effect on the relative performance of the two streams.

Scene	Small (1 spp)			Large (1 spp)
Scene	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Bicycle	24.0400	0.6710	0.2600	24.6107	0.6996	0.2745
Bonsai	30.7500	0.9110	0.1600	32.1253	0.9284	0.2303
Counter	28.2700	0.8690	0.1890	28.9801	0.8927	0.2317
Garden	25.4000	0.7470	0.1750	26.0137	0.7769	0.1891
Kitchen	29.2100	0.8750	0.1550	30.0836	0.8948	0.1703

Appendix

Table A2: Per-scene and per-dataset reconstruction quality of the Small and Large variants at 1 spp, measured against ground-truth images. Dataset averages are arithmetic means over their constituent scenes; the final row is the macro-average over all 11 scenes.

Scene	Small (4 spp)			Large (4 spp)
Scene	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Bicycle	24.5220	0.7018	0.2799	24.9213	0.7212	0.2614
Bonsai	31.2651	0.9232	0.2346	32.4673	0.9317	0.2279
Counter	28.5755	0.8876	0.2357	29.1298	0.8981	0.2263
Garden	26.1494	0.7916	0.1903	26.5915	0.8104	0.1662
Kitchen	29.9029	0.8948	0.1701	30.5050	0.9057	0.1595

Appendix

Table A3: Per-scene and per-dataset reconstruction quality of the Small and Large variants at 4 spp, measured against ground-truth images.

Scene	Small (16 spp)			Large (16 spp)
Scene	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Bicycle	24.9336	0.7303	0.2550	25.1800	0.7470	0.2080
Bonsai	31.7702	0.9304	0.2228	32.8700	0.9350	0.1430
Counter	28.7965	0.8974	0.2180	29.2900	0.9000	0.1690
Garden	26.7682	0.8296	0.1487	27.0900	0.8450	0.1000
Kitchen	30.4227	0.9098	0.1461	30.7800	0.9150	0.1120

Appendix

Table A4: Per-scene and per-dataset reconstruction quality of the Small and Large variants at 16 spp, measured against ground-truth images. Dataset averages are arithmetic means over their constituent scenes; the final row is the macro-average over all 11 scenes.

Variant	Bonsai	Counter	Train	Avg.
Forward Nearest.	30.75	28.23	21.50	26.83
Forward Bilinear.	30.60	28.19	21.38	26.72
Backward Nearest.	30.59	28.21	21.40	26.73
Backward Bilinear.	30.68	28.25	21.40	26.78

Appendix

Table A5: Temporal transport under stochastic visibility. Backward gathering produces structured artifacts; bilinear transport spreads samples, whereas nearest-forward transport preserves discrete observations. The best and second-best PSNR (dB) values are shown in bold and underlined ; ties share the same rank.

Scene	Hash Value
Pavil.	d3812aad538261e7f73c75762ff55f23b468bcc76f376d52ac86ca6cf3c44b4b
Stat.	c37109a55effe0000f8e40652ca935376e75bcb2a0b56de8eabd20a26e2a0f68
Flow.	1d6a9ed47cce39fd1c4d18f776bcc97e507b81bde921ca596bd91b0b02b5e414
Plant	dafa9c7cbda9d1ddaa8a2b51fc8c54f4eb44161f5e5c53685dc744580ca77751
House	6d81c5ab0d480fd43d78b75ff372a8113ad38e2c03f1d69627c009883054d4c2

Appendix

Table A6: Scene names and corresponding dataset hash values in DL3DV.

Figure A2: After 1024spp rendering using GPS’s source code, the result was compared with that of Gaussian rendering, and there was a loss. The only modification to Gaussian Point Splatting is that we removed the upper limit on individual Gaussian sampling.

Figure A3: Rendering failure of Gaussian Point Splatting: its workload distribution loses much detail when aligning the 3DGS bounding box at conventional rendering resolution.

Configuration	Ours-S	Ours-L
Input / output channels	$40/3$	$40/3$
Down-/upsampling stages	$3/3$	$4/4$
Resolution pyramid	$H$ to $H/8$	$H$ to $H/16$
Encoder widths, fine to coarse	$16,16,16,16$	$64,96,128,192,256$
Encoder/bottleneck convolutions per resolution	$2,1,1,2$ (first stage: $1{\times}1$ , then $3{\times}3$ )	$2,2,2,2,2$ (all $3{\times}3$ )
Decoder transitions, coarse to fine	$32\!\to\!16\!\to\!16$ at $H/4$ ; $32\!\to\!16\!\to\!16$ at $H/2$ ; $32\!\to\!16\!\to\!16$ at $H$	$384\!\to\!192\!\to\!192$ at $H/8$ ; $288\!\to\!128\!\to\!128$ at $H/4$ ; $192\!\to\!96\!\to\!96$ at $H/2$ ; $128\!\to\!64\!\to\!64$ at $H$

Appendix

Table A7: Detailed reconstruction-network architectures. A decoder transition

a\!\to\!b\!\to\!c

denotes the channel count after skip concatenation and the outputs of the two subsequent

3\times 3

convolutions. Both variants use concatenative encoder skips, a projected full-resolution input skip, max-pool downsampling, nearest-neighbor upsampling, and direct RGB prediction. Parameter counts include convolutional biases. FP16 kernel-weight sizes exclude biases, activations, and runtime workspace.

Figure A4: Temporal ablations in the evaluated stochastic-input setting. Backward reprojection exhibits structured artifacts, while forward reprojection leaves unfilled pixels.

Figure A5: Qualitative results of additional scenes of 16spp + Large network showing neural reconstruction quality

Explore similar work

TemporalGS: Training-Free Plug-and-Play Acceleration for 3D Gaussian Splatting Rendering via Temporal Priors
Jul 3, 2026Yuhongze Zhou, Zihao Yang, Xinxin Zuo +1Gaussian Splatting Training-Free
McGill University, Montr´eal, Canada · University of Waterloo, Waterloo, Canada · Concordia University, Montr´eal, Canada +1
BlitzGS: City-Scale Gaussian Splatting at Lightning Speed
May 13, 2026Zhongtao Wang, Huishan Au, Yilong Li +6Gaussian Primitives Speeds
Peking University, China
EffGS: Efficient and High-Fidelity Gaussian Splatting
Sep 30, 2026Changbai Li, Shuo Yang, Yichen Yang +2Gaussian Splatting Novel View Synthesis
Beihang University · Nanyang Technological University

Jul 3, 2026, cs.CVJ/K move · Enter open · S save

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

Yuhongze Zhou, Zihao Yang, Xinxin Zuo, Juwei Lu

McGill University, Montr´eal, Canada · University of Waterloo, Waterloo, Canada · Concordia University, Montr´eal, Canada+1

Gaussian Splatting Training-Free

Open paper AskPDF arXiv

3D Gaussian Splatting (3DGS) has revolutionized novel-view synthesis with its fast and high-fidelity rendering. However, rendering at high FPS and low latency across various scenes remains a challenge, especially when large amounts of 3D Gaussian ellipsoids appear in the scene. To address this issue, we introduce TemporalGS, to the best of our knowledge, the first training-free plug-and-play algorithmic approach to accelerate 3DGS rendering without any post-training or post-processing, implemented on top of tile-based software rasterization. The key idea is that, instead of rendering frames independently as 3DGS, we leverage the temporal priors, represented by novel geometry and appearance buffers, etc., to reduce redundancy of Gaussian preprocessing, sorting, and rasterization operations of consecutive frames. Specifically, we propose two acceleration strategies: (1) temporal dynamic culling, which filters out Gaussians that contribute less to current frame rendering; (2) selective rendering, which renders only a small portion of tiles that cannot be approximated by the temporal priors. By adapting and interleaving these two strategies, TemporalGS yields a simple but effective plug-and-play solution for 3DGS rendering speed-up without any training. Extensive experiments show that TemporalGS achieves comparable or even better performance compared to existing state-of-the-art post-training or post-processing-based 3DGS rendering acceleration approaches. TemporalGS can significantly enhance the rendering speed of various 3DGS methods, achieving up to

1.48\times

acceleration, while maintaining competitive rendering quality. We further extend our TemporalGS to hardware rasterization-based 3DGS to show the portability of our algorithm.

Abstract

Explore similar work

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

BlitzGS: City-Scale Gaussian Splatting at Lightning Speed

EffGS: Efficient and High-Fidelity Gaussian Splatting