Ultra-fast Neural Inference for Stochastic Gaussian Splatting Denoising

Chenxiao Hu; Hao Zhang; Yanchen Zhang; Meng Gai; Guoping Wang; Sheng Li

cs.CVSep 22, 2026

Ultra-fast Neural Inference for Stochastic Gaussian Splatting Denoising

Authors: Chenxiao Hu, Hao Zhang, Yanchen Zhang, Meng Gai, Guoping Wang, Sheng Li

Organizations: School of Computer Science, Peking University

Abstract

Stochastic rendering eliminates the sorting and alpha blending process in Gaussian splatting, at the cost of introducing spatial noise. Formulating temporal denoising over the pixel stream shared by view-consistent stochastic splatting renderers, we propose a temporal neural denoiser validated on stochastic 2D Gaussian Splatting rendering, combining dual-path exponential moving average accumulation, per-pixel learned trust prediction for history validation, a fixed anisotropic spatial filter and a variance-gated composition with stabilization. The denoiser suppresses the noise, achieving temporally stable, visually compelling outputs during free camera navigation, all while retaining the sort-free, blend-free rasterization performance. The combined pipeline retains a PSNR gap to sorted alpha-blending renderers, but the denoiser's overhead stays below the time saved by removing sorting and blending.

Figures & tables

Figure 1: Top left: Garden navigation. Bottom left: zoom-in by conventional $\alpha$ -blended 2DGS renderer, by stochastic rasterization with ST-TAA (see Sec. 5 ), and by our approach on the same setting. ST-TAA suffers from residual Monte Carlo noise, whereas our method visually resembles the conventional renderer. Right: per-frame time breakdown. Despite the additional cost of denoising, our approach achieves significant speedup by avoiding the expensive Gaussian depth sort and $\alpha$ -blending required by conventional splatting.

Figure 2: Workflow of our denoiser for each frame (Eq. ( 1 )) with trainable modules highlighted. The host stochastic rasterizer emits the 1 spp stream (Sec. 3.2 ). In the accumulated path (left), the previous accumulated state is reprojected into the current view and a convolutional predictor estimates the history trust $t_{s}$ from reprojection-derived motion, depth, and photometric cues. In the denoised path (middle), the current frame is spatially filtered by a fixed anisotropic mipmap kernel parameterised by smoothed neural-view scalars, and a lightweight head estimates the trust $t_{h}$ of the reprojected denoised mean. In postprocessing (right), both paths update their exponential moving average (EMA) states (Eq. ( 2 )), a variance gate (Eq. ( 3 )) composites the two means per pixel, and a stabilization pass ( STAB ) suppresses residual jitter. The updated states propagate to the next frame.

Figure 3: The two EMA paths and their composition at 4 (top) and 64 (bottom) accumulated frames under slow lateral camera motion; the zoomed region is marked on the reference. The accumulated path is noisy at low frame counts but improves with history, while the denoised path is stable but blurry; the variance gate blends the two adaptively (white: prefer denoised, black: prefer accumulated). At 4 frames the not-yet-calibrated gate dips the composite marginally below the denoised path; the dip vanishes within a few frames. PSNR is computed on the zoomed region.

	MipNeRF360			TanksAndTemples
Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
2DGS	29.45	0.875	0.199	21.50	0.700	0.379
Ref. (1024)	28.25	0.851	0.202	20.69	0.675	0.388
ST-TAA (128)	27.92	0.822	0.219	20.61	0.635	0.431
Ours (16)	26.89	0.741	0.330	20.33	0.548	0.496
Ours (98)	27.94	0.822	0.226	20.60	0.631	0.436

Table 1: Static quality, averaged per dataset; per-scene PSNR is reported in the appendix. Vanilla 2DGS is the deterministic tile renderer ( Huang et al. 2024 ) . Ref. (1024) is the Monte-Carlo accumulation reference (Sec. 5 ). ST-TAA is the adapted StochasticSplats baseline (Sec. 5 ), accumulated over 128 static frames. Ours renders through the learned denoiser at 16, 98, or 128 accumulated frames (spp counts in parentheses); the 98 spp point approximately matches ST-TAA (128) in wall-clock time, making it comparable (Table 3 ); bold marks the better of this equal-time pair in each column (both when equal after rounding). Ours (z.s.) deploys the room checkpoint on every scene with no per-scene denoiser training (zero-shot transfer) (Sec. 5.5 ).

Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	CVVDP $\uparrow$	ms $\downarrow$
Ref. (1024)	32.96	0.950	0.080	8.73	–
Raw 1 spp	18.26	0.282	0.671	4.94	2.71
SS 1 spp †	17.86	0.292	0.659	4.67	3.2 †
SS 4 spp †	22.54	0.526	0.540	6.87	6.5 †
ST-TAA	21.48	0.418	0.615	5.95	2.88
SVGF	25.57	0.785	0.297	6.29	2.86

Table 2: Quality under free camera navigation on MipNeRF360, averaged over all scene tracks (per-track numbers in the appendix), using the 2DGS tile renderer as the reference. Raw 1 spp is the single-sample stochastic output; ST-TAA is the adapted StochasticSplats baseline, temporal accumulation without learned trust or spatial filtering; SVGF is the adapted variance-guided filtering baseline (see appendix). Ref. (1024) is the per-frame Monte-Carlo ceiling, not a practical method. The last column is the mean per-frame time over the seven tracks on an RTX 3090. † SS 1 spp/4 spp are the official StochasticSplats model run on its own finetuned 3DGS scenes over the same camera tracks for 1 or 4 spp, measured against its own 1024 spp accumulation as the reference (no ground truth exists on free tracks); it is listed to show that the 1 spp noise level is inherent with the stochastic formulation. Its per-frame error distribution closely tracks our Raw 1 spp row (per-frame PSNR correlation

r=0.90

, and its CVVDP score likewise matches our Raw 1 spp row). SS timings come from the official viewer. Ours (z.s.) deploys the room checkpoint on every scene with no per-scene denoiser training (Sec. 5.5 ).

Panel of Figure 4 — Figure 4: Garden (top) and Room (bottom) navigation tracks. Left: vanilla-rendered reference crop with the zoom region marked; right: the region for each renderer.

Method (ms/frame)	7-track mean $\downarrow$
2DGS Huang et al. (2024)	17.96
2DGS (hardware raster.)	7.99
Raw 1 spp	2.71
+ TAA (ST-TAA)	2.88
Ours (full pipeline)	3.75

Table 3: Per-frame rendering time, averaged over the seven navigation tracks (per-scene numbers in the appendix). The 2DGS row is the official implementation measured with auxiliary channel outputs removed; the hardware raster. row is our hardware-rasterized reimplementation of 2DGS, included as a stronger baseline. The cost of denoiser is nearly constant across scenes.

Variant	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Full model	26.63	0.794	0.238
w/o $t_{s}$ prediction ( $t_{s}{\equiv}1$ )	22.51	0.695	0.323
w/o $t_{s}$ prediction ( $t_{s}=\mathrm{heur.}$ )	25.82	0.755	0.268
w/o spatial filter	23.52	0.676	0.359
w/o stabilization pass	25.94	0.736	0.311

Table 4: Ablations on the garden navigation track (evaluated on every 10th frame of the original 1563 frames, metrics against the vanilla tile renderer). The learned trust prediction is the most impactful single component, followed by the spatial filter. The stabilization pass improves all three metrics and, more importantly, removes most of the visible temporal jitter at the cost of slight ghosting. For the

t_{s}=\mathrm{heur.}

ablation, we replace the neural network with

t_{s}=\exp(-\sigma\cdot\frac{|d_{1}-d_{2}|}{\max(d_{1},d_{2})})

heuristic as a stronger baseline, where

d_{1},d_{2}

are pixel depths of the reprojected/current pixels and

\sigma

is grid-searched on the same track for best performance (oracle).

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Ch.	Input
0	$\|Y_{\text{cur}}-Y_{\text{stoch}}\|$ : luma diff to reprojected accumulated mean
1	reprojected previous $t_{s}$
2	squared compressed motion magnitude
3	accumulated-path history length (normalised)
4	compressed motion magnitude
5	relative depth difference

Appendix

Table 5: Accumulated-path trust predictor (

t_{s}

) inputs, 8 channels.

Ch.	Input
0	$\|Y_{\text{cur}}-Y_{\text{den}}\|$ : luma difference to reprojected denoised mean
1	reprojected denoised-path luma
2	reprojected previous $t_{h}$
3	denoised-path history length (normalised)
4	compressed motion magnitude
5	YCoCg distance to reprojected denoised mean

Appendix

Table 6: Denoised-path trust head (

t_{h}

) inputs, 8 channels.

Stage	Renderer	Optimised	Steps
1	Tile Rasterizer	all	30k
2	Tile Rasterizer	opacity	1k
3	Depth Peeler	full finetune	3k
4	Stochastic Rasterizer	denoiser + $V_{0,1,2,3}$	15k / 4k

Appendix

Table 7: Two-phase training recipe (four stages). Stage 4 trains 15k steps on room and transfers to every other scene with 4k steps, starting from the room denoiser. The 4k transfer is our default per-scene refinement. The room checkpoint can also be deployed with no per-scene denoiser training at some quality cost (main text, Zero-Shot Transfer).

Scene	SVGF	Ours (zero-shot)
bicycle	2.94	3.76
bonsai	2.63	3.44
counter	2.91	3.66
garden	2.97	3.75
kitchen	3.00	3.76
room	3.04	3.79

Appendix

Table 8: Per-scene per-frame rendering time (ms) for the SVGF baseline and the zero-shot variant, measured under the identical protocol as Table 9 . Both consume the same 1 spp stochastic stream; SVGF replaces the learned post-processing with hand-tuned variance-guided filtering, and the zero-shot variant skips per-scene denoiser training.

Scene	2DGS	2DGS (hw)	Raw	ST-TAA	Ours	Speedup
bicycle	29.32	13.73	2.86	3.03	3.95	3.5 $\times$
bonsai	17.66	8.14	2.49	2.66	3.52	2.3 $\times$
counter	16.24	6.33	2.74	2.89	3.75	1.7 $\times$
garden	16.27	6.57	2.79	2.97	3.85	1.7 $\times$
kitchen	10.38	4.49	2.84	3.01	3.86	1.2 $\times$
room	13.94	6.18	2.89	3.04	3.88	1.6 $\times$

Appendix

Table 9: Per-scene per-frame rendering time (ms) on the navigation tracks. 2DGS is the official CUDA tile renderer with auxiliary channels removed; 2DGS (hw) is our hardware-rasterized reimplementation of the same sorted alpha-blending algorithm; Raw is 1 spp stochastic rasterization; ST-TAA adds the baseline temporal accumulation; Ours is the full denoising pipeline. Speedup is 2DGS (hw) over Ours.

Figure 5: Static-view convergence on garden. PSNR versus the number of accumulated frames on a static camera, averaged over the 24 held-out views. The dashed line marks the deterministic vanilla 2DGS renderer, which renders each frame identically and independently of accumulation.

Variant	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Full model (learned $g$ )	26.63	0.794	0.238
$g\equiv 0$ (accumulated path only)	25.14	0.751	0.298
$g\equiv 1$ (denoised path only)	25.95	0.752	0.247

Appendix

Table 10: Variance-gate ablation on the garden navigation track (same protocol as the ablation table in the paper).

g

is the per-pixel weight of the denoised path in the final composition in the main text; forcing it to either path exclusively degrades all three metrics. The learned gate improves over the better fixed extreme by 0.68 dB PSNR.

	PSNR $\uparrow$ per scene							Average
Method	bicycle	bonsai	counter	garden	kitchen	room	stump	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
2DGS	24.71	32.12	28.82	26.73	32.49	31.53	29.74	29.45	0.875	0.199
Ref. (1024)	24.20	30.65	28.14	26.15	30.24	30.51	27.88	28.25	0.851	0.202
ST-TAA (128)	24.08	30.10	27.83	25.86	29.76	30.14	27.63	27.92	0.822	0.219
Ours (16)	23.58	28.80	26.90	24.69	28.40	29.34	26.54	26.89	0.741	0.330
Ours (98)	24.09	30.20	27.84	25.82	29.80	30.22	27.58	27.94	0.822	0.226

Appendix

Table 11: Static quality on the MipNeRF360 held-out test views. 2DGS is the deterministic tile renderer of 2D Gaussian Splatting (Huang et al.). ST-TAA is the baseline adapted from StochasticSplats (Kheradmand et al.): temporal accumulation over 128 static frames without a learned denoiser. Ref. (1024) is the Monte-Carlo accumulation reference (main text, Experimental Setup). Ours renders through the learned denoiser at 16, 98, or 128 accumulated frames; Ours (zero-shot) deploys the room checkpoint on every scene with no per-scene denoiser training (main text, Zero-Shot Transfer). Per-scene PSNR with dataset averages of all metrics; bold marks the better of the equal-time pair, ST-TAA (128) vs. Ours (98), in each column (both when equal after rounding).

	PSNR $\uparrow$ per scene							Average
Method	Barn	Caterpillar	Church	Courthouse	Ignatius	Meetingroom	Truck	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
2DGS	22.82	20.33	21.52	21.73	19.78	22.68	21.65	21.50	0.700	0.379
Ref. (1024)	22.39	19.84	19.87	21.17	19.11	21.42	21.04	20.69	0.675	0.388
ST-TAA (128)	22.35	19.75	19.75	21.11	19.05	21.30	20.96	20.61	0.635	0.431
Ours (16)	22.07	19.46	19.42	20.83	18.79	21.01	20.72	20.33	0.548	0.496
Ours (98)	22.32	19.74	19.75	21.09	19.03	21.32	20.98	20.60	0.631	0.436

Appendix

Table 12: Static quality on TanksAndTemples. All methods are trained on the full capture of each scene; we evaluate every 8th frame. Rows follow Table 11 ; bold marks the better of the equal-time pair, ST-TAA (128) vs. Ours (98), in each column (both when equal after rounding). Ours (zero-shot) deploys the room checkpoint with no per-scene denoiser training (main text, Zero-Shot Transfer).

	PSNR $\uparrow$ / LPIPS $\downarrow$					SSIM $\uparrow$ / CVVDP $\uparrow$
Track	1 spp	ST-TAA	SVGF	Ref.	Ours	1 spp	ST-TAA	SVGF	Ref.	Ours
bicycle (161 frames)	16.81	19.82	22.05	29.87	26.39	0.289	0.407	0.709	0.925	0.807
	0.639	0.588	0.346	0.088	0.284	4.28	5.29	5.23	8.34	6.84
bonsai (166 frames)	19.28	23.00	28.17	34.54	31.98	0.313	0.486	0.868	0.967	0.914
	0.690	0.629	0.241	0.060	0.253	5.57	6.66	7.23	8.96	8.25
counter (126 frames)	18.05	20.97	23.98	33.76	30.35	0.257	0.367	0.766	0.956	0.880

Appendix

Table 13: Track-averaged quality under free camera navigation on MipNeRF360, using the vanilla 2DGS tile renderer as the reference. Each track occupies two rows: the first reports PSNR (left) and SSIM (right), the second LPIPS (left) and CVVDP (right). Raw 1 spp is the single-sample stochastic output; ST-TAA applies generic temporal accumulation without learned trust or spatial filtering; SVGF is the adapted variance-guided filtering baseline (Sec. E ); Ref. (1024) is the per-frame Monte-Carlo ceiling (main text, Experimental Setup), not a practical method. The generic TAA warp fails to accumulate reliably under the heavy 1 spp noise, while our learned trust gating keeps the temporal history usable.

	PSNR $\uparrow$		SSIM $\uparrow$		LPIPS $\downarrow$		CVVDP $\uparrow$
Track	SS 1 spp	SS 4 spp	SS 1 spp	SS 4 spp	SS 1 spp	SS 4 spp	SS 1 spp	SS 4 spp
bicycle	15.70	20.08	0.275	0.483	0.634	0.523	3.65	6.18
bonsai	18.84	23.68	0.314	0.557	0.689	0.580	5.21	7.36
counter	18.13	22.60	0.273	0.475	0.702	0.597	4.31	6.47
garden	15.90	20.78	0.273	0.528	0.590	0.449	4.54	6.84
kitchen	18.83	24.48	0.313	0.594	0.641	0.488	5.62	7.71

Appendix

Table 14: Per-track breakdown of the official StochasticSplats rows of the main text’s navigation table: the official model on its own finetuned 3DGS scenes, measured against its own 1024 spp accumulation (no ground truth exists on free camera tracks). SS 4 spp renders through the official viewer’s multi-sample path with a small fix so that each MSAA sample draws an independent stochastic sample (as released, all samples share one draw). All metrics are computed on every 10th track frame (the stride-10 dumps underlying the corresponding main-table rows; ColorVideoVDP uses the same stride-10 videos).

	PSNR $\uparrow$		SSIM $\uparrow$		LPIPS $\downarrow$		CVVDP $\uparrow$
Track	Ours	zero-shot	Ours	zero-shot	Ours	zero-shot	Ours	zero-shot
bicycle	26.39	25.86	0.807	0.817	0.284	0.255	6.84	6.56
bonsai	31.98	30.91	0.914	0.914	0.253	0.205	8.25	7.83
counter	30.35	29.68	0.880	0.891	0.271	0.248	7.50	7.32
garden	26.63	26.82	0.794	0.808	0.238	0.235	7.15	7.12
kitchen	30.95	30.61	0.906	0.914	0.192	0.162	8.07	8.07

Appendix

Table 15: Per-track breakdown of the zero-shot transfer experiment (main text, Zero-Shot Transfer): the room checkpoint deployed verbatim on every scene, with the neural-view payload neutralised (the trust feature

V_{2}

is set to zero and the spatial filter falls back to a fixed isotropic kernel,

x=y=2

and

\rho=0

in the covariance parameterisation of the main text). Ours is the per-scene fine-tuned configuration of the main navigation table. The zero-shot variant trades a small PSNR/CVVDP margin for slightly better SSIM/LPIPS.

Parameter	Value
finetune steps	15k (4k for fast transfer)
loss	L1 + $0.05$ SSIM + $0.01$ LPIPS
auxiliary vanilla-colour loss (accumulated path)	$0.2$
temporal step ratio	$0.7$
stochastic suffix length $K$	$8$
chain-length curriculum	$2\to 128$

Appendix

Table 16: Stage-4 training configuration.

Image Denoising

Image denoising removes unwanted noise while preserving the edges, textures, and colors needed for faithful images.

See all 157 papers

SteadySplats: Resampling of Low-Variance Gaussians for High-Fidelity Stochastic Rendering
Oct 4, 2026Felix Windisch, Thomas Köhler, Lukas Radl +43D Gaussian Splatting 3DGS Acceleration
Graz University of Technology · University of Copenhagen
Gaussian Stippling: Efficient Sorting-Free 3D Gaussian Rendering through Hybrid Sampling and Spatiotemporal Reconstruction
Sep 29, 2026Zijian Huang, Suiliang Mai, Chuankun Zheng +23D Gaussian Splatting 3DGS Acceleration
State Key Laboratory of CAD&CG, Zhejiang University, Hangzhou, China
TemporalGS: Training-Free Plug-and-Play Acceleration for 3D Gaussian Splatting Rendering via Temporal Priors
Jul 3, 2026Yuhongze Zhou, Zihao Yang, Xinxin Zuo +13D Gaussian Splatting 3DGS Acceleration
McGill University, Montr´eal, Canada · University of Waterloo, Waterloo, Canada · Concordia University, Montr´eal, Canada +1

Oct 4, 2026J/K move · Enter open · S save

SteadySplats: Resampling of Low-Variance Gaussians for High-Fidelity Stochastic Rendering

Felix Windisch, Thomas Köhler, Lukas Radl, Chris Wyman+3

Graz University of Technology · University of Copenhagen

3D Gaussian Splatting 3DGS Acceleration Variance Reduction

Open paper AskPDF arXiv

Stochastic order-independent transparency enables efficient and elegant rendering of primitive-based radiance fields like 3D Gaussian Splatting models, but remains impractical due to the inherent visible noise in the output. We propose a principled approach to minimize high-frequency noise, addressing its sources at the representation and image synthesis level. During stochastic rendering, our history-based spatial resampling scheme drastically accelerates image convergence, while temporal importance resampling ensures coherence under camera movement. During training, a color regularizer implicitly reduces the variance along view rays in the 3DGS models. With these properties, our optimized, Vulkan-based renderer effectively mitigates output noise at low and high sample counts, achieving a substantial 13~dB PSNR increase in quality over previous stochastic methods at 1 sample per pixel and quickly converging to sorted 3DGS with an average L1 error of less than

10^{-4}

Abstract

Image Denoising

Explore similar work

SteadySplats: Resampling of Low-Variance Gaussians for High-Fidelity Stochastic Rendering

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

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