DenoiseSplat: Feed-Forward Gaussian Splatting for Noisy 3D Scene Reconstruction

Fuzhen Jiang; Zhuoran Li; Yinlin Zhang

cs.CVMar 10, 2026

DenoiseSplat: Feed-Forward Gaussian Splatting for Noisy 3D Scene Reconstruction

Authors: Fuzhen Jiang, Zhuoran Li, Yinlin Zhang

Organizations: Hangzhou Dianzi University, Hangzhou, Zhejiang, China

Abstract

3D scene reconstruction and novel-view synthesis are fundamental for VR, robotics, and content creation. However, most NeRF and 3D Gaussian Splatting pipelines assume clean inputs and degrade under real noise and artifacts. We therefore propose DenoiseSplat, a feed-forward 3D Gaussian splatting method for noisy multi-view images. We build a large-scale, scene-consistent noisy--clean benchmark on RE10K by injecting Gaussian, Poisson, speckle, and salt-and-pepper noise with controlled intensities. With a lightweight MVSplat-style feed-forward backbone, we train end-to-end using only clean 2D renderings as supervision and no 3D ground truth. On noisy RE10K, DenoiseSplat outperforms vanilla MVSplat and a strong two-stage baseline (IDF + MVSplat) in PSNR/SSIM and LPIPS across noise types and levels.

Figures & tables

Method	Enc. (ms)	Dec. (ms)	Peak Mem. (GiB)
MVSplat-Noisy	58.12	1.92	1.03
Denoise-Then-MVSplat	86.90	1.92	1.28
Ours	60.74	1.92	1.07

Table 1. Inference efficiency comparison under the RE10K evaluation setting (256

\times

256, 2 context views). Encoder time is measured per scene, while decoder time is measured per rendered target view.

Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
MVSplat-GT(upper bound)	26.38	0.869	0.128
MVSplat-Noisy	24.46	0.702	0.349
Denoise-Then-MVSplat (IDF)	24.77	0.788	0.272
Ours	25.05	0.814	0.260

Table 2. Overall reconstruction performance on the noisy RE10K test set. We report average PSNR

\uparrow

, SSIM

\uparrow

, and LPIPS

\downarrow

over all test views.

Methods	MVSplat-Noisy			Denoise-Then-MVSplat			DenoiseSplat(Ours)
PSNR	SSIM	LPIPS	PSNR	SSIM	LPIPS	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$

$\sigma$ = 0.05	25.19	0.777	0.298	25.28	0.801	0.256	25.44	0.814	0.243
$\sigma$ = 0.08	24.62	0.700	0.365	24.82	0.732	0.300	24.86	0.784	0.282
$\sigma$ = 0.12	23.54	0.612	0.432	24.26	0.665	0.396	24.33	0.742	0.328
$\sigma$ = 0.15	22.79	0.557	0.472	23.14	0.603	0.436	23.87	0.712	0.360

Table 3. Ablation on the standard deviation

\sigma

of Gaussian noise for our DenoiseSplat. We report average performance on the noisy RE10K test set.

Noise type	Parameter setting	PSNR	SSIM	LPIPS
Poisson	scale $=0.03$	24.10	0.740	0.331
Poisson	scale $=0.04$	23.74	0.723	0.350
Speckle	$\sigma=0.02$	25.43	0.846	0.177
Speckle	$\sigma=0.05$	25.36	0.832	0.211
Salt–Pepper	$\alpha=0.01$	24.98	0.816	0.241
Salt–Pepper	$\alpha=0.03$	24.53	0.776	0.298

Table 4. Ablation on the hyper-parameters of other noise types for our DenoiseSplat. Each entry reports performances on the noisy RE10K test set.

Explore similar work

ReconSplat: Generalizable 3D Scene Reconstruction Beyond Observed Views
Aug 28, 2026Giuseppe Stracquadanio, Kevin Raj, Julia Grabinski +1Feed-Forward 3D Reconstruction Novel View Synthesis
TU Darmstadt · Zuse School ELIZA
ViewSplat: View-Adaptive 3D Gaussian Splatting for Feed-Forward Synthesis
Mar 26, 2026Moonyeon Jeong, Seunggi Min, Suhyeon Lee +13D Gaussian Splatting Novel View Synthesis
University of Seoul · Korea Electronics Technology Institute
ZipSplat: Fewer Gaussians, Better Splats
Jun 3, 2026Alexander Veicht, Sunghwan Hong, Dániel Baráth +13D Gaussian Splatting Multi-View 3D Reconstruction
ETH Zürich · Microsoft

Aug 28, 2026J/K move · Enter open · S save

ReconSplat: Generalizable 3D Scene Reconstruction Beyond Observed Views

Giuseppe Stracquadanio, Kevin Raj, Julia Grabinski, Stefan Roth

TU Darmstadt · Zuse School ELIZA

Feed-Forward 3D Reconstruction Novel View Synthesis 3D Diffusion Models 3D Gaussian Splatting

Open paper AskPDF arXiv

We introduce ReconSplat, a feed-forward model for 3D scene reconstruction that aims to address the longstanding trade-off between plausible view generation for unobserved regions and geometric consistency, providing both geometrically aligned novel views and sharp depth estimates. Our approach builds on 3D Gaussian splatting (3DGS) as an intermediate differentiable scene representation and integrates it with a multi-view latent diffusion model (MV-LDM) trained to act simultaneously as a refiner and an inpainter for appearance and scene geometry. We enforce geometric consistency by guiding the diffusion process with variational 3D latent features for appearance and geometry, encoded by the feed-forward 3DGS representation and rasterized to 2D latent space. ReconSplat produces both photorealistic novel views and accurate depth maps on real-world benchmarks, RealEstate10K and DL3DV-10K, outperforming existing methods in challenging extrapolation setups. Notably, ReconSplat allows the extrapolation of unseen and challenging viewpoints jointly with coherent and precise scene geometry.

Abstract

Explore similar work

ReconSplat: Generalizable 3D Scene Reconstruction Beyond Observed Views

ViewSplat: View-Adaptive 3D Gaussian Splatting for Feed-Forward Synthesis

ZipSplat: Fewer Gaussians, Better Splats