DenoiseGS: 3D Gaussian Splatting for Burst Denoising

• DenoiseGS is a burst denoising framework that utilizes feed-forward 3D Gaussian reconstruction to efficiently restore images and generate novel views under challenging noisy conditions.
• It introduces a Gaussian self-consistency loss to stabilize 3D geometry and a log-weighted frequency loss to emphasize high-frequency details for improved image quality.
• Empirical results show DenoiseGS achieves superior restoration metrics and roughly a 250× speedup over NeRF-based methods, addressing key issues in traditional burst processing.

Searching arXiv for the relevant DenoiseGS paper and closely related Gaussian-splatting denoising work. DenoiseGS is a burst denoising framework that leverages 3D Gaussian Splatting (3DGS) for image restoration and novel view synthesis under noisy conditions. It is built on top of GS-LRM and targets two difficulties that arise when a feed-forward Gaussian reconstruction model is applied to noisy bursts: the degradation of Gaussian point clouds and the loss of fine details. To address these issues, the method introduces a Gaussian self-consistency (GSC) loss, which regularizes geometry predicted from noisy inputs using high-quality Gaussian point clouds generated from clean inputs by the same model, and a log-weighted frequency (LWF) loss, which strengthens spectral supervision and emphasizes difficult high-frequency content. The reported result is a 3DGS-based burst denoising system that exceeds state-of-the-art NeRF-based methods on burst denoising and novel view synthesis under noisy conditions while achieving about $250\times$ faster inference speed (Cheng et al., 28 Nov 2025).

1. Problem setting and position within burst denoising

Burst denoising uses multiple short-exposure images of the same scene to recover a clean target image. The method is motivated by the observation that neighboring frames contain complementary information, but that practical burst denoising remains difficult because large motion and parallax make 2D alignment or optical flow unreliable, heavy noise degrades both features and geometry estimation, and NeRF-based burst denoising methods are computationally expensive because of volume rendering. DenoiseGS is introduced against this background as the first framework to bring 3D Gaussian Splatting into burst denoising (Cheng et al., 28 Nov 2025).

The central architectural choice is to replace slow NeRF-style rendering with feed-forward Gaussian reconstruction. The paper identifies feed-forward Gaussian reconstruction models such as GS-LRM as particularly attractive because they can reconstruct scenes in a single forward pass without per-scene optimization. However, directly applying such models to noisy burst inputs causes Gaussian point clouds to become degraded under noise and causes fine details to be lost, especially at high noise levels. DenoiseGS is therefore framed not merely as a faster renderer, but as a method for preserving 3D structure and detail in the presence of burst noise.

A key conceptual consequence is that DenoiseGS treats burst denoising as a multi-view reconstruction problem rather than a framewise 2D restoration problem. The paper explicitly attributes two advantages to this formulation: robustness to large motion, because a 3D representation naturally handles parallax and misalignment better than 2D optical-flow-based methods, and efficiency, because Gaussian splatting is much faster than NeRF-style volume rendering.

2. Gaussian reconstruction model and scene parameterization

DenoiseGS is built on top of GS-LRM. Its processing pipeline takes as input a burst of noisy images together with camera conditions, embeds the images and camera cues, processes them with Transformer blocks, predicts Gaussian parameters for each pixel, constructs a 3D Gaussian point cloud, and then uses differentiable 3DGS rendering to generate either the denoised target view or novel views under noisy conditions (Cheng et al., 28 Nov 2025).

For each pixel $p$, the predicted Gaussian primitive is parameterized by depth $d_p \in \mathbb{R}$, scale $s_p \in \mathbb{R}^3$, rotation quaternion $r_p \in \mathbb{R}^4$, opacity $\alpha_p \in \mathbb{R}$, and RGB color $c_p \in \mathbb{R}^3$. The Gaussian center is given by

$$\boldsymbol{\mu}_p = \mathbf{o} + d_p \mathbf{d}_p,$$

where $\mathbf{o}$ is the camera origin and $\mathbf{d}_p$ is the ray direction through pixel $p$0. In this construction, each pixel corresponds to a Gaussian primitive, so the model reconstructs a 3D Gaussian scene representation directly from the burst and then renders a clean image from that representation.

This representation choice is central to how the method handles burst noise. Rather than denoising each frame independently in image space, DenoiseGS predicts a 3D Gaussian point cloud and performs differentiable Gaussian rasterization. The paper presents this as the mechanism that enables both geometric robustness and fast inference. A plausible implication is that the denoising objective is coupled to scene reconstruction quality rather than restricted to per-pixel averaging or alignment.

3. Gaussian self-consistency loss

The first principal innovation of DenoiseGS is the Gaussian self-consistency loss. Its purpose is to stabilize geometry predicted from noisy inputs without relying on an external depth estimator. During training, the model receives both noisy bursts and corresponding clean bursts. The clean bursts are passed through the same network, producing a higher-quality Gaussian point cloud, and that clean-generated point cloud is then used as self-guidance for the noisy branch (Cheng et al., 28 Nov 2025).

The Gaussian point cloud predicted from noisy images is written as

$p$1

and the Gaussian point cloud predicted from clean images is

$p$2

The GSC loss is

$p$3

where $p$4 is the stop-gradient operator.

The paper emphasizes that gradients from this loss flow only through the noisy branch. The clean branch is not updated by the self-consistency objective. This design makes the noisy-predicted depths move toward clean-predicted depths rather than forcing both branches toward a shared compromise. The stated rationale for this strategy is twofold. First, an external depth estimator introduces bias from the external model. Second, there is a domain gap because the geometry rendered by GS-LRM for denoising is not necessarily true scene depth; it is a scene representation optimized for image synthesis. The paper therefore characterizes same-model clean-input guidance as intrinsic, task-consistent, and less prone to bias or domain mismatch than supervision from an external model.

Only the depth attribute is used in the loss. The paper explicitly notes that, although it is possible in principle to supervise scale, rotation, opacity, or color, adding multiple attributes can destabilize training. Depth is identified as the most important factor for preserving the 3D structure of the Gaussian point cloud.

A practical training detail is the warm-up phase. The paper states that GSC is not applied at the start of training, because early in optimization the model may not yet reconstruct better Gaussians from clean inputs than from noisy inputs. GSC is activated only after the model has learned a reasonable reconstruction from clean data, so that the self-guidance is meaningful.

4. Log-weighted frequency supervision

The second principal innovation is the log-weighted frequency loss, introduced to preserve fine details that may still be lost even when geometry is improved. The paper motivates this loss by observing that low-frequency content is usually easy to reconstruct, whereas high-frequency edges and textures are harder and therefore require stronger supervision (Cheng et al., 28 Nov 2025).

Given the ground-truth image $p$5 and predicted image $p$6, the paper defines their discrete Fourier transforms as

$p$7

The spectral discrepancy at frequency $p$8 is

$p$9

and the logarithmic weighting is

$d_p \in \mathbb{R}$0

The resulting LWF loss is

$d_p \in \mathbb{R}$1

The role of the logarithmic weighting is described as adaptive and less explosive than a plain frequency loss while still focusing strongly on difficult components. The paper attributes two effects to this design: it encourages reconstruction of high-frequency details such as edges, texture, and thin structures, and it suppresses spurious high-frequency artifacts rather than only filling in missing detail. This suggests that LWF functions both as a detail-preservation term and as a spectral regularizer.

5. Training data, noise model, and optimization pipeline

The training set used for DenoiseGS is RE10K-N, which is created from RealEstate10K. The construction procedure consists of taking clean images, linearizing them to match the model’s linear RGB assumptions, adding synthetic noise using the KPN noise model,

$d_p \in \mathbb{R}$2

and then delinearizing the noisy images back to sRGB for model input (Cheng et al., 28 Nov 2025).

The overall training objective combines image-space and geometry-aware terms: $d_p \in \mathbb{R}$3 The reported fixed weights are $d_p \in \mathbb{R}$4, $d_p \in \mathbb{R}$5, $d_p \in \mathbb{R}$6, and $d_p \in \mathbb{R}$7.

The main training configuration uses burst size $d_p \in \mathbb{R}$8, starts GSC after $d_p \in \mathbb{R}$9k iterations, and trains for $s_p \in \mathbb{R}^3$0k iterations. Optimization uses Adam and FlashAttention via xFormers, and training is conducted on a single NVIDIA RTX A6000 at resolution $s_p \in \mathbb{R}^3$1. At inference time, the model runs in a single forward pass, reconstructs Gaussian point clouds, and renders the target denoised image or a novel view; no scene-specific optimization is required.

These design choices are closely tied to the method’s practical objective. The paper’s argument is not only that Gaussian splatting is faster than volume rendering, but that feed-forward Gaussian reconstruction can retain this efficiency even when extended to burst denoising, provided that geometry and frequency supervision compensate for the effects of noisy inputs.

6. Reported performance, efficiency, and ablation findings

On RE10K-N, DenoiseGS is reported to outperform 2D methods such as KPN, BPN, and Deeprep, the 3D method NAN, and the GS-LRM baseline. At gain $s_p \in \mathbb{R}^3$2, the burst denoising results are reported as PSNR $s_p \in \mathbb{R}^3$3, SSIM $s_p \in \mathbb{R}^3$4, and LPIPS $s_p \in \mathbb{R}^3$5 for DenoiseGS, compared with PSNR $s_p \in \mathbb{R}^3$6, SSIM $s_p \in \mathbb{R}^3$7, and LPIPS $s_p \in \mathbb{R}^3$8 for GS-LRM. Across all gain levels, the paper states that DenoiseGS achieves the best or near-best restoration metrics (Cheng et al., 28 Nov 2025).

The method is also evaluated on novel view synthesis under noisy conditions, where the target view is excluded from the input. At gain $s_p \in \mathbb{R}^3$9, DenoiseGS reports PSNR $r_p \in \mathbb{R}^4$0, SSIM $r_p \in \mathbb{R}^4$1, and LPIPS $r_p \in \mathbb{R}^4$2, compared with PSNR $r_p \in \mathbb{R}^4$3, SSIM $r_p \in \mathbb{R}^4$4, and LPIPS $r_p \in \mathbb{R}^4$5 for GS-LRM. The paper presents this as evidence that the method improves not only denoising of the target frame but also generation of clean novel views from noisy burst inputs.

The efficiency claim is a major part of the method’s empirical profile. DenoiseGS is reported at about $r_p \in \mathbb{R}^4$6 FPS, whereas NAN is reported at about $r_p \in \mathbb{R}^4$7 FPS, corresponding to roughly a $r_p \in \mathbb{R}^4$8 speedup. The paper attributes this difference to the use of feed-forward 3D Gaussian Splatting rather than NeRF volume rendering.

Qualitatively, the paper reports that competing methods often suffer from residual noise, over-smoothing, and floaters or other artifacts, whereas DenoiseGS produces sharper details, cleaner outputs, and fewer geometric artifacts. The ablation study further reports that GSC gives the biggest gain, warm-up is necessary, LWF improves detail preservation, RPPC improves geometric conditioning, and larger burst sizes improve quality but reduce FPS. The paper also compares GSC with an alternative Gaussian Loss based on external GS-LRM supervision and reports similar performance, while noting that GSC avoids reliance on a pretrained external model.

7. Scope, nomenclature, and relation to adjacent work

DenoiseGS specifically denotes the burst denoising method based on feed-forward Gaussian reconstruction described above. It should be distinguished from adjacent denoising-oriented Gaussian-splatting methods that operate in different regimes. “Denoising-GS: Gaussian Splatting with Spatial-aware Denoising” formulates static-scene 3DGS optimization as a primitive denoising process and jointly refines the mean and covariance of each primitive for novel view synthesis (Zhou et al., 14 May 2026). “DN-4DGS: Denoised Deformable Network with Temporal-Spatial Aggregation for Dynamic Scene Rendering” addresses dynamic scenes by denoising canonical Gaussian coordinates before and during deformation through a Noise Suppression Strategy and a Decoupled Temporal-Spatial Aggregation Module (Lu et al., 2024). A separate line of work, “Denoising Monte Carlo Renders with Diffusion Models,” uses a pretrained pixel-space diffusion model with ControlNet-like conditioning for Monte Carlo render denoising rather than Gaussian scene reconstruction (Vavilala et al., 2024).

This distinction matters because the denoising target, the representation, and the supervision differ substantially across these methods. DenoiseGS is centered on noisy handheld-style bursts and the reconstruction of a 3D Gaussian point cloud from them; Denoising-GS is concerned with denoising Gaussian primitives during static 3DGS optimization; DN-4DGS focuses on noisy canonical coordinates in dynamic deformation pipelines; and diffusion-based Monte Carlo denoising addresses physically based rendering noise in image space. Within that landscape, DenoiseGS occupies the specific niche of burst denoising and noisy-condition novel view synthesis using a fast feed-forward 3DGS reconstruction model.