sRGB Real Noise Modeling via Noise-Aware Sampling with Normalizing Flows

Dongjin Kim; Donggoo Jung; Sungyong Baik; Tae Hyun Kim

cs.CVAug 29, 2026

sRGB Real Noise Modeling via Noise-Aware Sampling with Normalizing Flows

Authors: Dongjin Kim, Donggoo Jung, Sungyong Baik, Tae Hyun Kim

Organizations: Dept. of Computer Science, Hanyang University · Dept. of Artificial Intelligence, Hanyang University · Dept. of Data Science, Hanyang University

Abstract

Noise poses a widespread challenge in signal processing, particularly when it comes to denoising images. Although convolutional neural networks (CNNs) have exhibited remarkable success in this field, they are predicated upon the belief that noise follows established distributions, which restricts their practicality when dealing with real-world noise. To overcome this limitation, several efforts have been taken to collect noisy image datasets from the real world. Generative methods, employing techniques such as generative adversarial networks (GANs) and normalizing flows (NFs), have emerged as a solution for generating realistic noisy images. Recent works model noise using camera metadata, however requiring metadata even for sampling phase. In contrast, in this work, we aim to estimate the underlying camera settings, enabling us to improve noise modeling and generate diverse noise distributions. To this end, we introduce a new NF framework that allows us to both classify noise based on camera settings and generate various noisy images. Through experimental results, our model demonstrates exceptional noise quality and leads in denoising performance on benchmark datasets.

Figures & tables

Key features	Flow-sRGB	NeCA	Ours
Sampling without metadata	✗	✗	✓
Single unified model	✓	✗	✓
Correlated noise generation	✗	✓	✓

Table 1: Comparison between sRGB noise modeling networks.

Figure 1: Overview of proposed framework. (a) The training procedure of our NAFlow, which uses conditional normalizing flow network $f$ to map noisy images to Gaussian distributions on the latent space, where each Gaussian distribution corresponds to the camera configuration used for capturing images. (b) The pipeline of inference procedure NAS. We use the Gaussian mixture model to obtain a more accurate latent representation without metadata of a given input noisy images.

Figure 2: Noise visualization from diverse device settings ( e.g., smartphone model, ISO). Each noise has a unique distribution due to the distinct characteristics of each image signal processor.

Camera	Metrics	C2N	Flow-sRGB	NeCA-S	NeCA-W	NAFlow
G4	KLD	0.1660	0.0507	0.4025	0.0242	0.0254
G4	AKLD	0.2007	0.1504	1.6803	0.1524	0.1367
GP	KLD	0.1315	0.0781	0.1713	0.0432	0.0352
GP	AKLD	0.1968	0.1797	0.7379	0.1273	0.1180
IP	KLD	0.0581	0.5128	0.3424	0.0410	0.0339
IP	AKLD	0.2929	1.7490	1.2924	0.1145	0.1522

Table 2: Quantitative results of synthetic noise on SIDD-Validation. The results are computed with KLD

\downarrow

and AKLD

\downarrow

. In the case of NeCA-S and NeCA-W ( Fu et al., 2023 ) , they require five different models for noise generation, one for each smartphone camera. In particular, NeCA-W requires three sub-models (GENet a , NPNet b , NCNet c ) for each camera. Best results are highlighted in bold .

Figure 4: Visualization of synthetic noisy images with different Camera configurations (Camera $\_$ ISO) on SIDD-Validation. From left to right: C2N, NeCA-W, Ours (NAFlow), and Real noisy images.

Metrics	C2N	NoiseFlow	Flow-sRGB	NeCA-S	NeCA-W	NAFlow	Real Noise
PSNR $\uparrow$	33.76	33.81	34.74	36.10	36.82	37.22	37.63
SSIM $\uparrow$	0.901	0.894	0.912	0.927	0.932	0.935	0.935

Table 4: Quantitative results of denoising performance on SIDD Benchmark in terms of PSNR

\uparrow

and SSIM

\uparrow

. All methods are trained under the same dataset conditions (synthetic noisy-clean pairs). Note that Real Noise are trained with real noisy-clean pairs. The best result is highlighted in bold .

Table 9

Figure 6: Denoising results in terms of PSNR $\uparrow$ and SSIM $\uparrow$ on SIDD-Validation from DnCNN trained on each method. For more qualitative results, please refer to the appendix.

Table 7: KLD score conditioning on learnable parameters of Gaussian distribution in Eq. 4 . Training

\mu

Training

\bm{\Sigma}

KLD

\downarrow

✗ ✗ 0.0544 ✓ ✗ 0.0449 ✗ ✓ 0.0340 ✓ ✓ 0.0291

Table 8: KLD score depending on different number of scale-level blocks

L

in Fig. 3 . Scale-level KLD ↓

L=1

0.0349

L=2

0.0313

L=3

0.0291

L=4

N/A

Table 7: KLD score conditioning on learnable parameters of Gaussian distribution in Eq. 4 .

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 7: Forward pass of Conditional Affine Coupling and Affine Injector.

Figure 8: Visualization of synthetic noise images on SIDD-Validation. The images are arranged from left to right in the following order: C2N, NeCA-W, Ours (NAFlow), and Real noisy image. For more detail, noise images are also provided at the bottom of each row.

Figure 9: Denoising results on SIDD-Validation from DnCNN trained on each method.

Figure 10: Generation results using camera configuration of NAFlow. The images is obtained from SIDD+ ( Abdelhamed et al., 2020 ) .

Image Denoising

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

See all 157 papers

Diffusion-Based sRGB Real Noise Generation via Prompt-Driven Noise Representation Learning
Mar 5, 2026Jaekyun Ko, Dongjin Kim, Soomin Lee +2Synthetic Data Generation Diffusion-Based Image Generation
Department of Computer Science, Hanyang University · Mobile Experience (MX) Division, Samsung Electronics · Department of Computer Science, Toronto Metropolitan University
YeTI: You Only Need Two Noisy Images for Real-World sRGB Noise Generation
Jul 10, 2026Jaekyun Ko, Byung Wan Lim, Soomin Lee +2Synthetic Data Augmentation Image Denoising
Department of Computer Science, Hanyang University · Mobile Experience (MX) Division, Samsung Electronics
Noise-Robust Conditional Flow Matching: Generating Clean Samples from Noisy Datasets
Jul 28, 2026Adrian Urbański, Gabriel della Maggiora, Artur YakimovichFlow Matching Image Generation
Center for Advanced Systems Understanding (CASUS), Görlitz, Germany · Helmholtz-Zentrum Dresden-Rossendorf e. V. (HZDR), Dresden, Germany · Institute of Computer Science, University of Wrocław, Wrocław, Poland +2

Mar 5, 2026J/K move · Enter open · S save

Diffusion-Based sRGB Real Noise Generation via Prompt-Driven Noise Representation Learning

Jaekyun Ko, Dongjin Kim, Soomin Lee, Guanghui Wang+1

Department of Computer Science, Hanyang University · Mobile Experience (MX) Division, Samsung Electronics · Department of Computer Science, Toronto Metropolitan University

Synthetic Data Generation Diffusion-Based Image Generation Image Denoising

Open paper AskPDF arXiv

Denoising in the sRGB image space is challenging due to large noise variability. Although end-to-end methods perform well, their effectiveness in real-world scenarios is limited by the scarcity of real noisy-clean image pairs, which are expensive and difficult to collect. To address this limitation, several generative methods have been developed to synthesize realistic noisy images from limited data. These approaches often rely on camera metadata during both training and testing to synthesize real-world noise. However, the lack of metadata or inconsistencies between devices restricts their usability. Therefore, we propose a novel framework called Prompt-Driven Noise Generation (PNG). This model is capable of acquiring high-dimensional prompt features that capture the characteristics of real-world input noise and creating a variety of realistic noisy images consistent with the distribution of the input noise. By eliminating the dependency on explicit camera metadata, our approach significantly enhances the generalizability and applicability of noise synthesis. Comprehensive experiments reveal that our model effectively produces realistic noisy images and show the successful application of these generated images in removing real-world noise across various benchmark datasets.

Abstract

Image Denoising

Explore similar work

Diffusion-Based sRGB Real Noise Generation via Prompt-Driven Noise Representation Learning

YeTI: You Only Need Two Noisy Images for Real-World sRGB Noise Generation

Noise-Robust Conditional Flow Matching: Generating Clean Samples from Noisy Datasets