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

Jaekyun Ko; Dongjin Kim; Soomin Lee; Guanghui Wang; Tae Hyun Kim

cs.CVMar 5, 2026

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

Authors: Jaekyun Ko, Dongjin Kim, Soomin Lee, Guanghui Wang, Tae Hyun Kim

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

Abstract

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.

Figures & tables

Figure 2: Overview of the proposed method. (a) Training pipeline. (b) Inference pipeline.

Figure 3: (a) Sketch of the Prompt Autoencoder (PAE). (b) Details of Global and Local Prompt Blocks.

Camera	Metrics	C2N	Flow-sRGB	NeCA-W	NAFlow	Ours
G4	KLD	0.1660	0.0507	0.0242	0.0254	0.0174
G4	AKLD	0.2007	0.1504	0.1524	0.1367	0.1283
GP	KLD	0.1315	0.0781	0.0432	0.0352	0.0143
GP	AKLD	0.1968	0.1797	0.1273	0.1180	0.1074
IP	KLD	0.0581	0.5128	0.0410	0.0339	0.0291
IP	AKLD	0.2929	1.7490	0.1145	0.1522	0.1128

Table 1: Quantitative results for synthetic noise on a subset of the SIDD validation set, in which the ISO values exist in the training set. The results are computed with KLD

\downarrow

and AKLD

\downarrow

. The best and second-best results are highlighted in bold and underline , respectively.

Figure 4: Visualization of synthetic noisy images on the SIDD validation set. From left to right: C2N, NeCA-W, NAFlow, Ours (PNG), and real noisy images.

Metrics	C2N	NoiseFlow	Flow-sRGB	NeCA-W	NAFlow	Ours	Real
PSNR $\uparrow$	33.76	33.81	34.74	36.82	37.22	37.55	37.63
SSIM $\uparrow$	0.901	0.894	0.912	0.932	0.935	0.937	0.936

Table 2: Denoising performance of DnCNN on SIDD-Benchmark in terms of PSNR

\uparrow

and SSIM

\uparrow

. All methods are trained with synthetic noisy-clean pairs. Note that Real indicates denoising results by training using real noisy-clean pairs. The best and second-best results are highlighted in bold and underline .

Figure 5: Visual comparison on denoising results with PSNR $\uparrow$ on SIDD validation set from DnCNN trained on each method. For more qualitative results, please refer to the supplementary material.

Model	Methods	PolyU		Nam		SIDD Validation		SIDD+		Average
Model	Methods	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$
DnCNN	Real (100%)	36.34	0.9204	35.35	0.8828	37.72	0.8905	35.68	0.8860	36.27	0.8949
	NeCA-W (100%)	36.96	0.9544	35.61	0.9403	36.26	0.8812	35.42	0.8900	36.06	0.9165
	NAFlow (100%)	36.85	0.9499	37.45	0.9524	37.27	0.8989	36.67	0.9249	37.06	0.9315
	Ours (100%)	37.93	0.9609	38.08	0.9630	37.63	0.8960	36.23	0.9072	37.47	0.9318
	NeCA-W-Mixed (50%)	37.92	0.9601	37.86	0.9580	37.51	0.8952	35.79	0.8961	37.27	0.9274

Table 3: Quantitative comparison of DnCNN denoising performance on the PolyU, Nam, SIDD validation, and SIDD+ benchmarks. The percentage (%) denotes the mixing ratio between the two training subsets. The best and second-best results are highlighted in bold and underline .

Methods	PolyU		Nam		MAI2021		Average
Methods	KLD $\downarrow$	AKLD $\downarrow$	KLD $\downarrow$	AKLD $\downarrow$	KLD $\downarrow$	AKLD $\downarrow$	KLD $\downarrow$	AKLD $\downarrow$
Metadata-Dependent	✗	✗	✗	✗	✗	✗	✗	✗
NAFlow	0.2304	1.7801	0.2055	1.9320	0.7304	2.6465	0.3888	2.1195
Ours	0.0822	0.5814	0.0966	0.4668	0.0747	0.3314	0.0845	0.4599

Table 4: Quantitative results of synthetic noise on the PolyU, Nam, and MAI2021. All methods are trained with SIDD training set. The results are computed with KLD

\downarrow

and AKLD

\downarrow

. The best results are shown in bold .

Test Dataset	Real (100%)		NAFlow-Mixed (50%)		Ours-Mixed (50%)
Test Dataset	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$
Nam	38.87	0.9712	39.69	0.9715	40.40	0.9761

Table 5: Quantitative denoising results on the external dataset.

Classifier	Camera Sensor (%)	Camera Sensor + ISO Level
Classifier	Camera Sensor (%)	Top-1 (%)	Top-3 (%)
Baseline (No Prompt)	75.80	66.27	93.91
GPB	82.37	68.35	94.71
GPB + LPB	94.47	75.48	98.64

Table 6: Quantitative results of metadata classification on SIDD validation with different combinations of prompt features in terms of accuracy

\uparrow

. We test the validity of the prompt features in two aspects: camera sensor type and ISO level. The best and second-best results are highlighted in bold and underlined .

Combination		Generation
GPB	LPB	KLD $\downarrow$	AKLD $\downarrow$
✗	✗	0.6182	0.4387
✓	✗	0.0287	0.1112
✓	✓	0.0261	0.1108

Table 7: Effect of GPB and LPB on SIDD validation noise generation. The best results are shown in bold .

Figure S1 : Overview of Prompt DiT (P-DiT). (a) Overall P-DiT structure. (b) P-DiT block.

Resolution	Inference Time (image / second)
Resolution	NeCA-W	NAFlow	Ours
$256{\times}256$	150	13	57
$512{\times}512$	38	8	21
$1024{\times}1024$	9	4	5

Table S1 : Inference speed comparison between NeCA-W, NAFlow, and PNG.

Conditioning		Generation
Timestep	Attention	KLD↓	AKLD↓
✗	✗	0.5661	0.4132
✓	✗	0.0287	0.1291
✓	✓	0.0261	0.1108

Table S2 : Effect of conditioning features on different components in P-DiT. The best results are shown in bold .

Combination		Reconstruction
GPB	LPB	PSNR $\uparrow$	SSIM $\uparrow$
✗	✗	37.58	0.9800
✓	✗	46.30	0.9982
✓	✓	46.54	0.9983

Table S3 : Effect of GPB and LPB on SIDD validation noisy image reconstruction. The best results are shown in bold .

# Blocks	KLD $\downarrow$
$B=4$	0.0350
$B=6$	0.0345
$B=8$	0.0261

Table S4 : KLD score depending on different number of blocks

B

Figure S2 : Overview of unpaired noise generation process.

# Samples	PolyU		Nam		SIDD Validation		SIDD+		Average
# Samples	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$
$\times$ 1	37.40	0.9569	37.29	0.9578	36.55	0.8887	35.72	0.8997	36.74	0.9258
$\times$ 2	37.61	0.9546	37.92	0.9580	36.95	0.8838	35.87	0.9078	37.09	0.9260
$\times$ 4	37.72	0.9542	37.94	0.9578	37.27	0.8955	36.00	0.9022	37.23	0.9274

Table S5 : Quantitative results on denoising generalization performance depending on the size of the synthesized dataset using PNG. The multiplication sign (

\times

) indicates the scaling applied to the original number of patches. The best and second-best results are denoted as bold and underline , respectively.

Methods	G4		GP		IP		N6		S6		Average
Methods	KLD $\downarrow$	AKLD $\downarrow$	KLD $\downarrow$	AKLD $\downarrow$	KLD $\downarrow$	AKLD $\downarrow$	KLD $\downarrow$	AKLD $\downarrow$	KLD $\downarrow$	AKLD $\downarrow$	KLD $\downarrow$	AKLD $\downarrow$
NeCA-W †	0.0513	0.1765	0.0487	0.1819	0.0258	0.1609	0.0603	0.1496	0.0581	0.2205	0.0489	0.1779
NAFlow †	0.4482	1.6918	0.2260	1.0135	0.3267	1.1612	0.2200	0.8725	0.1462	0.5456	0.2734	1.0569
Ours †	0.0350	0.1567	0.0312	0.1553	0.0222	0.1392	0.0291	0.1229	0.0204	0.1326	0.0276	0.1413

Table S6 : Quantitative results for the noise generation performance on SIDD validation set under unpaired settings. The results are reported using KLD

\downarrow

and AKLD

\downarrow

, with the best and second-best results highlighted in bold and underline , respectively. † indicates that the model is evaluated under an unpaired setting.

Methods	PolyU		Nam		SIDD Validation		SIDD+		Average
Methods	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$
NeCA-W †	37.39	0.9555	37.13	0.9495	32.86	0.7490	33.92	0.8342	35.33	0.8720
NAFlow †	32.40	0.9157	31.73	0.9381	30.89	0.7776	31.60	0.7953	31.66	0.8567
Ours †	38.00	0.9618	38.37	0.9633	32.93	0.7803	34.25	0.8557	35.89	0.8903

Table S7 : Quantitative results on denoising generalization performance trained with synthetic noisy-clean pairs generated in an unpaired manner. The best and second-best results are denoted as bold and underline , respectively. † indicates that noise is generated under an unpaired setting.

Figure S3 : Visualization of synthetic noisy images on the SIDD validation set. From left to right: C2N, NeCA-W, NAFlow, Ours (PNG), and real noisy images.

Figure S4 : Visual comparison on denoising results with PSNR $\uparrow$ on SIDD validation set from DnCNN trained on each method.

Synthetic Data Generation

Synthetic data generation creates artificial examples, simulations, and datasets to train, test, or protect machine-learning systems.

See all 620 papers

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
sRGB Real Noise Modeling via Noise-Aware Sampling with Normalizing Flows
Aug 29, 2026Dongjin Kim, Donggoo Jung, Sungyong Baik +1Image Denoising Normalizing Flows
Dept. of Computer Science, Hanyang University · Dept. of Artificial Intelligence, Hanyang University · Dept. of Data Science, Hanyang University
SCoRe: Clean Image Generation from Diffusion Models Trained on Noisy Images
Apr 10, 2026Yuta Matsuzaki, Seiichi Uchida, Shumpei TakezakiDiffusion-Based Image Generation Diffusion Models
Kyushu University, Fukuoka, Japan

Jul 10, 2026J/K move · Enter open · S save

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

Jaekyun Ko, Byung Wan Lim, Soomin Lee, Dongjin Kim+1

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

Synthetic Data Augmentation Image Denoising

Open paper AskPDF arXiv

Real-world sRGB image denoising remains challenging due to the nonlinear characteristics of sensor noise and the difficulty of acquiring aligned clean-noisy image pairs. Supervised denoisers often overfit to limited paired datasets, while self-supervised methods still depend on sufficiently diverse noisy observations. These limitations motivate scalable noise synthesis methods that can model real-world noise without clean ground truth or camera metadata. We propose YeTI, a real-world sRGB noise generation framework that learns from only two noisy observations of the same scene. YeTI uses a Reconstruction Autoencoder to disentangle scene structure and noise characteristics, and models the latent noise distribution with a one-step Conditional Diffusion Transformer trained using consistency objectives. Given a single noisy input at inference time, YeTI generates realistic, signal-dependent noise while preserving the underlying scene content. Extensive experiments demonstrate the effectiveness of YeTI across real-world benchmarks. We evaluate noise generation on SIDD and further assess generalization on SIDD+, MAI2021, and SID, covering smartphone and diverse consumer-camera sensors. Downstream denoising results on DND further show that denoisers trained with YeTI-synthesized images achieve strong real-world performance, highlighting the practical value of clean-image-free and metadata-free noise generation.

Abstract

Synthetic Data Generation

Explore similar work

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

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

SCoRe: Clean Image Generation from Diffusion Models Trained on Noisy Images