RealRestorer: Towards Generalizable Real-World Image Restoration with Large-Scale Image Editing Models

Yufeng Yang; Xianfang Zeng; Zhangqi Jiang; Fukun Yin; Jianzhuang Liu; Wei Cheng; jinghong lan; Shiyu Liu

cs.CVMar 26, 2026

RealRestorer: Towards Generalizable Real-World Image Restoration with Large-Scale Image Editing Models

Authors: Yufeng Yang, Xianfang Zeng, Zhangqi Jiang, Fukun Yin, Jianzhuang Liu, Wei Cheng, jinghong lan, Shiyu Liu, +3 more

Organizations: Southern University of Science and Technology · StepFun · Shenzhen University of Advanced Technology

Abstract

Image restoration under real-world degradations is critical for downstream tasks such as autonomous driving and object detection. However, existing restoration models are often limited by the scale and distribution of their training data, resulting in poor generalization to real-world scenarios. Recently, large-scale image editing models have shown strong generalization ability in restoration tasks, especially for closed-source models like Nano Banana Pro, which can restore images while preserving consistency. Nevertheless, achieving such performance with those large universal models requires substantial data and computational costs. To address this issue, we construct a large-scale dataset covering nine common real-world degradation types and train a state-of-the-art open-source model to narrow the gap with closed-source alternatives. Furthermore, we introduce RealIR-Bench, which contains 464 real-world degraded images and tailored evaluation metrics focusing on degradation removal and consistency preservation. Extensive experiments demonstrate our model ranks first among open-source methods, achieving state-of-the-art performance.

Figures & tables

Figure 2: Overview of our large-scale Synthetic Degradation Data pipeline. We construct nine representative degradation types, including blur, compression artifacts, moiré patterns, low-light, noise, flare, reflection, haze, and rain. Compared with previous synthetic-only pipelines, our upgraded framework incorporates granular noise modeling, segment-aware perturbations, and web-style degradation processes, significantly narrowing the gap between synthetic and real-world distributions. This comprehensive pipeline enables more robust and generalizable restoration learning.

Figure 3: Comparison with state-of-the-art image editing models across nine real-world degradations, including blur, compression artifacts, moiré patterns, low-light, noise, flare, reflection, haze, and rain. We compare our method with large-scale image editing models, such as Seedream 4.5, Nano Banana Pro, GPT-Image-1.5, Step1X-Edit, FLUX.1-Kontext-dev, Qwen-Image-Edit-2511, and LongCat-Image-Edit.

Method	Open-source	Rain Removal			Deblurring			Low-light Enhancement			Haze Removal			Reflection Removal
Method	Open-source	LPS $\downarrow$	RS $\uparrow$	FS $\uparrow$	LPS $\downarrow$	RS $\uparrow$	FS $\uparrow$	LPS $\downarrow$	RS $\uparrow$	FS $\uparrow$	LPS $\downarrow$	RS $\uparrow$	FS $\uparrow$	LPS $\downarrow$	RS $\uparrow$	FS $\uparrow$
Image Editing Methods
Nano Banana Pro [ 56 ]	No	0.429	2.063	0.236	0.326	1.068	0.144	0.467	0.720	0.077	0.492	1.920	0.195	0.358	1.368	0.176
GPT–Image–1.5 [ 45 ]	No	0.535	2.120	0.197	0.532	1.667	0.156	0.523	1.048	0.100	0.558	1.840	0.163	0.468	2.320	0.247
Seedream 4.5 [ 55 ]	No	0.438	1.500	0.169	0.254	0.255	0.038	0.423	0.600	0.069	0.418	1.140	0.133	0.291	1.156	0.164
LongCat–Image–Edit [ 57 ]	Yes	0.381	1.302	0.161	0.200	0.000	0.000	0.158	0.120	0.020	0.236	0.000	0.000	0.254	1.060	0.158

Table 1: Quantitative comparison on the Rain Removal, Deblurring, Low-light Enhancement, Haze Removal, and Reflection Removal tasks. We compared state-of-the-art (SOTA) image editing models. For each task, we reported LPS (

\downarrow

), RS (

\uparrow

), and FS (

\uparrow

). The best result is marked in bold , and underline indicates the second-best result. The best and second-best open-source results are highlighted with yellow and blue backgrounds, respectively.

Method	Open-source	Deflare			Moiré Patterns Removal			Denoise			Compression Restoration			Avg Total (9)
Method	Open-source	LPS $\downarrow$	RS $\uparrow$	FS $\uparrow$	LPS $\downarrow$	RS $\uparrow$	FS $\uparrow$	LPS $\downarrow$	RS $\uparrow$	FS $\uparrow$	LPS $\downarrow$	RS $\uparrow$	FS $\uparrow$	LPS $\downarrow$	RS $\uparrow$	FS $\uparrow$
Image Editing Methods
Nano Banana Pro [ 56 ]	No	0.214	1.222	0.192	0.562	1.560	0.137	0.386	0.712	0.087	0.483	1.122	0.116	0.413	1.306	0.153
GPT–Image–1.5 [ 45 ]	No	0.336	1.415	0.188	0.646	1.633	0.116	0.496	0.993	0.100	0.633	1.167	0.086	0.525	1.578	0.150
Seedream 4.5 [ 54 ]	No	0.225	1.104	0.171	0.548	1.600	0.145	0.387	0.770	0.094	0.529	1.136	0.107	0.390	1.029	0.125
LongCat–Image–Edit [ 57 ]	Yes	0.241	1.717	0.261	0.420	1.200	0.139	0.350	0.471	0.061	0.188	0.083	0.014	0.270	0.661	0.097

Table 2: Quantitative comparison on the Deflare, Demoiré, Denoise, and Compression-restoration tasks. The average results of all 9 tasks are reported in the last column. We compared state-of-the-art (SOTA) image editing models. For each task, we reported LPIPS (

\downarrow

), RS (

\uparrow

), and FS (

\uparrow

). The best result is marked in bold , and underline indicates the second-best result. The best and second-best open-source results are highlighted with yellow and blue backgrounds, respectively.

Method	Blur		Rain		Raindrops		Noise		Low-light		Haze		Compression		Average
Method	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$
Nano Banana Pro [ 56 ]	20.16	0.68	21.55	0.65	23.29	0.66	24.78	0.89	15.73	0.73	17.63	0.50	19.48	0.58	20.37	0.67
GPT–Image–1.5 [ 45 ]	13.25	0.41	12.52	0.34	13.89	0.31	14.15	0.68	11.00	0.55	12.39	0.28	13.06	0.39	12.89	0.42
Seedream 4.5 [ 55 ]	17.94	0.61	17.61	0.57	18.01	0.58	15.57	0.78	13.64	0.68	15.65	0.44	14.29	0.54	16.10	0.60
LongCat–Image–Edit [ 57 ]	18.53	0.65	17.09	0.52	18.22	0.48	19.77	0.81	15.24	0.70	10.52	0.36	15.64	0.47	16.43	0.57
Qwen–Image–Edit–2511 [ 65 ]	15.60	0.55	14.68	0.46	15.19	0.42	19.73	0.81	7.13	0.17	7.20	0.34	18.29	0.61	13.98	0.48

Table 3: Quantitative comparison on the FoundIR dataset across various real-world degradations. We report PSNR (

\uparrow

) and SSIM (

\uparrow

). The best results are highlighted in bold , and the second-best results are underlined .

Figure 4: Model performance with varying training steps and training data on RealIR-Bench. The blue line shows transfer training on synthetic degradation data, where the model gradually acquires basic restoration capability. The blue dashed line indicates performance degradation after prolonged training due to the limited diversity of synthetic data. The purple line represents supervised fine-tuning with real-world degradation data, which rapidly improves performance and generalization. The purple dashed segment indicates the onset of overfitting after around 2.5K steps.

Appendix figures & tables16 assets

Supplementary material from the paper’s appendix.

Appendix

Degradation Type	CLIP Text Prompt
Flare	a photo with lens flare, bright streaks of light
Haze	a hazy photo, foggy atmosphere, low contrast
Rain	a rainy photo with rain streaks or raindrops
Low-light	a dark photo, underexposed, low illumination
Blur	a blurry photo with motion blur or out-of-focus regions
Reflection	a photo with glass or mirror-like reflection artifacts

Appendix

Table 4: Semantic prompts used for CLIP-based degradation filtering.

Degradation	Synthetic	Real	Total
Rain	84,968	43,415	128,383
Blur	1,014,229	13,458	1,027,687
Low-light	5,000	7,005	12,005
Hazy	103,971	8,147	112,118
Reflection	68,227	7,604	75,831
Flare	59,520	7,956	67,476

Appendix

Table 5: Statistics of our dataset across different degradation types. The table reports the number of training image pairs from Synthetic Degradation Data and Real-World Degradation Data. The total column indicates the combined number of samples for each degradation category.

Hyperparameters	Transfer Training Stage	Supervised Fine-Tuning Stage
Learning rate	$1\times 10^{-5}$	$1\times 10^{-5}\rightarrow 0$
LR scheduler	Constant	Cosine
Weight decay	0.0	0.01
Gradient norm clip	1.0	1.0
Optimizer	AdamW ( $\beta_{1}=0.9,\beta_{2}=0.95,\epsilon=1e^{-8}$ )
Warm-up steps	100	100

Appendix

Table 6: Training hyperparameters used in the two training stages.

Method	Flare		Moiré		Refl (Postcard)		Refl (Solid)		Refl (Wildscene)		Average
Method	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$
Nano Banana Pro	25.28	0.889	18.86	0.726	20.11	0.821	24.53	0.880	22.27	0.835	22.21	0.830
GPT–Image–1.5	16.86	0.526	11.84	0.549	13.51	0.465	13.97	0.451	14.59	0.527	14.15	0.504
Seedream 4.5	23.88	0.831	15.30	0.643	21.38	0.807	22.16	0.809	21.45	0.783	20.83	0.775
LongCat–Image–Edit	22.57	0.809	15.79	0.670	20.28	0.706	17.12	0.632	21.28	0.805	19.41	0.724
Qwen–Image–Edit–2511	22.45	0.869	16.07	0.673	21.07	0.839	20.17	0.864	23.77	0.896	20.71	0.828

Appendix

Table 7: Quantitative comparison on extra public benchmarks for real-world image restoration tasks. We report PSNR (

\uparrow

) and SSIM (

\uparrow

). The evaluation is conducted on multiple datasets including Flare-R from Flare7K++ for deflare, UHDM for moiré pattern removalremoval, and the SIR 2 + reflection removal benchmark with three subsets: PostcardDataset, SolidObjectDataset, and WildScene. Flare-R contains real captured flare images, while UHDM provides ultra-high-definition paired moiré images. The SIR 2 + subsets represent different reflection scenarios with diverse scene contents. The best results are highlighted in bold , and the second-best results are underlined .

Metric	Correlation Coefficient	$p$ -value
Kendall’s $\tau_{b}$	0.2493	$2.96\times 10^{-124}$
SRCC	0.3010	$4.62\times 10^{-128}$
PLCC	0.2919	$3.21\times 10^{-120}$

Appendix

Table 8: Consistency evaluation between RS metric and subjective human perception.

Figure 5: Examples from our training dataset containing both synthetic and real-world degradation pairs. The upper rows with gray labels show synthesized degradations generated by our pipeline, while the bottom rows highlighted with orange labels correspond to real-world degraded images paired with clean references.

Figure 6: Additional qualitative results of RealRestorer under real-world degradations. Please zoom in for better visualization of details.

Figure 7: Additional qualitative results of RealRestorer under real-world degradations. Please zoom in for better visualization of details.

Figure 8: Examples from our RealIR-Bench. Each degradation category is evaluated using a fixed bilingual prompt.

Panel of Figure 9 — Figure 9: System instruction used for degradation evaluation.

Figure 10: More qualitative comparison results on RealIR-Bench. Zoom in to see more details.

Figure 11: More qualitative comparison results on RealIR-Bench. Zoom in to see more details.

Figure 12: More qualitative comparison results on RealIR-Bench. Zoom in to see more details.

Figure 13: More qualitative comparison results on RealIR-Bench. Zoom in to see more details.

Figure 14: Qualitative comparison of different training strategies. Models trained only with synthetic degradation data show limited ability to restore complex real-world degradations. In contrast, models trained solely on real-world degradation data tend to overfit, which may harm structural consistency. Our two-stage training strategy effectively balances restoration capability and content consistency.

Figure 15: User study interface used to evaluate the restoration results. Participants are presented with one degraded input image and five restored results generated by different models and are asked to rate them based on restoration quality and consistency.

Explore similar work

GGT-100K: Generative Ground Truth for Generalizable Real-World Image Restoration
May 29, 2026Xiangtao Kong, Jixin Zhao, Lingchen Sun +2Synthetic Data Augmentation Image Restoration
The Hong Kong Polytechnic University · OPPO Research Institute
HarnessIR: Harnessing Multimodal Foundation Models for Universal Real-World Image Restoration
Oct 7, 2026Xiangtao Kong, Shuaizheng Liu, Rongyuan Wu +5Image Restoration All-in-One Image Restoration
The Hong Kong Polytechnic University · OPPO Research Institute
Universal Image Restoration via Internalized Chain-of-Thought Reasoning
Jun 16, 2026Yu Guo, Zhengru Fang, Shengfeng He +4All-in-One Image Restoration Image Restoration
Hong Kong JC Lab of Smart City and the Department of Computer Science, City University of Hong Kong · School of Computing and Information Systems, Singapore Management University · Computer Science and Information Engineering, National Taiwan University

May 29, 2026J/K move · Enter open · S save

GGT-100K: Generative Ground Truth for Generalizable Real-World Image Restoration

Xiangtao Kong, Jixin Zhao, Lingchen Sun, Rongyuan Wu+1

The Hong Kong Polytechnic University · OPPO Research Institute

Synthetic Data Augmentation Image Restoration Multimodal Foundation Models

Open paper AskPDF arXiv

Real-world image restoration (IR) is bottlenecked by the scarcity of high-quality paired training data. Synthetic datasets are abundant but often fail to model real-world degradations, while real-world paired datasets are expensive and difficult to capture. As a result, IR models trained on these datasets show limited generalization in real-world scenarios. In this work, we propose Generative Ground Truth (GGT) by using generative multimodal foundation models (MFMs) to produce high-quality (HQ) targets from real-world low-quality (LQ) images. We first conduct a systematic evaluation of nine state-of-the-art MFMs, including Nano-Banana-2 and GPT-Image-2, on images of various scenes and degradation types. The results demonstrate that Nano-Banana-2 with VLM-based adaptive prompting shows the highest capability to synthesize perceptually realistic and content-faithful HQ targets, which can serve as the GGT for the LQ input. We then employ Nano-Banana-2 to build a GGT synthesis pipeline, which involves multi-stage quality control to ensure data reliability, and construct GGT-100K, an LQ-HQ paired dataset comprising 103,707 training pairs and covering diverse scenes and complex real-world degradations. A test set of 500 image pairs is also established. Extensive experiments show that GGT-100K consistently improves the real-world generalization of a wide range of IR models, with particularly strong benefits for finetuning generative models for IR tasks. Our results suggest that MFMs can serve as practical tools for restoration-oriented data generation, and GGT-100K is a useful resource to expand the generalization boundaries of real-world IR models.

Abstract

Explore similar work

GGT-100K: Generative Ground Truth for Generalizable Real-World Image Restoration

HarnessIR: Harnessing Multimodal Foundation Models for Universal Real-World Image Restoration

Universal Image Restoration via Internalized Chain-of-Thought Reasoning