Restore, Assess, Repeat: A Unified Framework for Iterative Image Restoration

I-Hsiang Chen; Isma Hadji; Enrique Sanchez; Adrian Bulat; Sy-Yen Kuo; Radu Timofte; Georgios Tzimiropoulos; Brais Martinez

cs.CVMar 27, 2026

Restore, Assess, Repeat: A Unified Framework for Iterative Image Restoration

Authors: I-Hsiang Chen, Isma Hadji, Enrique Sanchez, Adrian Bulat, Sy-Yen Kuo, Radu Timofte, Georgios Tzimiropoulos, Brais Martinez

Organizations: Samsung AI Center Cambridge · National Taiwan University · Technical University of Iasi · Chang Gung University · University of Wurzburg · Queen Mary University of London

Abstract

Image restoration aims to recover high quality images from inputs degraded by various factors, such as adverse weather, blur, or low light. While recent studies have shown remarkable progress across individual or unified restoration tasks, they still suffer from limited generalization and inefficiency when handling unknown or composite degradations. To address these limitations, we propose RAR, a Restore, Assess and Repeat process, that integrates Image Quality Assessment (IQA) and Image Restoration (IR) into a unified framework to iteratively and efficiently achieve high quality image restoration. Specifically, we introduce a restoration process that operates entirely in the latent domain to jointly perform degradation identification, image restoration, and quality verification. The resulting model is fully trainable end to end and allows for an all-in-one assess and restore approach that dynamically adapts the restoration process. Also, the tight integration of IQA and IR into a unified model minimizes the latency and information loss that typically arises from keeping the two modules disjoint, (e.g. during image and/or text decoding). Extensive experiments show that our approach consistent improvements under single, unknown and composite degradations, thereby establishing a new state-of-the-art.

Figures & tables

Figure 1: Overview of RAR . Given an input image with unknown composite degradations, RAR operates entirely in latent by iteratively assessing current image quality, restoring the image based on the assessment, and verifying the quality improvement.

Figure 2: To enable our RAR process, we define an LQA that fully integrates the input and out of the IQA into the latent space of the restoration module, using adapters $\mathcal{A}_{I}$ and $\mathcal{A}_{Q}$ , respectively. Then, we instore a feedback loop between the two modules to iteratively recover best image quality. The dashed-line is only used for the very first iteration to feed $\mathbf{z}^{0}_{deg}$ to the restoration model.

Figure 3: We illustrate our quality verification step. This is used at inference time to define a stopping criterion of the RAR process.

Panel of Figure 4 — Figure 4: Qualitative comparison on Composite Degradation.

Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	MANIQA $\uparrow$	CLIP-IQA $\uparrow$	MUSIQ $\uparrow$
Group A
AirNet	19.13	0.6019	0.4283	0.2581	0.3930	42.46
PromptIR	20.06	0.6088	0.4127	0.2633	0.4013	42.62
MiOIR	20.84	0.6558	0.3715	0.2451	0.3933	47.82
DA-CLIP	19.58	0.6032	0.4266	0.2418	0.4139	42.51
InstructIR	18.03	0.5751	0.4429	0.2660	0.3528	45.77

Table 1: Quantitative results for Composite Degradations . Results taken from AgenticIR [ 53 ]

Panel of Figure 5 — Figure 5: Qualitative analysis: Comparison with other multi-round methods (AutoDIR [ 16 ] , AgenticIR [ 53 ] ). Intermediate steps for our method are shown for illustrative purposes. First row: Composite degradations (haze + noise). RAR identifies and corrects them, also enhances contrast, making the result visually clearer. Second row: Single degradation (low light). AgenticIR fails due to inadequate tools for very dark scenes. Third row: Single degradation (haze). RAR successfully dehazes, but also improves low light, lowering the fidelity to ground truth. Fourth row: Unknown degradations. RAR demonstrates stronger generalization (also improving over ground truth).

Method	MUSIQ $\uparrow$	CLIP-IQA $\uparrow$	NIQE $\downarrow$	NIMA $\uparrow$
AirNet	41.20	0.265	7.681	4.048
PromptIR	45.40	0.303	8.294	4.344
AdaIR	21.14	0.225	7.688	4.023
DARCOT	20.38	0.211	8.224	4.039
LD	43.41	0.288	7.983	4.371
AutoDIR	49.27	0.311	6.392	4.427

Table 2: Results for Unknown Degradations on the UDC dataset.

Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	CLIP-IQA $\uparrow$	MUSIQ $\uparrow$	MANIQA $\uparrow$
AirNet	25.65	0.8230	0.1820	0.4516	48.69	0.3163
PromptIR	27.04	0.8445	0.1515	0.4196	48.77	0.3169
AdaIR	26.48	0.7825	0.2604	0.4368	53.51	0.3456
DARCOT	26.08	0.7772	0.2579	0.4409	54.34	0.3461
LD	22.17	0.6922	0.1935	0.4253	54.94	0.3438
AutoDIR	27.81	0.8703	0.1283	0.4119	54.37	0.3053

Table 3: Quantitative comparison on Single Degradation tasks.

Panel of Figure 6 — Figure 6: Qualitative comparison on single degradation: Row1: Kodak24 (noise), Row2: GoPro (blur), Row3: Rain100L (rain), Row4: SOTS (haze), In the single-degradation setting, our method produces clearer images that better align with human visual perception. This results lower fidelity scores but higher perceptual scores.

Backbone	Configuration				Unknown Degradations (UDC)			Composite Degradations (Group-C)
Backbone	IQA	Text Space	Image Space	noise-free cond.	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
SD1.5	CLIP	Text	Pixel	✗	20.67	0.6217	0.2211	16.70	0.4848	0.2997
SD1.5	DepictQA	Text	Pixel	✗	22.34	0.6596	0.2016	17.67	0.5260	0.2267
SD3.5	DepictQA	Text	Pixel	✗	24.41	0.7583	0.1652	17.88	0.5696	0.2306
SD3.5	DepictQA	Text	Pixel	✓	24.70	0.7766	0.1589	17.89	0.5776	0.2289
SD3.5	DepictQA	Text	Latent	✓	24.90	0.7952	0.1131	18.72	0.6278	0.1576

Table 4: Ablation of the integration process between the IQA and the UIR modules.

Panel of Figure 7 — Figure 7: Noise vs. noise-free flow matching formulation. Notice how noise affects the eye of the parrot (zoom required).

Panel of Figure 8 — Figure 8: Text vs. embeddings as conditioning inputs. Embeddings provide more efficient and informative conditioning.

Backbone	Training approach	Unknown (UDC)			Composite (Group-C)
Backbone	Iterative	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
SD1.5	-	25.57	0.747	0.132	18.37	0.555	0.175
SD1.5	✓	23.05	0.697	0.154	18.17	0.553	0.177
SD3.5	-	28.49	0.849	0.101	18.76	0.625	0.173
SD3.5	✓	28.60	0.856	0.083	19.16	0.649	0.149

Table 5: Role of the RAR iterative process. Intermediate assessments are not accurate for SD1.5 due to the noised latents. Instead, our noise-free FM variant improves across all metrics and datasets.

# RAR rounds	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	CLIPIQA $\uparrow$	MUSIQ $\uparrow$	MINIQA $\uparrow$
1	30.69	0.901	0.0583	0.591	54.87	0.417
2	29.55	0.873	0.073	0.601	56.29	0.428
3	29.12	0.866	0.077	0.608	56.80	0.438
4	28.60	0.856	0.083	0.614	57.19	0.446
2.4	29.69	0.874	0.071	0.603	55.97	0.423

Table 6: Stopping Criterion Analysis. Results for unknown composite degradations on the UDC dataset.

Degradations	Method	Wall clock time	# of Round	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Group A	AutoDIR	14.30	2.92	19.64	0.6286	0.3967
Group A	AgenticIR	48.00	3.37	21.04	0.6818	0.3148
	RAR	6.29	2.82	20.46	0.7144	0.1299
Group B	AutoDIR	10.59	1.98	19.90	0.6643	0.3542
Group B	AgenticIR	54.00	3.63	20.55	0.7009	0.3072
	RAR	6.71	2.84	21.04	0.7326	0.1269

Table 7: Analysis of Computational Efficiency. Comparison with iterative approaches in terms of latency and performance.

Appendix figures & tables16 assets

Supplementary material from the paper’s appendix.

Appendix

Figure S.1 : Illustrations of the three prompt-driven IQA tasks used in LQA training : distortion identification, assessment reasoning, and quality comparison. These tasks collectively preserve rich degradation-awareness within the unified VLM backbone.

Figure S.2 : Overview of the latent space alignment module , which allows us to use a single encoder for both IQA and restoration modules.

Figure S.3 : Overview of the conditioning alignment module , which maps LQA outputs into the UIR conditioning space.

Method	Precision $\uparrow$	Recall $\uparrow$	F1-score $\uparrow$
DepictQA (Pixel-based IQA)	0.7713	0.7564	0.7612
LQA	0.8750	0.8748	0.8749

Appendix

Table S.1 : Distortion identification on composite degradations. Evaluated on KADIS700K [ 24 ] dataset.

Method	Haze	Blur	Rain	LOL	Raindrop	Noise	Low-Res	Average
CLIP	0.674	0.721	0.889	0.882	0.757	0.805	0.723	0.778
SA-CLIP	1.000	0.944	0.995	1.000	1.000	1.000	0.998	0.991
LQA	1.000	1.000	1.000	1.000	1.000	1.000	0.975	0.996

Appendix

Table S.2 : Distortion identification on single degradations. F1-scores on each restoration dataset.

Degradation	Prompts	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	CLIPIQA $\uparrow$	MUSIQ $\uparrow$	MINIQA $\uparrow$
Group C	Fixed	19.33	0.6579	0.1489	0.6554	56.56	0.4653
Group C	Random	19.31	0.6572	0.1493	0.6566	56.72	0.4675

Appendix

Table S.3 : Sensitivity Analysis of User Prompts.

Figure S.4 : Training curves comparing noise-free flow matching with the standard noise-conditioned variant, showing significantly faster and stable convergence.

Backbone	Configuration				Unknwon (UDC)			Composite (Group-C)
Backbone	IQA	Text Space	Image Space	Iterative Training	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
AutoDIR	CLIP	Text	Pixel	✗	22.47	0.7267	0.2199	18.61	0.5443	0.5019
SD1.5	CLIP	Text	Pixel	✗	20.67	0.6217	0.2211	16.70	0.4848	0.2997
SD1.5	DepictQA	Text	Pixel	✗	22.34	0.6596	0.2016	17.67	0.5260	0.2267
SD1.5	DepictQA	Text	Latent	✗	23.66	0.7055	0.1498	18.06	0.5475	0.1792
SD1.5	DepictQA	Embeddings	Latent	✗	25.57	0.7473	0.1319	18.37	0.5546	0.1751

Appendix

Table S.4 : Ablation of the integration process between the IQA and the UIR modules.

Panel of Figure 24 — Figure S.5 : Comparison of RAR when iterative operating in pixel space versus latent space. Please zoom in on the eye region for detailed comparison.

Method	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	CLIP-IQA $\uparrow$	MUSIQ $\uparrow$	MANIQA $\uparrow$
Denoise (2 sets)
LD	22.58	0.6250	0.1860	0.4586	44.66	0.3481
AirNet	29.10	0.8030	0.2130	0.4246	47.29	0.2997
PromptIR	29.89	0.8240	0.1810	0.4124	46.22	0.2905
AutoDIR	29.68	0.8320	0.1670	0.4654	43.46	0.2996
AgenticIR	27.89	0.8020	0.2680	0.4653	59.30	0.3917

Appendix

Table S.5 : Quantitative comparison on Single Degradation tasks.

# rounds	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	CLIPIQA $\uparrow$	MUSIQ $\uparrow$	MINIQA $\uparrow$
Single Degradation
1	26.09	0.8485	0.0672	0.5463	55.78	0.4036
2	24.94	0.8334	0.0760	0.5836	58.15	0.4413
3	24.32	0.8212	0.0836	0.5970	59.31	0.4638
4	23.87	0.8127	0.0896	0.6033	59.91	0.4774
1.3	25.88	0.8378	0.0699	0.5566	56.00	0.4125

Appendix

Table S.6 : Stopping Criterion Analysis . Results for single and composite degradations.

Stopping	Nmax	# rounds	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	CLIPIQA $\uparrow$	MUSIQ $\uparrow$	MINIQA $\uparrow$
	4	4	19.20	0.6495	0.1508	0.6683	58.33	0.4982
✓	4	2.87	19.32	0.6578	0.1499	0.6567	56.88	0.4694
	16	16	17.86	0.5851	0.2243	0.6326	58.33	0.5115
✓	16	3.02	19.29	0.6562	0.1519	0.6560	56.97	0.4719

Appendix

Table S.7 : Sensitivity Analysis of Iteration Counts. Results for composite degradations on Group C.

Panel of Figure 28 — Figure S.6 : Qualitative analysis of iterative restoration methods on composite degradations , including haze, blur, noise, low-resolution conditions.

Panel of Figure 29 — Figure S.7 : Qualitative analysis of iterative restoration methods on composite degradations, including rain, noise, low-resolution, low-light.

Panel of Figure 30 — Figure S.8 : Qualitative analysis of iterative restoration methods on EUVP [ 15 ] dataset.

Panel of Figure 31 — Figure S.9 : Qualitative analysis of iterative restoration methods on UDC [ 52 ] dataset.

Explore similar work

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
TaskIR: Task-Driven Image Restoration via Degradation Adaptation and Task Feedback
Sep 25, 2026Yanjie Tu, Qingsen Yan, Axi Niu +4All-in-One Image Restoration Image Restoration
Northwestern Polytechnical University · Shenzhen Research Institute of Northwestern Polytechnical University · Xi’an University of Architecture and Technology
Degradation-Aware Adaptive Context Gating for Unified Image Restoration
May 2, 2026Lei He, Jielei Chu, Fengmao Lv +4All-in-One Image Restoration Image Restoration
the School of Computing and Artificial Intelligence, Southwest Jiaotong University, Chengdu 611756, China · the Engineering Research Center of Sustainable Urban Intelligent Transportation, Ministry of Education, Chengdu, Sichuan 611756, China · School of Computing and Artificial Intelligence, Jiangxi University of Finance and Economics, Nanchang, 330013, China +3

Jun 16, 2026J/K move · Enter open · S save

Universal Image Restoration via Internalized Chain-of-Thought Reasoning

Yu Guo, Zhengru Fang, Shengfeng He, Senkang Hu+3

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

All-in-One Image Restoration Image Restoration

Open paper AskPDF arXiv

Image restoration seeks to recover high-quality images from degraded inputs but becomes highly ill-posed under complex, mixed degradations. While unified all-in-one models are common, their performance declines as degradation complexity increases. Recent works adopt Chain-of-Thought (CoT) reasoning for multi-round restoration using specialized modules. However, this approach faces two key limitations: (i) increased computational cost due to multi-step processing, and (ii) weak modeling of interactions between degradations during stepwise inference. We introduce CoTIR, a universal image restoration framework that internalizes CoT reasoning within a single model. Concretely, we view image restoration as a specialized subtask of image editing, which implies that a large-scale pre-trained editing model provides a more favorable optimization starting point. Building on this, we fine-tune the model for restoration and further encode structured CoT-style reasoning into the learning objective via a differentiable formulation inspired by Lagrangian optimization, enabling holistic restoration without chaining specialized restorers. To facilitate training and evaluation, we further present CoTIR-Bench, a large-scale benchmark comprising 5.2 million samples with CoT-style reasoning traces. Extensive experiments on CoTIR-Bench and broad real composite degradation scenes show that CoTIR achieves stronger perceptual quality and more competitive fidelity than both all-in-one models and multi-round restoration methods. The source code is available at https://github.com/gy65896/CoTIR.

Abstract

Explore similar work

Universal Image Restoration via Internalized Chain-of-Thought Reasoning

TaskIR: Task-Driven Image Restoration via Degradation Adaptation and Task Feedback

Degradation-Aware Adaptive Context Gating for Unified Image Restoration