Universal image restoration aims to recover clean images from arbitrary real-world degradations using a single inference model. Despite significant progress, existing all-in-one restoration networks do not scale to multiple degradations. As the number of degradations increases, training becomes unstable, models grow excessively large, and performance drops across both seen and unseen domains. In this work, we show that scaling universal restoration is fundamentally limited by interference across degradations during joint learning, leading to catastrophic task forgetting. To address this challenge, we introduce a unified inference pipeline with a multi-branch mixture-of-experts architecture that decomposes restoration knowledge across specialized task-adaptable experts. Our approach enables scalable learning (over sixteen degradations), adapts and generalizes robustly to unseen domains, and supports user-controllable restoration across degradations. Beyond achieving superior performance across benchmarks, this work establishes a new design paradigm for scalable and controllable universal image restoration.
Figures & tables
Figure 1: We introduce UnSCAR , a universal image restoration model that can handle 2× more degradations than previously explored, achieving state-of-the-art performance across multiple benchmarks. Our framework also enables user-controllable restoration via degradation-control sliders and supports few-shot adaptation (down to one-shot) with strong out-of-distribution generalization on challenging medical imaging datasets.
Method
Venue
Unified Model
Mixed Restoration
Automatic
Scalability
Adaptability
Control Modality
#Primary Tasks
Restormer [ 80 ]
CVPR‘22
✗
✗
✗
✗
(✓)
—
1
AutoDIR [ 31 ]
ECCV‘24
✓
(✓)
(✓)
✗
✗
—
4
DA-CLIP [ 50 ]
ICLR‘24
✗
✗
✓
✗
✗
—
10
OneRestore [ 26 ]
ECCV‘24
✓
✓
(✓)
(✓)
✗
Limited
3
MPerceiver [ 5 ]
CVPR‘24
✓
✓
✓
✗
(✓)
—
8
CAPTNet [ 24 ]
TCVST‘24
✓
✗
✗
✗
✗
—
4
Table 1: Comparison with prior image restoration works. Factors are denoted as fully met ✓ , partially met (✓) , or unmet ✗ .
Figure 2: Low-level cues. (a) Low-level cues provide discriminative signals across degradation types, enabling automatic restoration. (b) Our framework leverages diverse cues such as edges, color statistics, noise, and transmission, to characterize the input scene.
Figure 3: UnSCAR architecture. We leverage low-level visual guidance within a bidirectional feedback, controllable generative network with mixture-of-experts. This enables unified image restoration across multiple simultaneous degradations.
Figure 4: Degradation-aware embeddings. Our degradation-aware training framework (left) learns an auxilliary degradation encoder cloned from the CLIP image encoder trained with SigLIP loss [ 82 ] . The resulting embeddings provide clear separation between single- and mixed-degradation tasks while maintaining proximity across related degradations.
Figure 6: Generalization to unseen degradations. Zero-shot evaluation of baseline methods and our approach on unseen mixed-degradation datasets.
Dehazing: RESIDE [ 38 ]
Motion Deblur : RHM [ 84 ]
Low-Light: LOL [ 74 ]
Denoising : CBSD68 [ 52 ]
Method
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
PSNR ↑
SSIM ↑
LPIPS ↓
{\color[rgb]{0,0,1}\blacksquare}
Restormer [ 80 ]
26.57
0.95
0.05
26.41
0.81
0.35
20.43
0.83
0.30
28.60
0.79
0.23
NAFNet [ 12 ]
27.62
0.96
0.04
27.87
0.84
0.24
22.19
0.88
0.16
32.71
0.89
0.08
{\color[rgb]{0.5,0.5,0}\blacksquare}
PromptIR [ 59 ]
26.27
0.95
0.03
28.69
0.86
0.18
18.58
0.83
0.20
30.52
0.84
0.16
OneRestore [ 26 ]
30.68
0.97
0.02
27.95
0.82
0.27
22.21
0.88
0.19
32.27
0.88
0.11
InstructIR [ 14 ]
29.78
0.97
0.02
27.14
0.83
0.24
22.33
0.89
0.14
32.10
0.88
0.10
Table 2: Single-degradation restoration performance. All models are trained jointly on all tasks to evaluate scalability. Bold denotes best performance; underline indicates second best. {{\color[rgb]{0,0,1}\blacksquare}} and {{\color[rgb]{0.5,0.5,0}\blacksquare}} denotes prior single-task and all-in-one restoration models.
Figure 7: Mixed restoration. Zero-shot evaluation of baseline methods vs ours for mixed restoration dataset with individual tasks seen during training.
Figure 8: In-the-wild degradation removal. Comparison with the diffusion based DA-CLIP method on heavy indoor smoke removal [ 88 ] .
Method
TOLED [ 95 ]
POLED [ 95 ]
EUVP [ 30 ]
NIQE ↓
BRISQUE ↓
NIQE ↓
BRISQUE ↓
NIQE ↓
BRISQUE ↓
Restormer [ 80 ]
6.95
46.69
10.38
27.58
7.30
57.93
NAFNet [ 12 ]
6.90
49.58
6.65
34.42
6.76
50.37
OneRestore [ 26 ]
6.93
52.97
6.49
16.57
7.43
56.45
InstructIR [ 14 ]
7.16
49.05
7.06
35.08
6.83
50.89
AdaIR [ 16 ]
7.07
44.99
6.96
25.06
7.40
54.60
Table 3: Zero-shot performance on unseen-mixed (left) and mixed datasets (right). The best performing models across all tasks from Table 1 were used for evaluations.
Figure 9: Fewshot generalization. One (K=1) to Few (K<=50) shot generalization performance over demoiring and demosaicing tasks.
Table 13Table 14
Figure 10: Medical. Our model adapts to out-of-distribution data such as medical images.
Figure 11: Controllability. Our model supports slider-based control of the level of restoration per degradation. α denotes the aberration strength (haze vs. rain, noise vs. gray).
Figure 12: Ablation studies. We analyze model performance under reduced sampling steps (left) and fewer low-level cues under automatic versus user controlled restoration (middle), and the impact of decoder refinement on structural preservation (right).
Restoring images affected by various types of degradation, such as noise, blur, or improper exposure, remains a significant challenge in computer vision. While recent trends favor complex monolithic all-in-one architectures, these models often suffer from negative task interference and require extensive joint training cycles on high-end computing clusters. In this paper, we propose a modular, task-decoupled image restoration framework based on an explicit diagnostic routing mechanism. The architecture consists of a lightweight Convolutional Neural Network (CNN) classifier that evaluates the input image and dynamically directs it to a specialized restoration node. A key advantage of this framework is its model-agnostic extensibility: while we demonstrate it using three independent U-Net experts, the system allows for the integration of any restoration method tailored to specific tasks. By isolating reconstruction paths, the framework prevents feature conflicts and significantly reduces training overhead. Unlike monolithic models, adding new degradation types in our framework only requires training a single expert and updating the router, rather than a full system retraining. Experimental results demonstrate that this computationally accessible approach offers a scalable and efficient solution for multi-degradation restoration on standard local hardware. The code will be published upon paper acceptance.
Joanna Wiekiera, Martyna Zur
Faculty of Automatic Control, Electronics and Computer Science, Silesian University of Technology, Gliwice, Poland
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.
Yu Guo, Zhengru Fang, Shengfeng He +4
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
Recent progress in image restoration has converged on all-in-one architectures that jointly handle multiple degradations within a single network. These methods are effective on static benchmarks but target a closed-world setting that assumes simultaneous access to every target degradation at training time. In practice, degradations are encountered sequentially as field-deployed systems progressively face new environmental conditions, and historical training data is often unavailable due to privacy or storage constraints. Accommodating a new degradation then requires either retraining on the union of all prior data, which is often costly or infeasible, or fine-tuning, which causes catastrophic forgetting. We formulate multi-degradation image restoration as a continual domain-incremental learning problem, in which degradations arrive incrementally and prior data is unavailable. Our proposed Restoring without Forgetting (RwF) framework learns a lightweight adapter for each new degradation, eliminating forgetting by construction at a fraction of the cost of dedicated per-domain networks. To isolate degradation learning from dataset variation, we construct a benchmark spanning five degradation domains under shared image content. At test time, an unsupervised routing mechanism identifies the appropriate restoration path for unknown inputs without requiring domain labels. Across the five-domain sequence, RwF improves final average PSNR over naive sequential fine-tuning by 15.25 dB and 11.83 dB on the Restormer and NAFNet backbones, respectively. The framework transfers to eleven canonical real-degradation benchmarks (3,465 images) at 89.5% routing accuracy with only a +0.94 dB oracle PSNR gap, establishing, to our knowledge, the first systematic baseline for continual multi-degradation image restoration.