UniRestorer++: Scalable Universal Image Restoration with Multi-Granularity Degradation Modeling

Jingbo Lin; Zhilu Zhang; Fan Li; Renjing Pei; Wenbo Li; Hang Xu; Hongzhi Zhang; Wangmeng Zuo

cs.CVNov 4, 2026

UniRestorer++: Scalable Universal Image Restoration with Multi-Granularity Degradation Modeling

Authors: Jingbo Lin, Zhilu Zhang, Fan Li, Renjing Pei, Wenbo Li, Hang Xu, Hongzhi Zhang, Wangmeng Zuo

Abstract

Universal image restoration aims to handle diverse degradations within a single model, but a reliable scaling strategy remains underexplored. Increasing backbone depth and width can lead to performance saturation, while degradation-aware specialization depends on accurate degradation estimation and is vulnerable to errors on out-of-distribution and real-world inputs. We propose UniRestorer++, a scalable framework based on multi-granularity degradation modeling. It progressively refines degradation groups and introduces group-specific experts, coupling capacity growth with specialization. Degradation-estimation-based routing selects a root-to-leaf path, while granularity-estimation-based routing chooses a level along that path to balance specialization and robustness. Spatially dense degradation representation maps further enable region-aware routing for spatially non-uniform corruptions. We develop CDD-50, a benchmark covering 50 degradation categories from single to penta-degradation. Experiments show approximately log-linear PSNR trends over the evaluated model-size range and stronger scaling gains on out-of-distribution and real-world data than vanilla MoE scaling. UniRestorer++ achieves state-of-the-art performance across benchmarks and demonstrates zero-shot generalization and few-shot transfer on real-world images. Code: https://github.com/mrluin/UniRestorer_plusplus

Figures & tables

Fig. 1: Representative universal image restoration frameworks. (a) Degradation-agnostic methods use a shared backbone without explicit degradation conditioning. Degradation-aware methods guide restoration through conditioning (b) or expert-based processing (c) that routes features toward degradation-specific experts. These representative methods typically use a fixed degradation granularity, which can make them sensitive to degradation-estimation errors. (d) UniRestorer $++$ proposes multi-granularity degradation modeling, expanding model capacity by introducing finer degradation-specific restoration experts. Degradation-estimation-based routing determines a hierarchical path, while granularity-estimation-based routing selects an expert level along that path, balancing degradation-specific specialization with robustness to degradation-estimation errors.

Fig. 4: Overview of UniRestorer $++$ . Globally pooled degradation representations (DRs) are hierarchically clustered to construct a multi-granularity degradation set, while dense DR maps are retained for region-aware expert routing. At each restoration block, the degradation-estimation-based and granularity-estimation-based routers are conditioned on both dense DR priors and the current restoration feature, enabling expert routing to adapt throughout the network. Their joint decision produces an expert assignment at each spatial location.

Fig. 5: Illustration of the degradation pipeline. Weather artifacts are first introduced, after which the remaining degradation types are randomly selected, shuffled, and sequentially applied.

1st-Order	2nd-Order	3rd-Order	4th-Order	5th-Order
Noise (N)	I-RG	B-I-J	I-RL-J-RS	RS-RL-B-N-RG
Blur (B)	J-B	RS-RG-J	RS-B-J-RG	B-B-RL-N-J
JPEG (J)	B-N	B-N-RS	RL-RL-RG-RG	SH-I-RL-RL-RL
Rain (R)	RS-B	RL-RL-J	RS-RG-RS-J	H-RL-B-RS-J
Snow (S)	SH-RL	B-RS-RL	RH-RS-RG-J	RH-RS-J-RL-RG
Haze (H)	H-RG	RS-RS-RL	N-B-J-RG	J-J-I-RL-RG

Table I: Degradation combinations in the five degradation-order groups of CDD-50. ‘RH’ and ‘SH’ denote mixed rain-haze and snow-haze degradations, respectively.

Type

Method

Derain

Dehaze

Denoise

Deblur

Low-light

Desnow

CAR

Average

Rain100L

SOTS

BSD68

GoPro

LOLv1

Snow100K

LIVE1

AirNet [ 17 ]

34.90

.967

27.94

.962

31.06

.872

–

29.29

–

PromptIR [ 13 ]

36.37

.972

30.58

.974

31.12

.873

–

31.50

–

InstructIR [ 87 ]

37.98

.978

30.22

.959

31.32

.875

–

31.49

–

DaAIR [ 44 ]

37.10

.978

32.30

.981

31.06

.868

–

32.51

–

MoCE-IR [ 46 ]

38.57

.984

31.34

.979

31.24

.873

–

32.73

–

Table II: Universal image restoration results in the single-degradation setting. CAR denotes compression artifact removal. PSNR (dB,

\uparrow

) and SSIM (

\uparrow

) are reported in RGB space. ‘

n

T’ denotes a model jointly trained on

n

tasks (3T: Derain, Dehaze, and Denoise; 5T: Derain, Dehaze, Denoise, Deblur, and Lowlight; 7T: all seven tasks). The weighted averages of benchmarks are reported. The best performance is highlighted.

Method	1st-Order								2nd-Order										3rd-Order				Average
Method	L		H		R		S		L-H		L-R		L-S		H-R		H-S		L-H-R		L-H-S		Average
AirNet [ 17 ]	24.83	.778	24.21	.951	26.55	.891	26.79	.919	23.23	.779	22.82	.710	23.29	.723	22.21	.868	23.29	.901	21.80	.708	22.24	.725	23.75
PromptIR [ 13 ]	26.32	.805	26.10	.969	31.56	.946	31.53	.960	24.49	.789	25.05	.771	24.51	.761	24.54	.924	23.70	.925	23.74	.752	23.33	.747	25.90
OneRestore [ 15 ]	26.48	.826	32.52	.990	33.40	.964	34.31	.973	25.79	.822	25.58	.799	25.19	.789	29.99	.957	30.21	.964	24.78	.788	24.90	.791	28.47
MoCE-IR [ 46 ]	27.26	.824	32.66	.990	34.31	.970	35.91	.980	26.24	.817	26.25	.800	26.04	.793	29.93	.964	30.19	.970	25.41	.789	25.39	.790	29.05
UniRestorer [ 31 ]	27.71	.900	34.53	.991	36.07	.972	36.84	.983	26.88	.893	26.93	.894	26.73	.892	32.02	.936	30.82	.976	25.78	.884	25.87	.885	30.02

Table III: Universal image restoration results on CDD-11. L, H, R, and S denote low-light, haze, rain, and snow, respectively. PSNR (dB,

\uparrow

) and SSIM (

\uparrow

) are reported in RGB space. The best performance is highlighted.

Method	In-dist. Degradation												Out-of-dist. Degradation
Method	1st-Order		2nd-Order		3rd-Order		4th-Order		5th-Order		Average		1st-Order		2nd-Order		3rd-Order		4th-Order		5th-Order		Average
AirNet [ 17 ]	28.25	.912	23.10	.808	20.50	.758	20.20	.760	19.05	.714	22.22	.790	25.60	.868	20.35	.755	18.80	.710	18.45	.690	16.70	.655	19.98	.736
PromptIR [ 13 ]	30.78	.943	25.24	.848	22.27	.801	22.11	.806	20.92	.761	24.26	.832	27.92	.902	22.45	.795	20.66	.751	19.92	.724	18.01	.689	21.79	.772
OneRestore [ 15 ]	28.28	.926	24.02	.846	22.54	.793	22.15	.792	20.72	.735	23.54	.818	26.08	.897	21.42	.788	20.19	.731	20.08	.727	17.85	.642	21.12	.757
MoCE-IR [ 46 ]	31.15	.948	26.36	.882	23.52	.823	23.32	.821	22.07	.775	25.28	.850	28.56	.923	23.66	.830	21.24	.776	20.57	.758	18.60	.721	22.53	.802
UniRestorer [ 31 ]	32.58	.959	26.98	.901	24.29	.842	23.81	.837	22.53	.792	26.04	.866	29.72	.935	23.91	.845	21.57	.781	21.13	.769	19.12	.713	23.09	.809

Table IV: Universal image restoration results on CDD-50 under in-dist. and out-of-dist. degradation parameters. PSNR (dB,

\uparrow

) and SSIM (

\uparrow

) are reported in RGB space. The best performance is highlighted.

Method	SIDD	RealBlur-J	DPDD	RealRain	RealSnow	WeatherStream	FoundIR
Method	[ 82 ]	[ 83 ]	[ 84 ]	[ 85 ]	[ 23 ]	[ 86 ]	[ 30 ]
AirNet [ 17 ]	23.65	24.92	20.17	22.56	23.63	21.68	20.56
PromptIR [ 13 ]	24.32	25.47	21.76	24.12	24.94	22.09	21.54
OneRestore [ 15 ]	11.46	18.79	20.96	22.73	20.72	19.49	22.94
MoCE-IR [ 46 ]	15.96	14.29	15.61	28.10	24.63	20.54	22.63
UniRestorer [ 31 ]	24.70	25.55	23.60	29.85	27.73	22.25	23.80

Table V: Zero-shot generalization comparison on real-world datasets. PSNR (dB,

\uparrow

) is reported in RGB space. All methods are trained in the single-degradation setting and evaluated without additional training. The best performance is highlighted.

Method	#Params. (M)	#FLOPs (G)	Latency (s)
AirNet [ 17 ]	9	2414	0.548
PromptIR [ 13 ]	36	1266	0.355
InstructIR [ 87 ]	16	130	0.055
PromptIR-DCPT [ 88 ]	36	1266	0.355
MoCE-IR [ 46 ]	25	716	0.347
UniRestorer [ 31 ]	171	1156	0.484

Table VI: Computational cost comparison with state-of-the-art universal methods in the single-degradation setting.

Fig. 6: Visual comparison across the single-degradation (GoPro deblurring), triple-degradation (CDD-11), penta-degradation (CDD-50), and real-world settings. UniRestorer $++$ removes degradations and recovers textures and colors, as highlighted by the red boxes and enlarged regions.

Evaluation

Degradation Order / Test Set

Baseline

UniRestorer

++

UniRestorer

++

UniRestorer

++

UniRestorer

++

UniRestorer

++

UniRestorer

++

Setting

238M

154M

244M

423M

781M

1.5B

2.9B

In-dist. Synthetic Degradation in CDD-50

1st-Order

33.39

.965

+0.15

+.002

+0.36

+.003

+0.56

+.003

+0.81

+.004

+0.90

+.005

+0.94

+.006

2nd-Order

27.35

.912

+0.05

+.001

+0.15

+.002

+0.31

+.003

+0.46

+.005

+0.55

+.005

+0.58

+.006

3rd-Order

24.76

.854

+0.04

+.001

+0.10

+.002

+0.21

+.003

+0.31

+.005

+0.40

+.007

+0.44

+.008

4th-Order

24.01

.845

+0.13

+.003

+0.19

+.003

+0.30

+.005

+0.44

+.017

+0.51

+.018

+0.55

+.019

5th-Order

22.78

.803

+0.06

+.001

+0.14

+.003

+0.23

+.004

+0.36

+.006

+0.43

+.008

+0.46

+.009

Table VII: Performance of hierarchically scaled UniRestorer

++

variants. PSNR (dB,

\uparrow

) and SSIM (

\uparrow

) are reported in RGB space. All UniRestorer

++

variants are trained on the complete CDD-50 degradation space. Each degradation-order row uses its own

238

M baseline trained only on the corresponding order, and all improvements are computed relative to the baseline in the same row. In the real-world setting, the

237.8

M baseline is trained on the complete CDD-50 degradation space containing all five orders.

Test Sets	Task	Specialist Methods	Performance		UniRestorer $++$		UniRestorer $++$		UniRestorer $++$
Test Sets	Task	Specialist Methods	of Specialists		(2.9B, zero-shot)		(2.9B, 1-shot)		(2.9B, 5-shot)
SIDD [ 82 ]	Denoising	NAFNet [ 11 ]	40.30	.962	38.72	.942	38.83 $\pm$ .21	.945 $\pm$ .009	39.26 $\pm$ .13	.957 $\pm$ .004
RealBlur-J [ 83 ]	Motion Deblurring	X-Restormer [ 12 ]	36.27	.958	30.81	.905	36.20 $\pm$ .11	.952 $\pm$ .005	36.28 $\pm$ .05	.958 $\pm$ .002
DPDD [ 84 ]	Defocus Deblurring	Restormer [ 10 ]	25.98	.811	24.91	.782	25.81 $\pm$ .28	.804 $\pm$ .015	26.28 $\pm$ .14	.815 $\pm$ .006
RealRain [ 85 ]	Deraining	Restormer [ 10 ]	43.93	.993	31.37	.930	41.95 $\pm$ .37	.990 $\pm$ .003	44.06 $\pm$ .16	.991 $\pm$ .002
RealSnow [ 23 ]	Snow Removal	Restormer [ 10 ]	30.94	.913	28.88	.895	31.58 $\pm$ .42	.946 $\pm$ .011	33.28 $\pm$ .19	.969 $\pm$ .004

Table VIII: Zero-, one-, and five-shot comparison with state-of-the-art specialist models on real-world benchmarks. Specialist results are taken from the corresponding papers or trained from scratch. PSNR (dB,

\uparrow

) and SSIM (

\uparrow

) are reported in RGB space. We repeat each few-shot experiment five times and report

\operatorname{mean}\pm\operatorname{std}

. The best performance is highlighted.

Method	Blur	Noise	JPEG	Haze	Rain	Raindrop	B+N	B+J	N+J	R+H	B+N+J
FoundIR [ 30 ]	25.93	38.98	34.03	22.19	34.29	29.74	23.87	29.58	40.43	21.80	27.98
FoundIR [ 30 ]	.791	.964	.942	.849	.943	.916	.773	.854	.979	.758	.745
UniRestorer $++$	25.80 $\pm$ .22	40.10 $\pm$ .11	34.80 $\pm$ .08	21.50 $\pm$ .37	34.80 $\pm$ .26	32.80 $\pm$ .31	24.90 $\pm$ .28	26.80 $\pm$ .34	40.20 $\pm$ .14	23.00 $\pm$ .36	27.10 $\pm$ .32
(2.9B, 1-shot)	.775 $\pm$ .006	.970 $\pm$ .003	.975 $\pm$ .002	.852 $\pm$ .008	.949 $\pm$ .005	.945 $\pm$ .006	.832 $\pm$ .007	.818 $\pm$ .007	.973 $\pm$ .002	.812 $\pm$ .010	.725 $\pm$ .009
UniRestorer $++$	27.62 $\pm$ .14	41.89 $\pm$ .04	35.23 $\pm$ .02	23.34 $\pm$ .23	37.10 $\pm$ .15	37.22 $\pm$ .18	27.02 $\pm$ .21	25.94 $\pm$ .26	41.01 $\pm$ .05	26.81 $\pm$ .29	27.30 $\pm$ .12
(2.9B, 5-shot)	.881 $\pm$ .004	.985 $\pm$ .002	.976 $\pm$ .002	.894 $\pm$ .006	.969 $\pm$ .003	.979 $\pm$ .003	.864 $\pm$ .005	.803 $\pm$ .006	.985 $\pm$ .002	.891 $\pm$ .006	.742 $\pm$ .007

Table IX: One- and five-shot performance on FoundIR [ 30 ] . PSNR (dB,

\uparrow

) and SSIM (

\uparrow

) are reported in RGB space. We repeat each experiment five times and report

\operatorname{mean}\pm\operatorname{std}

. The best performance is highlighted.

Fig. 7: Visual comparison of zero-shot and five-shot UniRestorer $++$ against specialist models. Five-shot fine-tuning improves degradation removal and detail recovery, as highlighted by the red boxes .

Routers	# Levels	$\{$ # DR Groups $\}$	PSNR (dB)
Routers	# Levels	$\{$ # DR Groups $\}$	In-dist.	Out-of-dist.
DER	1	$\{$ 1 $\}$	22.06	17.23
	1	$\{$ 2 $\}$	22.42	17.51
	1	$\{$ 4 $\}$	23.75	18.49
	1	$\{$ 8 $\}$	24.22	18.35
	1	$\{$ 16 $\}$	24.30	18.44

Table X: Ablation of DR fineness and the multi-granularity degradation set. Average PSNR (dB) is reported under in-dist. and out-of-dist. degradation parameters.

Fig. 8: Average selected hierarchy level and corresponding PSNR for the full and Vanilla variants as degradation parameters shift from in-dist. to out-of-dist.

Variant	Rain	Haze	Noise	Blur	Low-light	Snow	CAR	Avg.
Baseline	40.22	34.94	31.39	30.86	24.42	30.53	30.31	31.81
+ Dense DR	40.31	35.09	31.44	31.02	24.52	30.59	31.27	32.03
+ Region-Aware Expert Routing	40.35	35.51	31.48	31.59	24.64	30.66	31.48	32.24

Table XI: Ablation of the dense DR map and region-aware expert routing. PSNR (dB) is reported per task, and Average is the mean over the seven tasks.

$\ell_{1}$	$\mathcal{L}_{\mathrm{dg}}$	$\mathcal{L}_{\mathrm{aux}}$	PSNR (dB)
$\ell_{1}$	$\mathcal{L}_{\mathrm{dg}}$	$\mathcal{L}_{\mathrm{aux}}$	In-dist.	Out-of-dist.
✓	✗	✗	24.14	18.37
✓	✓	✗	24.35	18.76
✓	✓	✓	24.46	19.45

Table XII: Ablation of the loss functions.

Fig. 9: Routing frequencies on CDD-11. Left: degradation-estimation-based router (DER) assignments to the eight finest-level DR groups, whose dominant degradation types are shown on the horizontal axis. Right: granularity-estimation-based router (GER) assignments to the hierarchy levels. Rows correspond to the 11 CDD-11 degradation subsets.

Fig. 10: Visualization of the dense DR map for real-world spatially varying degradations, i.e . , motion blur (ReLoBlur [ 90 ] ), defocus blur (DPDD [ 84 ] ), and haze (FoundIR [ 30 ] ). Gray and black indicate relatively lower and higher degradation severity, respectively.

Image Restoration

Image restoration uses computational methods to recover useful visual content from noisy, blurred, low-resolution, or otherwise degraded images.

See all 386 papers

UnSCAR: Universal, Scalable, Controllable, and Adaptable Image Restoration
Mar 8, 2026Debabrata Mandal, Soumitri Chattopadhyay, Yujie Wang +2All-in-One Image Restoration Catastrophic Forgetting
University of North Carolina at Chapel Hill · University of California, San Diego
Divide and Restore: A Modular Task-Decoupled Framework for Universal Image Restoration
Mar 30, 2026Joanna Wiekiera, Martyna ZurImage Restoration All-in-One Image Restoration
Faculty of Automatic Control, Electronics and Computer Science, Silesian University of Technology, Gliwice, Poland
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

Mar 8, 2026J/K move · Enter open · S save

UnSCAR: Universal, Scalable, Controllable, and Adaptable Image Restoration

Debabrata Mandal, Soumitri Chattopadhyay, Yujie Wang, Marc Niethammer+1

University of North Carolina at Chapel Hill · University of California, San Diego

All-in-One Image Restoration Catastrophic Forgetting

Open paper AskPDF arXiv

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.

Abstract

Image Restoration

Explore similar work

UnSCAR: Universal, Scalable, Controllable, and Adaptable Image Restoration

Divide and Restore: A Modular Task-Decoupled Framework for Universal Image Restoration

Universal Image Restoration via Internalized Chain-of-Thought Reasoning