All-in-one image restoration aims to recover clean images blindly from inputs with various degradation types and levels using a single model. However, the adaptation policy in most existing methods remains implicitly encoded in feature representations, leaving the restoration decision neither explicitly represented nor independently controllable. Since real-world degradations vary spatially, applying uniform updates across the entire image leads to over-smoothing of clean areas and under-correction of localized degradations. To address these limitations, we present ExplicitIR, which represents the restoration policy as two explicit per-pixel fields: Budget, which determines where and how strongly to update, and Style, which selects between narrow and wide context windows. Specifically, ExplicitIR incorporates a Spatial Policy Estimator (SPE) to predict these fields, explicitly representing the restoration policy and capturing spatially-varying degradations, which are then applied by Policy-Adaptive Decoder Blocks (PADB) through adaptive fusion of narrow- and wide-window responses, enabling independent control and precise local modulation. This design enables spatially-adaptive restoration in a single forward pass, without requiring task labels, prompts, or test-time degradation estimators. Extensive experiments on three-task and five-task benchmarks demonstrate that ExplicitIR achieves performance competitive with the strongest all-in-one baselines.
Figures & tables
Figure 1: Motivation. (a) A hazy street scene. Several regions are highlighted with bounding boxes: the red box marks a heavily degraded region, while the blue box marks a mildly degraded region. (b) The absolute residual map against the clean reference, where brighter areas indicate more severe degradation. (c, d) Magnified views of the boxed regions. (e) One strong uniform correction applied everywhere: the tower is recovered, but the pavement is over-corrected. (f) One weak uniform correction: the pavement is preserved, but haze remains over the tower. This demonstrates that a single uniform strength cannot serve both regions, and the correction must vary with location.
Figure 2: Overview of ExplicitIR. (a) Overall architecture. A Restormer backbone equipped with a Spatial Policy Estimator (SPE) and Policy-Adaptive Decoder Blocks (PADBs). (b) SPE module. Multi-scale dilated convolutions ( D1,D2,D4 ) and dual independent heads predict explicit per-pixel Budget and Style fields. (c) PADB module. Narrow- and wide-window features are mixed via the Style field, and the result is scaled by the Budget field before injection.
Figure 3: Mechanism visualization of the Spatial Policy Estimator (SPE). Given the encoder features F0 , F2 and F3 , the Budget field d controls the update strength, and the Style field λ determines the window size. (a) Deraining: d concentrates on localized rain streaks. (b) Dehazing: d distributes broadly over the homogeneous haze.
Figure 4: Mechanism verification of the Policy-Adaptive Decoder Block (PADB) on (a) deraining and (b) dehazing inputs. Top row: Direction fields of the narrow-window branch Δ1nar , the wide-window branch Δ1wid , and their Style-mixed output m^1 under a shared PCA basis. Bottom row: The measured update magnitude, the predicted magnitude d1κ(1) , and their scatter plot. The points align strictly on the y=x diagonal.
Category
Method
Dehazing
Deraining
Denoising (CBSD68)
Average
SOTS
Rain100L
σ=15
σ=25
σ=50
Blind
DL [ 8 ]
26.92 / 0.931
32.62 / 0.931
33.05 / 0.914
30.41 / 0.861
26.90 / 0.740
29.98 / 0.875
FD-GAN [ 6 ]
24.71 / 0.924
29.89 / 0.933
30.25 / 0.910
28.81 / 0.868
26.43 / 0.776
28.02 / 0.883
MPRNet [ 44 ]
25.28 / 0.954
33.57 / 0.954
33.54 / 0.927
30.89 / 0.880
27.56 / 0.779
30.17 / 0.899
Restormer [ 45 ]
30.43 / 0.975
36.55 / 0.975
33.84 / 0.931
31.18 / 0.885
27.90 / 0.790
31.98 / 0.911
AirNet [ 15 ]
27.94 / 0.962
34.90 / 0.967
33.92 / 0.933
31.26 / 0.888
28.00 / 0.797
31.20 / 0.910
Table 1: Comparison with state-of-the-art on the three-task setting: dehazing on SOTS, deraining on Rain100L, and denoising on CBSD68 ( σ=15,25,50 ); PSNR (dB) / SSIM [ 37 ] on full RGB images. The first column separates blind methods, which restore from the degraded image alone, and externally-guided ones, which receive an additional signal at test time.
Category
Method
Dehazing
Deraining
Denoising
Deblurring
Low-light
Average
SOTS
Rain100L
σ=25
GoPro
LOL
DGUNet ∗ [ 26 ]
24.78 / 0.940
36.62 / 0.971
31.10 / 0.883
27.25 / 0.837
21.87 / 0.823
28.32 / 0.891
SwinIR ∗ [ 17 ]
21.50 / 0.891
30.78 / 0.923
30.59 / 0.868
24.52 / 0.773
17.81 / 0.723
25.04 / 0.835
Restormer ∗ [ 45 ]
24.09 / 0.927
34.81 / 0.962
31.49 / 0.884
27.22 / 0.829
20.41 / 0.806
27.60 / 0.881
NAFNet ∗ [ 2 ]
25.23 / 0.939
35.56 / 0.967
31.02 / 0.883
26.53 / 0.808
20.49 / 0.809
27.76 / 0.881
Blind
DL [ 8 ]
20.54 / 0.826
21.96 / 0.762
23.09 / 0.745
19.86 / 0.672
19.83 / 0.712
21.05 / 0.743
Table 2: Comparison with state-of-the-art on the five-task setting: dehazing on SOTS, deraining on Rain100L, denoising on CBSD68 ( σ=25 ), deblurring on GoPro, and low-light enhancement on LOL-v1. A star ( ∗ ) marks a general-purpose restorer trained separately for each task; the remaining methods are all-in-one models.
Figure 5: Qualitative comparison on the three-task setting for dehazing, deraining, and denoising. ExplicitIR removes the haze, rain streaks, and noise while restoring the underlying colors and structures, and avoids the artifacts of the compared methods. Zoom in for better view.
Figure 6: Editing the Budget field at inference. The columns show the degraded input, re-decodes with the Budget in the marked region clamped to 0 and 0.5, the model prediction, and forced to 2 and 4, and the clean reference. The restoration improves as the Budget rises toward the predicted value, but degrades when forced beyond the valid range (0,1) , demonstrating that the field directly controls the correction strength.
Figure 7: Editing the Style field globally at inference. The columns show the degraded input, re-decodes with the Style fixed to 0 (narrow window only) and 1 (wide window only), the model prediction, and the clean reference. Clamping to either extreme degrades restoration in a task-characteristic way: narrow-only leaves streaks and grain, while wide-only over-smooths the texture.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Item
Value
Architecture
Backbone
Restormer [ 45 ]
Backbone blocks per stage
[4,6,6,8]
Attention heads per stage
[1,2,4,8]
Encoder embedding dim
[48,96,192,384]
Decoder stages with PADB
3
Appendix
Table 3: Model architecture and training configuration.
Figure 8: Budget fields across decoder stages. Each row shows one degradation; the columns give the degraded input, the residual against the clean reference, and the Budget fields at stages 1 to 3, with each stage shown on its own scale and the numeric range indicated above each column. The allocation coarsens with depth: stage 1 follows fine image structure, while the deeper stages move to region-level updates.
Figure 9: Style fields across decoder stages, with the pooled Style distribution over each test set (rightmost column). Each row shows one degradation. The predicted Style stays close to the balanced point at stage 1 and follows image structure at the deeper stages.
Figure 10: Accuracy versus compute on the three-task setting: average PSNR against the number of operations measured on the official implementations at a 128×128 input. ExplicitIR is marked with a star, and the dashed line tracks its average.
Variant
Avg PSNR
Mask-gap
Module effectiveness
Full ExplicitIR
32.78
0.14
Backbone only
31.98
n/a
SPE only
32.08
0.03
PADB only
32.35
0.00
Field separation
Appendix
Table 4: Ablations on the three-task setting, grouped by the claim each variant tests. Module : whether the gain comes from backbone capacity, field prediction, or field execution; Backbone only removes SPE and PADB, SPE only predicts without execution, and PADB only executes without prediction. Separation : whether the two fields contribute independently and whether the readable form carries the gain; Free gating keeps per-pixel capacity but removes the readable form, and Budget-only / Style-only fix the other field. Variation : whether spatial variation matters beyond a global amplitude change, and whether the dual-window design is necessary. Every variant is trained with the reconstruction loss of Eq. (6).
All-in-one image restoration aims to recover clean images degraded by multiple corruption types using a single unified model. Existing methods typically rely on image-level prompts or shared guidance to handle diverse degradations. However, such a paradigm becomes inadequate when degradations are spatially heterogeneous or even coexist in mixed forms within a single image. Yet spatially adaptive guidance alone is not sufficient, since accurate restoration also requires each spatial query to reliably aggregate complementary information from local neighborhoods and global contexts. To this end, we propose QuReC, a unified framework for all-in-one image restoration. QuReC consists of a Degradation-Guided Query Reconstruction Module (DQRM) and a Local-Global Response Calibration Module (LGRCM). Specifically, DQRM matches each spatial query against a degradation prototype space to reconstruct a query-specific degradation-aware representation, thereby providing fine-grained spatially adaptive restoration guidance. To further stabilize this query-wise matching process, we introduce a weakly supervised prototype matching learning strategy to improve optimization stability and degradation semantic consistency. Meanwhile, LGRCM performs local-global dual-branch aggregation and calibrates the aggregated responses with learnable priors, improving the reliability of feature aggregation and the coordination between local detail modeling and global context modeling. Extensive experiments demonstrate that QuReC achieves superior performance on multiple all-in-one image restoration benchmarks. The code is released at https://github.com/zhoushen1/QuReC.
Shen Zhou, Jinghui Zhang, Wenbo Huang +7
Southeast University Nanjing, China · Hainan University Hainan, China
All-in-one image restoration is a unified low-level vision task that aims to effectively recover high-quality images from inputs degraded by various types and levels of corruption using a single model. Recent works have achieved remarkable progress by learning degradation-adaptive prompts or network architectures. However, these methods typically apply a uniform restoration strategy across the entire image, neglecting the fact that different regions may suffer from distinct degradation types and varying degrees of severity. In contrast, we propose to perform restoration at the pixel level, thereby enabling more fine-grained and precise control over the restoration process. Specifically, we present MGN-AIR, a novel pixel-level restoration framework for all-in-one image restoration. Our approach first learns to estimate a pixel-level visual prompt. Then, it leverages both textual and visual prompts to provide global and local degradation cues, guiding the model on where to look and how to restore at each pixel. We conduct extensive experiments on multiple all-in-one image restoration benchmarks, covering a wide range of tasks including denoising, deraining, deblurring, dehazing, desnowing, and low-light enhancement. Experimental results demonstrate that our proposed method consistently and significantly outperforms existing approaches.
All-in-one image restoration aims to handle diverse degradations within a unified framework. Existing methods commonly encode heterogeneous degradation conditions in a shared latent space, where degradation-related cues and scene content can remain entangled. We characterize the resulting challenge as dual ambiguity: semantic ambiguity in channel-wise modulation and spatial ambiguity in restoration responses, which can lead to content corruption and residual artifacts. To mitigate this issue, we propose DAR-Net, a Dual-Ambiguity Rectification Network for all-in-one image restoration. DAR-Net first introduces a Degradation Archetype Representation (DAR) module to construct a structured degradation state through simplex-constrained archetype mixture modeling. Based on this state, a Semantic Ambiguity Rectification (SeAR) module generates degradation-aware prompts to improve channel-wise conditioning in the decoder. A Spatial Ambiguity Rectification (SpAR) module further regularizes degradation-aware and complementary features toward orthogonal response subspaces, reducing spatial interference between removal and preservation cues. Extensive experiments on standard all-in-one restoration benchmarks show that DAR-Net achieves the best overall performance under both three-degradation and five-degradation settings, improving the average PSNR over the strongest competitor by 0.14 dB and 0.34 dB, respectively; it additionally shows superior performance on CDD-11 and WeatherBench.
Cencen Liu, Wen Yin, Dongyang Zhang +6
University of Electronic Science and Technology of China · University of Electronic Science and Technology of China, Chengdu, Sichuan, China · Jiigan Technology +1