acceptodds
Under review as a conference paper at ICLR 2027

ReDU-IR: Reconstructible Feature Decomposition and State Complementary Updating for All-in-One Blind Image Restoration

Abstract

All-in-One Image Restoration (AiOIR) aims to recover high-quality images from inputs corrupted by heterogeneous degradations using a single model. In hierarchical restoration networks, downsampling may discard information needed for reconstruction. Standard skip connections typically pass encoder information directly to the corresponding decoder stage without updating it based on the restored main state. To address these issues, we propose ReDU-IR. It combines Reconstructible Feature Decomposition (RFD) with State Complementary Updating (SCU). RFD decomposes encoder features into main and complementary representations. The matched inverse transform reconstructs features at the corresponding decoder stage. This design avoids irreversible information loss during scale transitions. SCU applies a shared mapping to the encoder and decoder main states. It uses the difference between the two mapping outputs to update the complementary representation without explicit degradation identification. Extensive experiments on multiple datasets demonstrate the effectiveness of our proposed method and its individual components.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
All-in-one blind image restoration, reconstructible feature decomposition, state complementary updating, hierarchical encoder-decoder networks

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_2aHFOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes ReDU-IR, an all-in-one blind image restoration framework based on Reconstructible Feature Decomposition (RFD) and State Complementary Updating (SCU). RFD decomposes encoder features into main and complementary representations and reconstructs them through an inverse transform. SCU updates the complementary features based on the differences between encoder and decoder main states. Experiments on multiple restoration benchmarks demonstrate competitive performance, particularly under composite degradations.

Critical strengths
  1. The paper addresses a meaningful problem of information preservation and feature reuse in hierarchical image restoration networks. The proposed RFD and SCU modules are well motivated and technically sound, particularly the use of paired feature reconstruction and state-dependent complementary updates.
  2. The experimental results are strong across several restoration benchmarks, particularly CDD11 and WeatherBench. The relatively comprehensive ablation studies further support the effectiveness of the proposed components.
Critical weaknesses
  1. The methodological novelty of RFD is somewhat limited, as invertible and wavelet-based representations have been explored in previous restoration methods. Moreover, the nearly perfect feature reconstruction is an expected property of invertible transformations, and its practical advantages over conventional feature reuse are not sufficiently demonstrated.
  2. The effectiveness of SCU and the consistency of the improvements require further justification. In Table 5, SCU provides only a 0.08 dB improvement in average PSNR, raising questions about its additional complexity. Furthermore, the method falls behind several existing approaches on deblurring and low-light enhancement in the 5D setting, suggesting that its benefits may be task-dependent.
Questions for authors
  1. What are the main advantages of RFD compared with existing wavelet-based or invertible feature decomposition methods, beyond the learnable orthogonal transformation?
  2. Given the relatively small improvements from SCU in Table 5, could the authors further explain why the state-dependent update is necessary compared with simpler feature fusion strategies?
  3. Could the authors provide additional analysis showing how the reconstruction property of RFD affects restoration quality, particularly the preservation of high-frequency details?
  4. Why does ReDU-IR show relatively limited improvements on deblurring and low-light enhancement compared with its stronger results on other degradation types?
3 comments
AuthorsOpenReview ↗

Response Part 1 of 3 — Q1

Thank you for your valuable comments, which have helped us improve our manuscript. We have revised the manuscript based on your feedback. Below, we provide details of our revisions. Please let us know whether we have addressed your concerns.

Q1:

Weakness 1: The methodological novelty of RFD is somewhat limited, as invertible and wavelet-based representations have been explored in previous restoration methods. Moreover, the nearly perfect feature reconstruction is an expected property of invertible transformations, and its practical advantages over conventional feature reuse are not sufficiently demonstrated.

Question 1: What are the main advantages of RFD compared with existing wavelet-based or invertible feature decomposition methods, beyond the learnable orthogonal transformation?

A1:

We thank you for this helpful comment. We agree that near-perfect reconstruction of unchanged features is an expected property of invertible transforms. The contribution of RFD lies in how it organizes information for restoration while retaining a matched reconstruction path.

MWCNN (Liu et al., 2018) uses DWT/IWT for resolution changes, while WINNet (Huang & Dragotti, 2022) combines a lifting-inspired transform (LINN) with detail-coefficient denoising. In RFD, a portion of the mixed high-frequency coefficients is processed jointly with the low-frequency component along the main restoration pathway; the remaining coefficients form a complementary representation for matched reconstruction. This design specifies which information receives deeper restoration and which information is retained outside that pathway. The learnable orthogonal transform adapts the high-frequency combinations assigned to each representation while retaining the matched inverse transform. In the full ReDU-IR model, SCU additionally updates the complementary representation before synthesis; its contribution is evaluated separately from RFD.

Without SCU, Table 5(a) shows that RFD improves PSNR over the SymUNet PixelUnshuffle baseline from 32.95 to 33.17 dB (+0.22 dB), with 0.34M fewer parameters. RFD also exceeds fixed Haar decomposition by 0.11 dB. Table 15 further shows RFD-only PSNR gains of 1.43 dB on CDD11 and 1.14 dB on WeatherBench. These results support the practical value of the RFD design in restoration.

We additionally compare RFD with DWT/IWT and LINN as replacement modules within SymUNet (dim=32), without SCU (Table 17 below). RFD achieves 31.62 dB/0.881 with the fewest parameters (9.87M) and FLOPs (35.07G). Its quality gain over LINN is modest (+0.06 dB/+0.001), accompanied by lower latency (17.11 versus 39.97 ms). DWT/IWT is faster (14.47 ms), while RFD improves PSNR/SSIM by 1.33 dB/0.015. This comparison supports a favorable balance between restoration quality and computational cost in the evaluated backbone.

Table 17. RFD versus DWT/IWT and LINN on WeatherBench (SymUNet, dim=32; without SCU).

ModulePSNR ↑ (dB)SSIM ↑Params. ↓ (M)FLOPs ↓ (G)Time ↓ (ms)
DWT/IWT (MWCNN)30.290.86612.7842.3114.47
LINN (WINNet)31.560.88012.1957.2039.97
RFD (Ours)31.620.8819.8735.0717.11

PSNR/SSIM are averaged over haze, rain, and snow. Params., FLOPs, and time refer to complete models. Time is latency in ms on an RTX 4090 with 256×256 inputs and batch size 1. FLOPs are reported using the THOP MAC-count convention.

We have clarified RFD’s design in the final paragraph of Section 2 (page 3) and added the component and wavelet-module comparisons in Appendix F.2 (Tables 15 and 17, page 19; analysis, pages 19–20).

AuthorsOpenReview ↗

Response Part 2 of 3 — Q2–Q3

Q2:

Weakness 2: The effectiveness of SCU and the consistency of the improvements require further justification. In Table 5, SCU provides only a 0.08 dB improvement in average PSNR, raising questions about its additional complexity. Furthermore, the method falls behind several existing approaches on deblurring and low-light enhancement in the 5D setting, suggesting that its benefits may be task-dependent.

Question 2: Given the relatively small improvements from SCU in Table 5, could the authors further explain why the state-dependent update is necessary compared with simpler feature fusion strategies?

A2:

Thank you for your suggestion. SCU updates the complementary features using the difference between shared RSR responses to the encoder and decoder main states. This ties the update to the change in the main representation during restoration. We have repeated all five SCU ablations using the same three random seeds and now report mean PSNR ± standard deviation in Tab. 5(b). Compared with concatenation-based fusion, SCU raises mean PSNR from 33.15 ± 0.025 to 33.22 ± 0.036 dB with 0.20M additional parameters. Decoder-only updating achieves 33.16 ± 0.045 dB at the same parameter count as SCU.

Table 5(b). SCU ablation in the ReDU-IR-B configuration under the 3D setting, with the learned RFD fixed. All five variants use the same training protocol and the same three random seeds. PSNR is reported as mean ± standard deviation.

VariantModule configurationParams. (M)PSNR (dB) ↑
w/o SCUDirect complementary feature reuse21.9533.13 ± 0.034
w/o encoder-side RSRDecoder-only RSR22.5433.16 ± 0.045
w/o SCU (fusion replacement)Concatenation-based fusion22.3433.15 ± 0.025
w/o RSR weight sharingSeparate encoder and decoder RSRs23.1433.19 ± 0.038
ReDU-IR (full model)SCU with shared RSR22.5433.22 ± 0.036

We have updated Table 5(b) and the “SCU Update Design” discussion in Section 4.2 (page 9) with three-seed statistics and comparisons with alternative update mechanisms.


Q3:

Question 3: Could the authors provide additional analysis showing how the reconstruction property of RFD affects restoration quality, particularly the preservation of high-frequency details?

A3:

We thank you for this suggestion. We further explain the role of reconstructibility in terms of information preservation and restoration performance. RFD distributes high-frequency information between the main and complementary representations and recombines them at the decoder through a matched inverse transform. This allows information that does not enter the main restoration pathway to still contribute to subsequent reconstruction, thereby preserving information that can support the restoration of fine details. When both the main and complementary representations are left unchanged, the relative reconstruction error for the illustrated sample is 9.25×10−89.25\times10^{-8}, verifying this information-preserving property.

In restoration experiments, Table 15 shows that adding RFD without SCU improves PSNR by 1.43 dB on CDD11 and 1.14 dB on WeatherBench, supporting the contribution of the RFD module to restoration quality. Table 5(a) further shows that, while maintaining a reconstructible decomposition, RFD with learnable high-frequency allocation improves PSNR by 0.11 dB over fixed Haar decomposition. Together, these results support the RFD design in terms of both information preservation and restoration performance: matched reconstruction retains detail information for subsequent restoration, while learnable high-frequency allocation adapts the organization of this information to the restoration objective.

The reconstruction evidence in Section 3.2/Figure 3 (page 5) and Table 5(a) (Section 4.2, page 9) was already in the submission. We have added the component ablation in Table 15 (Appendix F.2, page 19).

AuthorsOpenReview ↗

Response Part 3 of 3 — Q4

Q4:

Question 4: Why does ReDU-IR show relatively limited improvements on deblurring and low-light enhancement compared with its strong results on other degradation types?

A4:

We thank you for this observation. We agree that ReDU-IR does not achieve the best results on deblurring and low-light enhancement, but it is useful to distinguish improvements over the baseline from rankings against other methods on individual tasks. Compared with SymUNet, ReDU-IR-B improves deblurring PSNR from 28.12 to 28.75 dB (+0.63 dB) and SSIM from 0.855 to 0.871 (+0.016), showing clear improvements in both metrics for this task.

The five-task results also show that different unified restoration methods have different task-specific strengths. For example, MoCE-IR achieves the highest deblurring PSNR, whereas ReDU-IR-B achieves higher PSNR on each of the other four tasks. MIRAGE-S and DFPIR achieve higher low-light PSNR, but perform worse than ReDU-IR-B on several other tasks. ReDU-IR-B achieves the highest average PSNR across the five tasks, at 31.03 dB. These results indicate strong overall restoration performance, although the gains are uneven across tasks.

On low-light enhancement, PSNR increases by 0.20 dB over SymUNet, while SSIM decreases by 0.001. These comparisons establish uneven task-specific gains but do not isolate their causes.

5D comparison from revised Table 2. Entries are PSNR (dB) / SSIM; bold indicates the better score between the two models. Denoising is evaluated on CBSD68 (σ = 25).

ModelParams. (M)DehazingDerainingDenoisingDeblurringLow-LightAverage
SymUNet22M31.31/0.97938.05/0.98131.38/0.89128.12/0.85523.27/0.85830.43/0.913
ReDU-IR-B (Ours)23M32.29/0.98239.10/0.98531.56/0.89428.75/0.87123.47/0.85731.03/0.918

We have added SymUNet (22M), its per-task and average PSNR/SSIM scores, and the Du et al. (2026) citation to Tables 1 and 2 in Section 4.1 (pages 7–8).

Reviewer_j7cqOpenReview ↗
Rating
3(+1)
Confidence
3(−1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ReDU-IR, an all-in-one blind image restoration framework designed to address information preservation and effective feature reuse across scales in hierarchical encoder–decoder networks. The proposed method consists of two core components: Reconstructible Feature Decomposition (RFD) and State Complementary Updating (SCU). RFD employs the Haar wavelet transform and a learnable orthogonal transform to decompose encoder features into main and complementary representations. The main representation is processed by the deeper restoration network, while the complementary representation is retained for subsequent feature reconstruction. SCU dynamically updates the complementary representation using the difference between shared mapping responses to the encoder and decoder main states, allowing it to adapt to changes in the main representation during restoration. The updated representations are then combined through inverse transforms to reconstruct the features. Experimental results demonstrate that ReDU-IR achieves competitive restoration performance across various single and composite degradation tasks.

Critical strengths

(1) The paper presents a plausible mechanism for cross-scale feature decomposition and dynamic updating.The paper combines Reconstructible Feature Decomposition (RFD) and State Complementary Updating (SCU) to improve feature propagation in hierarchical encoder–decoder networks. The two serve complementary roles and provide an alternative design for feature propagation in conventional encoder–decoder networks.

(2) The authors evaluate the proposed method under four experimental settings: 3D, 5D, CDD11, and WeatherBench, covering single degradations, composite degradations, and real-world adverse weather conditions. The results indicate that the proposed method achieves competitive restoration performance across multiple benchmarks. The paper also includes component ablation studies, sensitivity analysis of the update strength, and computational complexity comparisons.

Critical weaknesses

(1) RFD uses a learnable orthogonal transform to achieve reconstructible feature decomposition, but the authors do not sufficiently justify the chosen allocation ratio of high-frequency channels.

(2) The SCU update depends only on changes in the main representation and does not directly account for degradation-related information in the complementary representation. When high-frequency degradation information, such as noise and rain streaks, is mainly retained in the complementary representation but is not adequately reflected in the main representation, can SCU effectively correct these degradations? Table 5 shows that SCU improves PSNR by only 0.08 dB over directly reusing complementary features and by 0.05 dB over using only the decoder-side RSR. These results do not sufficiently establish its advantages over these alternatives.

(3) All qualitative comparisons are presented in the appendix, with no representative restoration results included in the main paper, making it difficult to directly assess the visual quality of the restored images.

Questions for authors

(1) RFD assigns all low-frequency features and part of the high-frequency features to the main representation, while retaining the remaining high-frequency features as the complementary representation. What is the rationale for the chosen high-frequency channel allocation ratio? Could the authors provide ablation studies with different partitioning ratios and explain how they verify that the main representation learns features that are more beneficial for image restoration?

(2) The SCU update depends solely on the difference between the mapped responses of the main representation before and after restoration. When high-frequency degradation-related information, such as noise and rain streaks, is primarily retained in the complementary representation but is not adequately reflected in the main representation, how can SCU effectively correct these degradations? Could the authors provide feature visualizations before and after the update and compare SCU with alternative feature updating mechanisms?

(3) Why do the qualitative comparisons for CDD11 composite degradation and single-task restoration include only MIRAGE? I suggest that the authors include representative qualitative results against recent competitive methods.

(4) In Table 3, the PSNR and SSIM values reported for ExDA and MIRAGE-S are identical across the two triple-degradation settings, L+H+R and L+H+S. The authors should verify these results and clarify whether the identical values are expected.

7 comments
AuthorsOpenReview ↗

Response Part 1 of 3 — Q1

Thank you for your valuable comments, which have helped us improve our manuscript. We have revised the manuscript based on your feedback. Below, we provide details of our revisions. Please let us know whether we have addressed your concerns.

Q1:

Weakness 1: RFD uses a learnable orthogonal transform to achieve reconstructible feature decomposition, but the authors do not sufficiently justify the chosen allocation ratio of high-frequency channels.

Question 1: RFD assigns all low-frequency features and part of the high-frequency features to the main representation, while retaining the remaining high-frequency features as the complementary representation. What is the rationale for the chosen high-frequency channel allocation ratio? Could the authors provide ablation studies with different partitioning ratios and explain how they verify that the main representation learns features that are more beneficial for image restoration?

A1:

Thank you for your suggestions. We use a high-frequency allocation ratio of 1/3 as an empirical choice balancing restoration quality and model size. We have added a WeatherBench allocation-ratio ablation with five ratios—0, 1/6, 1/3, 1/2, and 2/3—using RFD variants trained without SCU under the same training protocol (Table 16). The 1/3 setting achieves 31.62 dB with 9.87M parameters, compared with 31.97 dB and 31.76M parameters for 1/2. Thus, the default setting uses 68.92% fewer parameters with a PSNR difference of 0.35 dB. Increasing the ratio to 2/3 further improves PSNR to 32.24 dB, with the parameter count increasing to 86.65M.

The ratio specifies the number of high-frequency channels entering the main representation, while their combinations are learned through the orthogonal transform under the restoration objective. Across all tested ratios, the main and complementary representations jointly retain all transformed coefficients, preserving the reconstruction property of the paired decomposition. The existing 3D ablation (Table 5) compares fixed Haar decomposition with learnable RFD, yielding 33.06 and 33.17 dB, respectively. We have also added component ablations on CDD11 and WeatherBench (Table 15). Adding RFD to SymUNet increases PSNR from 29.98 to 31.41 dB on CDD11 and from 30.48 to 31.62 dB on WeatherBench. These experiments provide task-level evidence for the contribution of RFD and its learned allocation.

Table 15 (revised manuscript). Component ablation on CDD11 and WeatherBench.

VariantCDD11 PSNR (dB) ↑CDD11 SSIM ↑WeatherBench PSNR (dB) ↑WeatherBench SSIM ↑
SymUNet (baseline)29.980.89030.480.866
+ RFD31.410.89531.620.881
Ours31.550.89731.800.883
  • RFD denotes SymUNet with RFD and without SCU; Ours includes both RFD and SCU.

Table 16 (revised manuscript). High-frequency allocation-ratio ablation on WeatherBench.

ρPSNR (dB) ↑SSIM ↑Params. (M) ↓
030.370.8690.62
1/630.720.8732.57
1/3†31.620.8819.87
1/231.970.88631.76
2/332.240.88886.65

All ratio variants were trained without SCU under the same training protocol. † denotes the default ratio; Params. gives the total parameter count.

We have added Tables 15 and 16 and the analysis of the default allocation ratio in Appendix F.2 (page 19).

AuthorsOpenReview ↗

Response Part 2 of 3 — Q2

Q2:

Weakness 2: The SCU update depends only on changes in the main representation and does not directly account for degradation-related information in the complementary representation. When high-frequency degradation information, such as noise and rain streaks, is mainly retained in the complementary representation but is not adequately reflected in the main representation, can SCU effectively correct these degradations? Table 5 shows that SCU improves PSNR by only 0.08 dB over directly reusing complementary features and by 0.05 dB over using only the decoder-side RSR. These results do not sufficiently establish its advantages over these alternatives.

Question 2: The SCU update depends solely on the difference between the mapped responses of the main representation before and after restoration. When high-frequency degradation-related information, such as noise and rain streaks, is primarily retained in the complementary representation but is not adequately reflected in the main representation, how can SCU effectively correct these degradations? Could the authors provide feature visualizations before and after the update and compare SCU with alternative feature updating mechanisms?

A2:

Thank you for raising this question. We have clarified the relationship between the paired representations and added feature visualizations and component ablations to explain SCU's update mechanism and evaluate its contribution.

First, the main representation also contains high-frequency information. RFD applies learnable orthogonal mixing to the high-frequency coefficients and allocates a subset to the main representation together with the low-frequency coefficients. The remaining coefficients form the complementary representation. This allocation allows high-frequency cues to participate in the main restoration pathway. SCU applies a shared mapping to the original and restored main representations and uses the response difference to update the complementary representation before reconstruction. Changes in the paired main states provide cues for updating the complementary features through the learned relationship between the two representations. If the mapped main-state responses are unchanged, the SCU correction is zero.

Second, following your suggestion, Fig. 5 now visualizes the complementary features before and after SCU, the update maps, restored outputs, and ground truth. The examples cover rain streaks on Rain100L and Gaussian noise on CBSD68 with σ = 25. The update maps show spatially structured responses, while recognizable scene structures remain visible in the updated features. These visualizations make SCU's modifications explicit and provide qualitative evidence alongside the quantitative evaluation. Together with the ablations, these examples provide evidence for the evaluated samples and tasks, without guaranteeing correction of every degradation confined to the complementary representation.

Finally, the revised Table 5(b) reports means and standard deviations over the same three random seeds for the five existing update-design variants. SCU improves mean PSNR by 0.09 dB over direct complementary-feature reuse and by 0.06 dB over decoder-side RSR. To further assess SCU's contribution, we added the component ablation in Table 15, reproduced below. Adding SCU to RFD improves PSNR from 31.41 to 31.55 dB on CDD11 and from 31.62 to 31.80 dB on WeatherBench, with SSIM increasing by 0.002 on both datasets. Table 5 provides direct comparisons with alternative feature processing schemes, while Table 15 further evaluates SCU's additional contribution on top of RFD.

Table 15. Component ablation on CDD11 and WeatherBench. The RFD variants use the default high-frequency allocation ratio ρ = 1/3.

VariantCDD11 PSNR (dB) ↑CDD11 SSIM ↑WeatherBench PSNR (dB) ↑WeatherBench SSIM ↑
SymUNet (baseline)29.980.89030.480.866
+ RFD31.410.89531.620.881
Ours (RFD + SCU)31.550.89731.800.883

We have clarified the paired representations in Section 2 (page 3), updated Table 5(b) and the SCU discussion in Section 4.2 (page 9), and added Figure 5 with feature-update analysis (page 10) and Table 15 (Appendix F.2, page 19).

AuthorsOpenReview ↗

Response Part 3 of 3 — Q3–Q4

Q3:

Weakness 3: All qualitative comparisons are presented in the appendix, with no representative restoration results included in the main paper, making it difficult to directly assess the visual quality of the restored images.

Question 3: Why do the qualitative comparisons for CDD11 composite degradation and single-task restoration include only MIRAGE? I suggest that the authors include representative qualitative results against recent competitive methods.

A3:

Thank you for your suggestions.

We have moved the WeatherBench haze/rain/snow comparison from the appendix to Figure 4 (page 9), with analysis in Section 4.1 (page 7). We have added methods to the CDD11 and single-task comparisons in Appendix J, Figures 9–11 and 13 (pages 21–23).


Q4:

Question 4: In Table 3, the PSNR and SSIM values reported for ExDA and MIRAGE-S are identical across the two triple-degradation settings, L+H+R and L+H+S. The authors should verify these results and clarify whether the identical values are expected.

A4:

Thank you for pointing this out. We checked the entries in Table 3 against the corresponding original papers and confirmed that both ExDA and MIRAGE-S report 25.59/0.801 for L+H+R and 25.86/0.799 for L+H+S.

We have verified the relevant entries in Table 3, Section 4.1 (page 8); they match the cited papers and remain unchanged.

Reviewer_xpwLOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

ReDU-IR combines a reconstructible Haar-based feature decomposition with a complementary-feature update computed from a shared mapping of encoder and decoder states. Experiments cover four all-in-one restoration settings, including composite degradations and real-world adverse weather, with ablations examining decomposition choices and update designs.

Critical strengths
  1. The decomposition and update fit together naturally: reconstruction is exact when both representations remain unchanged, and the complementary correction vanishes when the main state does not change. This gives the architecture a clear, testable structure.
  2. Improvements extend across four benchmark settings, with particularly substantial gains on CDD11 and consistent gains on WeatherBench. Achieving these results with the 10M-parameter variant supports the method’s practical value.
Critical weaknesses
  1. The architectural novelty is modest given prior wavelet and lifting-based invertible restoration methods. The distinction from approaches such as WINNet needs sharper analysis beyond the new terminology and combination of components.
  2. SCU improves over decoder-only updating by 0.05 dB and separate encoder–decoder mappings by just 0.02 dB in Table 5. Without variability across training seeds, these differences do not firmly establish the claimed advantage of the shared response-difference design.
  3. Matched backbone and retrained component ablations are limited to the 3D setting. The much larger CDD11 gain lacks a matched SymUNet comparison, while disabling SCU in an already trained model cannot replace that comparison.
  4. All eleven CDD11 degradation combinations are included during training. These results support restoration of known mixtures, but leave generalization to unseen combinations and degradation families untested.
Questions for authors
  1. What is the variation across training seeds for the SCU alternatives in Table 5, particularly the 0.02–0.05 dB differences?
  2. How do matched SymUNet, RFD-only, and full ReDU-IR models compare on CDD11 under identical training settings?
3 comments
AuthorsOpenReview ↗

Response Part 1 of 2 — Q1–Q2

Thank you for your valuable comments, which have helped us improve our manuscript. We have revised the manuscript based on your feedback. Below, we provide details of our revisions. Please let us know whether we have addressed your concerns.

Q1:

Weakness 1: The architectural novelty is modest given prior wavelet and lifting-based invertible restoration methods. The distinction from approaches such as WINNet needs sharper analysis beyond the new terminology and combination of components.

A1:

Thank you for your suggestion. WINNet focuses on lifting-inspired multiscale reconstructible transforms and detail-coefficient denoising. ReDU-IR addresses how to adjust the complementary representation that is paired with the main representation for reconstruction after the latter undergoes restoration. Specifically, RFD uses learnable orthogonal mixing to allocate high-frequency information between the main restoration pathway and the complementary representation. SCU then updates the complementary representation using the response difference of the same shared mapping evaluated at the main states before and after restoration, making the update explicitly dependent on the change in the main state.

To examine the distinction from existing wavelet and lifting modules, Table 17 compares RFD with DWT/IWT from MWCNN and LINN from WINNet within the same backbone. RFD achieves a PSNR of 31.62 dB, compared with 30.29 dB for DWT/IWT and 31.56 dB for LINN. Relative to LINN, RFD achieves slightly higher restoration accuracy while reducing the parameter count from 12.19M to 9.87M and FLOPs from 57.20G to 35.07G, demonstrating the trade-off between restoration accuracy and computational cost offered by this decomposition design. FLOPs are reported using the THOP MAC-count convention.

Table 5 further examines the effects of the specific design choices: without complementary updates, learnable high-frequency mixing improves PSNR from 33.06 to 33.17 dB over fixed Haar decomposition. With the same 22.54M parameters, SCU improves the mean PSNR over three runs from 33.16 to 33.22 dB compared with updating based only on the decoder-side state. In addition, Table 15 shows the incremental gains from RFD and SCU on CDD11 and WeatherBench, while Figure 5 visualizes the spatial distribution of complementary-feature updates, providing an illustration of the mechanism.

We have added the discussion of distinctions from prior methods to Section 2, "Reconstructible Representations and Feature Reuse" (p. 3); updated the SCU ablations with three-seed statistics and the corresponding analysis in Section 4.2 (Table 5(b), p. 9), and added feature visualizations (Figure 5, p. 10); and added component ablations across datasets and wavelet-module comparisons to Appendix F.2 (Tables 15 and 17 and the corresponding analysis, pp. 19-20).


Q2:

Weakness 2: SCU improves over decoder-only updating by 0.05 dB and separate encoder-decoder mappings by just 0.02 dB in Table 5. Without variability across training seeds, these differences do not firmly establish the claimed advantage of the shared response-difference design.

Question 1: What is the variation across training seeds for the SCU alternatives in Table 5, particularly the 0.02–0.05 dB differences?

A2:

Thank you for your suggestion. We have repeated the five SCU ablations using the same three random seeds and now report the mean PSNR and standard deviation in Tab. 5(b). SCU achieves 33.22 ± 0.036 dB, compared with 33.16 ± 0.045 dB for decoder-only updating and 33.19 ± 0.038 dB for separate encoder and decoder RSRs. The mean improvements are 0.06 and 0.03 dB, respectively. We have updated Tab. 5 and the corresponding analysis in Section 4.2 on page 9 of the revised manuscript.

Table 5(b). SCU ablation in the ReDU-IR-B configuration under the 3D setting, with the learned RFD fixed. All five variants use the same training protocol and the same three random seeds. PSNR is reported as mean ± standard deviation.

VariantModule configurationParams. (M)PSNR (dB) ↑
w/o SCUDirect complementary feature reuse21.9533.13 ± 0.034
w/o encoder-side RSRDecoder-only RSR22.5433.16 ± 0.045
w/o SCU (fusion replacement)Concatenation-based fusion22.3433.15 ± 0.025
w/o RSR weight sharingSeparate encoder and decoder RSRs23.1433.19 ± 0.038
ReDU-IR (full model)SCU with shared RSR22.5433.22 ± 0.036
AuthorsOpenReview ↗

Response Part 2 of 2 — Q3–Q4

Q3:

Weakness 3: Matched backbone and retrained component ablations are limited to the 3D setting. The much larger CDD11 gain lacks a matched SymUNet comparison, while disabling SCU in an already trained model cannot replace that comparison.

Question 2: How do matched SymUNet, RFD-only, and full ReDU-IR models compare on CDD11 under identical training settings?

A3:

Thank you for your suggestion. We have added a component comparison on CDD11 in Tab. 15. SymUNet, RFD-only (SymUNet with RFD and without SCU), and full ReDU-IR (with both RFD and SCU) achieve 29.98, 31.41, and 31.55 dB, respectively. The experimental results and component analysis are provided in Appendix F.2 on page 19 of the revised manuscript.

Table 15. Component ablation on CDD11 and WeatherBench. The RFD variants use the default high-frequency allocation ratio ρ = 1/3.

VariantCDD11 PSNR (dB) ↑CDD11 SSIM ↑WeatherBench PSNR (dB) ↑WeatherBench SSIM ↑
SymUNet (baseline)29.980.89030.480.866
+ RFD31.410.89531.620.881
Ours31.550.89731.800.883
  • RFD denotes SymUNet with RFD and without SCU; Ours includes both RFD and SCU.

Q4:

Weakness 4: All eleven CDD11 degradation combinations are included during training. These results support restoration of known mixtures, but leave generalization to unseen combinations and degradation families untested.

A4:

Thank you for your suggestion. We have added generalization experiments using three-degradation models. Tab. 18 reports denoising results at unseen noise levels of σ = 60 and 75, and Tab. 19 reports results on the unseen desnowing task on CSD. Our method achieves the highest PSNR among the compared methods in all three tests. The experimental settings and results are provided in Appendix H, “Generalization to Unseen Degradations,” on page 20 of the revised manuscript. These tests cover unseen degradation severities and an unseen degradation family (desnowing). Generalization to unseen degradation combinations remains untested and is listed as future work in Appendix I on page 21.

Thanks for the careful response. I'll keep my positive score.

Comments on the paper

1 comment
Area_Chair_s2FdOpenReview ↗

AC Reminder: Please Follow Up on the Authors’ Responses

Dear Reviewers,

Thank you for your time and thoughtful reviews. The authors have posted responses to your questions and concerns.

If you have not yet done so, please read their responses and any relevant revisions, and leave a follow-up comment visible to the authors indicating which concerns have been addressed and which remain unresolved. Please update your review and rating if warranted. If your assessment remains unchanged, a brief explanation would still be helpful.

Thank you for your continued engagement in the discussion.

Best regards,

Area Chair

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.