acceptodds
Under review as a conference paper at ICLR 2027

Degradation Frequency Curve: An Explicit Frequency-Quantified Representation for All-in-One Image Restoration

Abstract

A fundamental difficulty in all-in-one blind image restoration is that degradation is usually treated as an implicit factor hidden in degraded-to-clean mapping, rather than as an explicit object that can be measured and manipulated. This limitation becomes more pronounced under mixed, compound, or unseen degradation conditions, where degradation effects are hard to assign to predefined labels or task-specific parameters. We propose the Degradation Frequency Curve (DFC), a structured spectral representation that quantifies degradation responses by measuring band-wise residual-to-degraded energy ratios in the frequency domain. DFC maps visually entangled degradation responses into a measurable degradation coordinate space. Moreover, DFC can be adaptively decomposed into band-wise spectral tokens, allowing local degradation responses to be represented as reusable restoration priors. Based on this representation, we develop the DFC-guided Image Restorer (DFC-IR), a token-conditioned multi-scale framework that progressively estimates DFCs from intermediate restorations and uses the resulting spectral tokens to guide degradation-aware restoration in a coarse-to-fine manner. Extensive experiments on standard, composite, unseen, and real-world degradation benchmarks show that DFC provides an effective representation basis for all-in-one restoration, leading to state-of-the-art performance and improved generalization under complex degradation profiles.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Degradation Frequency Curve, spectral tokens, degradation representation, all-in-one image restoration.

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_obesOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes the Degradation Frequency Curve (DFC), an explicit spectral representation that describes image degradation using band-wise residual-to-degraded energy ratios. Based on DFC, the authors develop DFC-IR, which converts scale-specific degradation curves into spectral tokens and uses them to guide coarse-to-fine image restoration. The method aims to improve restoration under single, composite, and unseen degradation conditions without requiring degradation labels during inference. Experiments on standard all-in-one benchmarks, composite degradations, unseen tasks, and real-world images show competitive performance, while the ablation studies and quantitative analyses support the effectiveness of the proposed degradation representation.

Critical strengths

The paper provides a meaningful new perspective on degradation modeling. Instead of representing degradation implicitly through features or prompts, it models degradation as an explicit frequency curve that can be measured, compared, and decomposed into reusable spectral tokens.

Critical weaknesses
  1. DFC depends on intermediate restoration quality during inference. Since clean images are unavailable, inaccurate early restoration results may lead to inaccurate DFC estimation, especially for unseen or complex degradations.

  2. The analysis of DFC properties is limited. Its content robustness and degradation sensitivity are mainly validated on synthetic haze, rain, and noise, without equivalent analysis for blur, low-light, snow, or composite degradations.

Questions for authors
  1. DFC construction and training supervision rely on clean ground-truth images. In real-world datasets, the ground truth may contain noise, misalignment, or remaining degradation, so the robustness of DFC-IR to imperfect ground truth is unclear.

  2. During inference, DFC is estimated from intermediate restoration results. It remains unclear how estimation errors from early stages affect the subsequent restoration process, especially under unseen degradations.

4 comments
AuthorsOpenReview ↗

Response to reviewer obes (W1, W2, Q2)

We sincerely appreciate your insightful comments and questions, which have greatly contributed to improving our work. We address your concerns as follows.

[W1, Q2] Impact of Early-Stage DFC Estimation Errors on Subsequent Restoration

We evaluate the impact of early-stage DFC estimation errors on subsequent restoration under unseen degradations (UIEB and CSD) using the pretrained five-degradation model. Specifically, we control the DFC estimation error at the early decoding stage (s2) by interpolating between the estimated and reference DFCs:

Dλs2=(1−λ)Dests2+λDrefs2,D_{\lambda}^{s2}=(1-\lambda)D_{\mathrm{est}}^{s2}+\lambda D_{\mathrm{ref}}^{s2} ,

while the subsequent stage (s1) estimates its DFC normally without intervention. The scale-wise DFC errors and restoration PSNR are reported below.

Tab. R4-1. Effect of Early-Stage DFC Estimation Errors on Subsequent Restoration under Unseen Degradations.

DatasetDFC Conditions2: NMAE↓ / W₁↓s1: NMAE↓ / W₁↓PSNR↑
UIEBEst. (λ=0)0.0551 / 0.01980.0494 / 0.017619.24
Interp. (λ=0.3)0.0386 / 0.01390.0369 / 0.013519.27
Interp. (λ=0.6)0.0220 / 0.00790.0207 / 0.007319.32
Ref. (λ=1)0 / 00.0114 / 0.003919.36
CSDEst. (λ=0)0.0370 / 0.01260.0342 / 0.011221.05
Interp. (λ=0.3)0.0259 / 0.00880.0244 / 0.008021.07
Interp. (λ=0.6)0.0148 / 0.00500.0141 / 0.004721.11
Ref. (λ=1)0 / 00.0089 / 0.002921.17

The results show that reducing early-stage DFC estimation errors improves subsequent estimation accuracy and restoration performance on both UIEB and CSD. This confirms that more accurate early-stage DFC guidance leads to better subsequent restoration under unseen degradations.

[W2] Analysis of DFC Properties across More Degradation Types

We extend the DFC property analysis to blur, low-light, snow, and composite haze+snow degradations on the MiO100, visualizing the mean normalized DFC profiles with standard deviation bands. Within each degradation, the standard deviation bands remain narrow , while the mean normalized DFC profiles differ across degradation types. The extended visualizations are provided in Fig. 12 (Appendix B.5) of the revision.

We further conduct a quantitative analysis of DFC's content robustness and degradation sensitivity across all seven degradation types using the Wasserstein-1 distance. The results are reported below.

Tab. R4-2. Within- and Between-Degradation Wasserstein-1 Distances across Seven Degradation Types.

MetricHazeRainNoiseBlurLow-lightSnowHaze+SnowAverage
Within Dist.↓0.01920.04110.02590.02430.01770.04640.07110.0351
Between Dist.↑0.14770.11430.13350.11840.15360.16240.21060.1486

The results show that the between-degradation distance consistently exceeds the within-degradation distance across all seven degradation types, with averages of 0.1486 and 0.0351, respectively, yielding a separation ratio of 4.23. These results provide additional evidence for the content robustness and degradation sensitivity of DFC. We have added this analysis to Table 17 (Appendix B.5) of the revision.

AuthorsOpenReview ↗

Response to reviewer obes (Q1)

[Q1] Impact of Imperfect Ground Truth

We clarify that the DFCs used for restoration guidance during both training and inference are estimated from degraded images and intermediate restorations rather than directly constructed from ground truth. If real-world datasets are used only for zero-shot evaluation without training or fine-tuning, imperfections in their ground truth do not affect DFC estimation or the restoration process. However, if imperfect ground truth is used for training, it may affect supervision and potentially influence restoration performance. This is a common challenge for supervised image restoration methods. Investigating robustness to imperfect training targets remains an important direction for future work.

We sincerely appreciate the careful thought behind your comments and your recognition of the value of explicit frequency-based degradation modeling. Your questions encouraged us to more directly investigate how early-stage DFC estimation affects subsequent restoration, extend the analysis of DFC properties to a broader range of degradations, and clarify the impact of imperfect ground truth. These additional analyses and clarifications have strengthened the experimental foundation of our work and made our discussion more comprehensive and rigorous. We are truly grateful for your constructive guidance in improving our paper. We hope the new evidence and clarifications address your concerns and would sincerely appreciate it if you could reconsider your assessment.

The authors' response has addressed my concerns. Therefore, I have decided to maintain my original score.

We are pleased that our response has addressed your concerns. We sincerely appreciate your thoughtful review and constructive suggestions, which have helped us strengthen our analysis and improve the manuscript. Thank you again for your time and consideration.

Reviewer_wxAxOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces the Degradation Frequency Curve (DFC), which describes degradation through the ratio of residual energy to degraded-image energy across frequency bands. The proposed restoration network, DFC-IR, estimates these curves from intermediate restoration outputs and uses them to adaptively extract frequency-band tokens. These tokens guide feature modulation and aggregation during coarse-to-fine decoding. The method is evaluated on standard all-in-one restoration benchmarks, composite degradations, unseen degradation types and severities, and real-world images.

Critical strengths
  1. The degradation representation has a clear meaning. Each DFC value measures a specific band-wise energy ratio. The paper also examines content sensitivity, differences between degradation types, and responses to degradation severity, providing empirical support for this representation.

  2. The evaluation covers a useful range of settings. Besides standard three- and five-degradation benchmarks, the paper tests composite degradations, unseen tasks, and real-world images. The reported results show the highest average PSNR in Tables 1 and 2 and encouraging generalization to underwater enhancement and desnowing without retraining.

  3. The ablations examine several important parts of the method. The paper studies the two proposed modules, token count, alternative conditioning strategies, and DFC estimation accuracy under different training objectives. These experiments provide useful evidence about how the complete system works.

Critical weaknesses
  1. The reliability of estimated DFCs under challenging degradations needs further examination. The representation analysis uses clean reference images, whereas the restoration network estimates DFCs from its own intermediate predictions. These estimates can therefore include restoration errors. Table 9 provides useful accuracy measurements, but does not break them down by decoding stage or challenging degradation setting. It remains unclear how reliably DFC guides restoration when the intermediate predictions are poor.

  2. The specific contribution of the DFC representation is not fully isolated. The ablations support the complete approach, but the multi-band decoder already accounts for a substantial improvement. Table 8 compares DFC-guided tokens with uniform tokens and learned prompts, while Table 13 compares spectral representations through separation statistics. Comparing DFC, residual spectra, and degraded-image spectra as guidance within the same restoration architecture would more directly establish the value of the proposed ratio formulation.

  3. The performance gains are uneven across tasks, while the model is larger than several competitive baselines. Although Tables 1 and 2 report the highest average PSNR, the clearest advantage over the strongest task-wise baselines is in dehazing. Deraining, denoising, and deblurring PSNR do not consistently lead. The model also has more parameters than MoCE-IR, AdaIR, and DFPIR. Comparisons at similar parameter or computational budgets would help assess the benefit relative to its cost.

  4. The sensitivity to the fixed frequency design remains unclear. DFC construction uses predefined radial Gaussian masks, 25 measurement bands, and a prescribed bandwidth schedule. Token count is studied, but these construction choices are not similarly evaluated. The real-world results are encouraging; further analysis would help establish whether the representation remains reliable across these design choices and provides useful guidance for spatially varying or directional degradations.

Questions for authors

As described in "Critical Weaknesses".

7 comments
AuthorsOpenReview ↗

Response to reviewer wxAx (W1, W2)

Thank you for your thoughtful review and valuable feedback. We address your concerns in detail below.

[W1] Reliability of DFC Estimation under Challenging Degradations

We appreciate your concern regarding the reliability of DFC estimation from imperfect intermediate restorations. Tab. 9 of the original submission reports DFC estimation errors averaged over scales 2 and 1. At scale 3, uniform frequency partitioning is used without DFC estimation. We provide the scale-wise NMAE and W1W_1 results below and have added them to Tab. 9 in the revision.

Tab. R3-1. Scale-wise DFC Estimation Errors under Different Training Objectives.

VariantMSLf\mathcal{L}_{f}s2: NMAE↓ / W₁↓s1: NMAE↓ / W₁↓
(a)✗✗0.3977 / 0.11940.3147 / 0.0938
(b)✗✓0.2943 / 0.08840.2609 / 0.0786
(c)✓✗0.0594 / 0.02170.0492 / 0.0175
(d)✓✓0.0175 / 0.00510.0152 / 0.0043

To further assess DFC estimation reliability under challenging degradations, we evaluate the pretrained five-degradation model on unseen degradations (UIEB and CSD) and real-world degradations (NH-HAZE and LOL-v2-real) without adaptation. The scale-wise results are reported below.

Tab. R3-2. DFC Estimation Errors on Unseen and Real-World Degradations.

SettingDatasets2: NMAE↓ / W₁↓s1: NMAE↓ / W₁↓
Unseen deg.UIEB0.0551 / 0.01980.0494 / 0.0176
CSD0.0370 / 0.01260.0342 / 0.0112
Real-world deg.NH-HAZE0.0648 / 0.01960.0602 / 0.0179
LOL-v2-real0.0289 / 0.00980.0256 / 0.0081

Across all four datasets, DFC estimation errors remain relatively low and consistently decrease from scale 2 to scale 1 in both NMAE and W1W_1. These results support the reliability of DFC estimation under challenging degradations.

[W2] Isolating the Contribution of the DFC Ratio Formulation

We sincerely appreciate your insightful suggestion to more directly isolate the contribution of the DFC ratio formulation. Following your suggestion, we compare restoration performance using the degraded-image spectrum, residual spectrum, and DFC as guidance in FCMS, while keeping the architecture and training supervision identical. All variants are trained and evaluated under the three-degradation setting. The results are reported below.

Tab. R3-3. Comparison of Spectral Representations for Restoration Guidance (PSNR↑ / SSIM↑).

RepresentationSep. Ratio↑HazeRainNoise (σ=15\sigma=15)Noise (σ=25\sigma=25)Noise (σ=50\sigma=50)Avg.
Degraded Spectrum1.7331.86 / 0.97938.45 / 0.98234.11 / 0.93131.41 / 0.88828.12 / 0.80032.79 / 0.916
Residual Spectrum3.5032.07 / 0.98138.68 / 0.98434.14 / 0.93331.46 / 0.88928.20 / 0.80232.91 / 0.918
DFC (Ours)6.0732.28 / 0.98238.74 / 0.98534.19 / 0.93331.53 / 0.89028.28 / 0.80433.00 / 0.919

The results show that DFC consistently outperforms both degraded-image and residual spectra across all three degradation tasks, with average PSNR gains of 0.21 dB and 0.09 dB, respectively. Together with DFC's superior degradation separability, these results demonstrate that the DFC ratio formulation provides more discriminative degradation information and more effective restoration guidance than either spectrum alone.

AuthorsOpenReview ↗

Response to reviewer wxAx (W3, W4)

[W3] Performance under Comparable Parameter Budgets

We appreciate your thoughtful suggestion to assess DFC-IR's performance across restoration tasks under comparable parameter budgets. To this end, we construct DFC-IR (Light) by reducing the encoder blocks from [4,6,6,8][4, 6, 6, 8] to [2,2,2,4][2, 2, 2, 4], while keeping the decoder [2,4,4][2, 4, 4] and other settings unchanged. This reduces the parameters from 33.1M to 25.1M. The five-degradation results are reported below.

Tab. R3-4. Performance Comparison under Similar Parameter Budgets (PSNR↑).

MethodParams.HazeRainNoiseDeblurLow-lightAvg. PSNR↑
AdaIR28.8M30.5338.0231.3528.1223.0030.20
MoCE-IR25.4M30.4838.0431.3430.0523.0030.58
DFPIR31.1M31.6437.6231.2928.8223.8230.64
DFC-IR (Light)25.1M31.8637.7931.5029.5523.6430.87

The results show that DFC-IR (Light) achieves the highest average PSNR with the fewest parameters among the compared methods, outperforming MoCE-IR by 0.29 dB with a comparable parameter count. While the gains vary across tasks and are most pronounced in dehazing, these results demonstrate a favorable performance–parameter trade-off.

[W4] Robustness to DFC Design Choices and Adaptability to Complex Degradations

We sincerely thank you for highlighting this important point. To examine the sensitivity of DFC to its frequency design choices, we vary the number of measurement bands and the bandwidth schedule while keeping other settings unchanged. We evaluate both degradation separability and restoration performance under the three-degradation setting. The results are reported below.

Tab. R3-5. Ablation Study of DFC Design Choices.

Hyperparam.SettingSep. ratio↑Avg. PSNR↑
Band Count156.0132.97
25 (Ours)6.0733.00
356.0733.00
Bandwidth ScheduleConstant (0.175)5.9432.94
Linear ↑ (0.05→0.30, Ours)6.0733.00
Linear ↓ (0.30→0.05)5.7732.87

The results show that DFC maintains stable degradation separability and restoration guidance across different frequency configurations, with average PSNR varying by only 0.03 dB across band counts and 0.13 dB across bandwidth schedules. We have added these ablation experiments to Tab. 8 in our revision.

For spatially varying and directional degradations, DFC provides global radial frequency guidance, which is combined with image features through spectral token extraction in Eq. (8). These tokens are subsequently used by MMD for feature modulation, allowing restoration to leverage image-dependent spatial information. As shown in Tab. 15 of the original submission, DFC-IR achieves the best restoration performance on NH-HAZE (spatially varying real-world haze) and SPANet (directional real-world rain streaks), demonstrating its effectiveness under these challenging degradations.

We sincerely appreciate the careful thought and valuable insights behind your comments. Your suggestions encouraged us to more thoroughly examine DFC estimation reliability under challenging degradations, directly validate the contribution of the DFC ratio formulation, and further investigate performance under comparable parameter budgets and sensitivity to frequency design choices. These additional analyses have strengthened the experimental evidence supporting our approach and made our evaluation more comprehensive and rigorous. We are truly grateful for your constructive guidance in improving our work. We hope the new results and clarifications address your concerns and would sincerely appreciate it if you could reconsider your assessment.

AuthorsOpenReview ↗

Eagerly Await Your Response

Dear Reviewer wxAx,

Thank you again for your valuable feedback and thoughtful review. We have carefully addressed your concerns through detailed clarifications and additional experiments, and incorporated the corresponding revisions into the updated manuscript.

We understand that this is a busy period, and would greatly appreciate hearing whether our rebuttal addresses the points you raised or whether any concerns remain. We would be happy to provide further clarifications or discuss any open questions before the discussion period ends.

If our new evidence and clarifications satisfactorily address your concerns, we would sincerely appreciate it if you could reconsider your rating.

Thank you again for your time and consideration.

Best regards,

The Authors

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

The paper introduces the Degradation Frequency Curve (DFC), which measures the band-wise ratio of residual to degraded image energy using ring-shaped Gaussian masks. On synthetic haze, rain, and noise, the authors argue that DFC is robust to image content while sensitive to degradation type and severity. Based on this, DFC-IR is a multi-scale encoder-decoder in which DFCs are estimated from intermediate restorations (an initialized curve is used at the coarsest scale). FCMS uses each DFC to adaptively partition the spectrum and extract band-wise tokens from latent features, and MMD uses these tokens to modulate and aggregate decoder features. DFC-IR reports the best average PSNR on the three-degradation, five-degradation, and CDD11 benchmarks. The paper also tests unseen and real-world degradations and includes ablations of FCMS and MMD, the number of tokens, the degradation representation, and DFC estimation accuracy.

Critical strengths
  1. DFC characterizes degradation through band-wise energy ratios. Table 13 shows a higher separation ratio for DFC (6.07) than for the degraded spectrum (1.73) or the residual spectrum (3.50), suggesting that the ratio formulation captures degradation information beyond raw spectral features. This could make it useful beyond this specific architecture.
  2. DFC-IR achieves the highest average PSNR on the three-degradation, five-degradation, and CDD11 benchmarks, with a 0.90 dB gain over R2R on CDD11. The paper also evaluates unseen and real-world degradations, which is relevant given its focus on compound degradations.
Critical weaknesses
  1. In Tab. 6, FCMS alone adds only 0.21 dB, while MMD alone adds 0.78 dB at a much higher cost. The extra 0.24 dB of the full model over MMD alone does not isolate the effect of DFC guidance either. Tab. 8 compares DFC-guided and uniform tokens only on unseen degradations, not on the three-degradation, five-degradation, or CDD11 benchmarks. So it is unclear how much of the main results comes from DFC guidance, which is the paper's core contribution.
  2. Sec. 3.2 analyzes DFCs computed from ground-truth residuals, but at inference DFCs are estimated from intermediate restorations, and the coarsest scale starts from an initialized curve. Tab. 9 reports only average errors, with no per-scale results and nothing on unseen or real-world degradations. Moreover, the equal-area partition is invariant to scaling, and Eq. 6 takes only band parameters as input. So the DFC magnitude, which the severity analysis relies on, does not directly affect band selection. Overall, it is unclear how the properties analyzed in Sec. 3.2 carry over to the actual restoration.
Questions for authors
  1. Tab. 8 compares DFC-guided and uniform tokens only on unseen degradations. Could the authors provide the same comparison with an identical MMD architecture on the three-degradation, five-degradation, and CDD11 benchmarks? This would help isolate the contribution of DFC-guided partitioning to the main results.
  2. Tab. 9 reports low average DFC estimation errors but does not break them down by decoding scale. It also does not evaluate them on unseen or real-world degradations with references (e.g., UIEB, CSD, NH-HAZE, LOL-v2-real). How do NMAE and W1 vary across scales and these settings? Since the coarsest scale starts from an initialized curve, is the frequency partition at the first FCMS stage independent of the input?
  3. Equal-area partitioning is invariant to uniform scaling of DFC, and Eq. 6 takes only candidate band parameters as input. Where does the overall DFC magnitude enter FCMS or MMD, and how is the severity information analyzed in Sec. 3.2 used during restoration?
  4. Fig. 3 illustrates reusable band-wise priors across degradations, but Eq. 8 extracts tokens independently for each input. What mechanism associates or reuses tokens across degradation types? Is there evidence that unseen degradations benefit from shared band-wise patterns rather than simply from shared convolutional weights?
4 comments
AuthorsOpenReview ↗

Response to reviewer xnds (W1, W2, Q1, Q2, Q3)

Thank you for your thoughtful comments and questions, which have helped improve our work. We address your concerns as follows.

[W1, Q1] Evaluating the Contribution of DFC Guidance across Main Benchmarks

We first clarify the purposes of Tabs. 6 and 8 in the original submission: Tab. 6 presents component ablations, while Tab. 8 evaluates the contribution of DFC guidance by comparing different degradation representations under comparable model capacity and identical supervision.

We sincerely appreciate your thoughtful suggestion to more directly evaluate the contribution of DFC guidance across the main benchmarks. Accordingly, we extend the comparison between uniform tokens and DFC-guided tokens to the three-degradation, five-degradation, and CDD11 settings. Both variants use the same MMD architecture and training supervision, differing only in DFC guidance. The results are reported below.

Tab. R2-1. Effect of DFC Guidance on Main Restoration Benchmarks (Avg. PSNR↑ / SSIM↑).

Representation3-Degradation5-DegradationCDD11
Uniform Tokens32.70 / 0.91630.75 / 0.92029.93 / 0.889
DFC-guided Tokens33.00 / 0.91931.03 / 0.92230.22 / 0.890

The results show that DFC-guided tokens consistently outperform uniform tokens across all three benchmarks, with average PSNR gains of 0.28–0.30 dB, demonstrating the effectiveness of DFC guidance in improving restoration performance. Detailed per-task results are provided in Tabs. 23–25 (Appendix F.1) of the revision.

[W2, Q2] DFC Estimation Reliability across Scales and Degradations

Tab. 9 of the original submission reports DFC estimation errors averaged over scales 2 and 1. We present the scale-wise NMAE and W1W_1 results below and have added them to Tab. 9 in our revision.

Tab. R2-2. Scale-wise DFC Estimation Errors under Different Training Objectives.

VariantMSLf\mathcal{L}_{f}s2: NMAE↓ / W₁↓s1: NMAE↓ / W₁↓
(a)✗✗0.3977 / 0.11940.3147 / 0.0938
(b)✗✓0.2943 / 0.08840.2609 / 0.0786
(c)✓✗0.0594 / 0.02170.0492 / 0.0175
(d)✓✓0.0175 / 0.00510.0152 / 0.0043

The results show that multi-scale supervision and frequency loss reduce DFC estimation errors at both scales. Across all variants, NMAE and W1W_1 consistently decrease from scale 2 to scale 1.

We further evaluate DFC estimation reliability on unseen degradations (UIEB and CSD) and real-world degradations (NH-HAZE and LOL-v2-real) using the pretrained five-degradation model without adaptation. The scale-wise results are reported below.

Tab. R2-3. DFC Estimation Errors on Unseen and Real-World Degradations.

SettingDatasets2: NMAE↓ / W₁↓s1: NMAE↓ / W₁↓
Unseen deg.UIEB0.0551 / 0.01980.0494 / 0.0176
CSD0.0370 / 0.01260.0342 / 0.0112
Real-world deg.NH-HAZE0.0648 / 0.01960.0602 / 0.0179
LOL-v2-real0.0289 / 0.00980.0256 / 0.0081

The results show that DFC estimation errors remain relatively low on unseen and real-world degradations, with both NMAE and W1W_1 consistently decreasing from scale 2 to scale 1 across all four datasets.

At scale 3, uniform frequency partitioning is used without DFC estimation. Thus, the frequency partitioning at the first FCMS stage is input-independent. However, this input independence applies only to frequency partitioning, as the spectral tokens are still constructed from the current image features via Eq. (8).

[W2, Q3] How Degradation Severity Informs Restoration

We clarify that the overall DFC magnitude does not directly enter FCMS or MMD as a separate input. Instead, severity information is conveyed through two pathways:

  • Relative DFC Profile Changes. Different degradation severity levels can alter the DFC's relative frequency distribution rather than merely scaling its overall magnitude. These changes affect the Gaussian masks through Eqs. (6)–(7) and consequently the spectral tokens generated by Eq. (8), which then guide feature modulation in MMD through Eq. (9). As shown in Fig. 11 (Appendix B.3) of the revision, the normalized DFC profiles exhibit different shapes across severity levels of haze, rain, and noise.
  • Image Feature Conditioning. Eq. (8) also generates spectral tokens from the current image features zs{z}^s, allowing degradation severity information to be conveyed through these tokens and subsequently used for feature modulation in MMD.

Furthermore, DFC-IR achieves the best restoration performance on unseen degradation levels (Rain100H and noise σ=75\sigma=75), as shown in Tab. 14 (Appendix D.1) of the original submission. These results demonstrate its effectiveness in handling varying degradation severities.

AuthorsOpenReview ↗

Response to reviewer xnds (Q4)

[Q4] Mechanism and Evidence for Band-wise Prior Reuse

We apologize for the unclear explanation of prior reuse. It refers to reusing the band-specific restoration knowledge learned by the band-wise branches in MMD across degradations, rather than spectral tokens themselves. The tokens are generated independently for each input via Eq. (8) and guide the corresponding branches to apply this knowledge. This reuse is enabled by a fixed correspondence between spectral tokens and MMD branches: the ii-th token consistently modulates the ii-th band-wise branch via Eq. (9). This allows each branch to accumulate band-specific restoration knowledge across degradations during training. We have revised the corresponding discussion in Sec. 3.3.

To further examine whether such knowledge reuse benefits unseen degradations, we compare original routing with random routing, which randomly assigns tokens to branches during training. Both variants use identical architecture and supervision under the five-degradation setting. The results are reported below.

Tab. R2-4. Effect of Band-wise Routing on PSNR↑. Random Routing: mean ± std. over 10 random-routing trials.

VariantSeen: HazeSeen: RainSeen: NoiseUnseen: UIEUnseen: Desnowing
Original Routing32.2737.9131.5519.2421.05
Random Routing32.09 ± 0.0637.75 ± 0.0331.45 ± 0.0418.72 ± 0.1320.54 ± 0.11

The results show that original routing consistently outperforms random routing, particularly on unseen degradations. This supports the benefit of reusing band-specific restoration knowledge across degradations, beyond shared convolutional weights alone.

We sincerely appreciate the careful thought behind your comments. Your questions prompted us to more systematically validate the independent contribution of DFC guidance across the main benchmarks, investigate DFC estimation reliability under unseen and real-world degradations, and clarify how degradation severity and band-wise prior reuse contribute to restoration. These additions have strengthened the experimental foundation of our work and made the underlying mechanisms clearer. We are truly grateful for the valuable perspective your review has brought to our research. We hope the new evidence and clarifications address your concerns and would greatly appreciate it if you could reconsider your assessment.

AuthorsOpenReview ↗

Eagerly Await Your Response

Dear Reviewer xnds,

Thank you again for your valuable feedback and thoughtful review. We have carefully addressed your concerns through detailed clarifications and additional experiments, and incorporated the corresponding revisions into the updated manuscript.

We understand that this is a busy period, and would greatly appreciate hearing whether our rebuttal addresses the points you raised or whether any concerns remain. We would be happy to provide further clarifications or discuss any open questions before the discussion period ends.

If our new evidence and clarifications satisfactorily address your concerns, we would sincerely appreciate it if you could reconsider your rating.

Thank you again for your time and consideration.

Best regards,

The Authors

AuthorsOpenReview ↗

Waiting for Your Response

Dear Reviewer xnds,

Since the rebuttal period is coming to an end, we would sincerely appreciate your feedback on whether our responses have adequately addressed your concerns. If so, we kindly hope you could reconsider your rating. If any other concerns remain, we would like to have the chance to address them before the deadline.

Thank you again for your support and constructive feedback.

Best regards,

The Authors

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

The paper addresses the difficulty of explicitly quantifying degradation information in all-in-one image restoration. It introduces the Degradation Frequency Curve (DFC), a structured representation based on the residual-to-degraded image energy ratio within radial frequency bands, to compare degradation patterns and severity. The model estimates the curve at successive scales from intermediate restoration outputs, adaptively partitions frequency bands, extracts spectral tokens, and performs coarse-to-fine restoration through feature modulation and fusion. Training combines multi-scale reconstruction supervision with a frequency-domain consistency loss and requires no degradation-type labels. Experiments cover three-degradation, five-degradation, composite, unseen-type, and real-world mixed-degradation settings, with ablations of band partitioning and curve estimation. The results support its restoration and generalization value, while attribution of the mechanism's independent contribution and reuse of frequency-band priors require clarification.

Critical strengths
  1. DFC explicitly connects degradation representation to restoration: Equations (1)–(3) define the band-wise energy ratio, and Section 4.1 estimates the curve from intermediate restoration outputs to guide adaptive band partitioning. In Table 8, it improves underwater enhancement and desnowing by 0.48/0.54 dB over uniform partitioning, providing direct evidence that input-dependent spectral guidance improves restoration under unseen degradations.
  2. Tables 1–5 cover standard, composite, and unseen degradations, as well as real-world mixed-degradation images. In Table 3, average PSNR on CDD11 improves by 0.90 dB over R2R; in Table 4, underwater enhancement and desnowing improve by 0.59/0.45 dB over MoCE-IR, both without retraining. These results across settings support the method's applicability to composite and unseen degradations.
Critical weaknesses
  1. The ablations do not fully isolate DFC's independent contribution. The gains in Table 6 accompany an increase from 26.5M to 33.1M parameters, while Table 9 also changes the supervision applied directly to restored images. The uniform-partitioning comparison in Table 8 provides supporting evidence, but the capacity of the substituted modules is unclear. Curve guidance should be evaluated with matched capacity and identical supervision to rule out confounding effects from increased network capacity or stronger supervision.
  2. AdaIR already uses adaptive frequency partitioning and band-wise modulation. The identifiable advances here are the residual energy ratio, feedback across scales, and equal-area partitioning into multiple frequency bands. However, Equations (6)–(8) use the curve to adjust Gaussian masks and filter current features, without explaining how reusable degradation units emerge. Thus, the representation and partitioning innovations do not directly substantiate the stronger claims of prior reuse in Figure 3 and Section 3.3.
Questions for authors
  1. Which modules do the three representations in Table 8 replace, and do they have identical trainable parameter counts? With supervision held fixed, how do restoration results differ between reference DFCs, estimated DFCs, and DFCs shuffled across images?
  2. Does 'prior reuse' in Figure 3 refer to shared network weights or frequency-band units with correspondence across samples? As the adaptive boundaries change, what mechanism in Equations (6)–(8) enables the latter?
  3. Over which datasets and scales is the DFC error in Table 9 computed? In the underwater enhancement and desnowing settings of Table 4, is the relationship between curve estimation error and final restoration gains still consistent?
  4. The equal-area partitioning in Section 4.1 is invariant to multiplying the entire DFC by a positive constant. How does absolute degradation severity enter the spectral tokens: through direct encoding of the curve's magnitude or indirectly through current image features?
5 comments
AuthorsOpenReview ↗

Response to reviewer LYGP (W1, Q1)

Thank you for your insightful feedback and constructive suggestions. We address your concerns and questions in detail below.

[W1, Q1] Isolating the Independent Contribution of DFC Guidance

We first clarify that Tab. 6 of the original submission presents component ablations, while Tab. 9 evaluates DFC estimation reliability under different training objectives. In Tab. 8 of the original submission, we compare three degradation representations under comparable model capacities and identical training supervision to isolate the contribution of DFC guidance. Specifically, the three variants differ only in the degradation representation used for token construction in FCMS:

  • Learned Prompt: PromptIR-style prompt generation.
  • Uniform Tokens: Replacing the DFC with a uniform curve.
  • DFC-guided Tokens: Original FCMS (unchanged).

Architectural details are provided in Fig. 10 (Appendix A.3) of our revision, and the parameter counts are summarized below.

Tab. R1-1. Trainable parameter comparison.

VariantParams (M)
Learned Prompt33.4
Uniform Tokens33.1
DFC-guided Tokens33.1

To further isolate the contribution of DFC guidance, we compare four DFC conditions under the three-degradation setting. We use the same pretrained model and vary only the DFCs during inference:

  • Ref.: Reference DFC.
  • Est.: Estimated DFC.
  • Same-Shuf.: DFC from another image with the same degradation type.
  • Cross-Shuf.: DFC from another image with a different degradation type.

The DFC errors at each decoding scale and the final restoration PSNR are reported below.

Tab. R1-2. Effect of Different DFC Guidance on Restoration Performance.

DegradationDFC Guidances2: NMAE↓ / W₁↓s1: NMAE↓ / W₁↓PSNR↑
HazeRef.0 / 00 / 032.30
Est.0.0172 / 0.00500.0164 / 0.004732.28
Same-Shuf.0.0397 / 0.01260.0377 / 0.011832.21
Cross-Shuf.0.3732 / 0.11850.3442 / 0.104631.04
RainRef.0 / 00 / 038.74
Est.0.0117 / 0.00310.0095 / 0.002438.74
Same-Shuf.0.0464/ 0.01780.0577 / 0.020438.68
Cross-Shuf.0.2730 / 0.08020.3006 / 0.087737.81
Noise (σ=25\sigma=25)Ref.0 / 00 / 031.54
Est.0.0237 / 0.00690.0196 / 0.005831.53
Same-Shuf.0.0323 / 0.00910.0301 / 0.008731.50
Cross-Shuf.0.2280 / 0.05680.2219/ 0.069530.88

The results show that: (1) Est. achieves nearly identical PSNR to Ref. (within 0.02 dB); (2) Shuffled DFCs reduce PSNR, with drops of 0.03–0.07 dB under Same-Shuf. and more pronounced drops of 0.65–1.24 dB under Cross-Shuf. Furthermore, across all three degradation types, restoration performance consistently decreases as DFC errors increase. These results support the effectiveness of reliable DFC guidance for image restoration.

To present these results more clearly and systematically, we have divided the original Sec. 5.3 into two sections in the revised manuscript: Sec. 5.3 (Ablation Studies) and Sec. 5.4 (Effect of DFC Guidance). The latter brings together the degradation representation comparison from Tab. 8 of the original submission and the newly added DFC guidance experiment.

AuthorsOpenReview ↗

Response to reviewer LYGP (W2, Q2, Q3)

[W2, Q2] Clarifying Band-wise Prior Reuse

We apologize for the unclear explanation of prior reuse. It refers to reusing the band-specific restoration knowledge learned by the band-wise branches in MMD across degradations. The DFC-guided tokens of each input serve as band-wise conditions, guiding the corresponding branches to apply this knowledge. Although adaptive frequency boundaries vary across images, the ii-th token generated by Eq. (8) consistently modulates the ii-th band-wise branch via Eq. (9). This fixed correspondence enables each branch to accumulate band-specific restoration knowledge across degradations during training. We have revised the corresponding discussion in Sec. 3.3.

To further validate this mechanism, we compare original routing with random routing, which randomly assigns tokens to branches during training. Both variants use identical architecture and supervision under the five-degradation setting. The results are reported below.

Tab. R1-3. Effect of Band-wise Routing on PSNR↑. Random Routing: mean ± std. over 10 random routing trials.

VariantSeen: HazeSeen: RainSeen: NoiseUnseen: UIEUnseen: Desnowing
Original Routing32.2737.9131.5519.2421.05
Random Routing32.09 ± 0.0637.75 ± 0.0331.45 ± 0.0418.72 ± 0.1320.54 ± 0.11

The results show that original routing consistently outperforms random routing, particularly on unseen degradations, providing empirical support for this mechanism.

[Q3] DFC Error–Performance Relationship in Unseen Degradations

Regarding Tab. 9 of the original submission, the DFC estimation errors are averaged over scales 2 and 1 on Rain100L (deraining), SOTS (dehazing), and BSD68 (denoising, σ=15,25,50\sigma=15,25,50). At scale 3, uniform frequency partitioning is used without DFC estimation. We have added scale-wise DFC errors to Tab. 9 in our revision.

We further examine the relationship between DFC estimation errors and restoration performance on unseen degradations (UIEB and CSD). Using the pretrained five-degradation model, we construct DFC guidance with varying errors through interpolation at each scale:

Dλ=(1−λ)Dest+λDref,λ∈[0,1].D_{\lambda}=(1-\lambda)D_{\mathrm{est}}+\lambda D_{\mathrm{ref}},\quad \lambda\in[0,1].

The scale-wise DFC errors and restoration PSNR are reported below.

Tab. R1-4. DFC Errors and Restoration Performance on Unseen Degradations.

DatasetDFC Conditions2: NMAE↓ / W₁↓s1: NMAE↓ / W₁↓PSNR↑
UIEBEst. (λ=0)0.0551 / 0.01980.0494 / 0.017619.24
Interp. (λ=0.3)0.0386 / 0.01390.0346 / 0.012319.29
Interp. (λ=0.6)0.0220 / 0.00790.0198 / 0.007019.32
Ref. (λ=1)0 / 00 / 019.39
CSDEst. (λ=0)0.0370 / 0.01260.0342 / 0.011221.05
Interp. (λ=0.3)0.0259 / 0.00880.0239 / 0.007821.09
Interp. (λ=0.6)0.0148 / 0.00500.0137 / 0.004521.16
Ref. (λ=1)0 / 00 / 021.23

The results show that as DFC errors decrease, restoration PSNR consistently improves on both UIEB and CSD, confirming that this relationship also holds for unseen degradations.

AuthorsOpenReview ↗

Response to reviewer LYGP (Q4)

[Q4] How Degradation Severity Enters Spectral Tokens

We clarify that absolute degradation severity is not directly encoded into spectral tokens through the DFC magnitude. Instead, severity information is conveyed through two pathways:

  • Relative DFC Profile Changes. Different degradation severity levels can alter the relative frequency distribution of DFC, rather than merely scaling its overall magnitude. These variations affect the Gaussian masks through Eqs. (6)–(7) and consequently the spectral tokens generated by Eq. (8). As shown in Fig. 11 (Appendix B.3) of the revision, the normalized DFC profiles exhibit different shapes across severity levels of haze, rain, and noise.
  • Image Feature Conditioning. Eq. (8) also generates spectral tokens from the current image features zsz^s, allowing degradation severity information to be conveyed through these features.

Furthermore, DFC-IR achieves the best restoration performance on unseen degradation levels (Rain100H and noise σ=75\sigma=75), as shown in Tab. 14 (Appendix D.1) of the original submission. These results demonstrate its effectiveness in handling varying degradation severities.

We sincerely appreciate the careful thought behind your comments. Your questions encouraged us to look beyond overall restoration performance and more directly examine the contribution of DFC guidance, the impact of DFC estimation errors, and the nature of band-wise prior reuse. This process has helped us provide more targeted experimental evidence and a clearer explanation of the underlying mechanisms. We believe these additions have made our paper more rigorous and convincing, and we are grateful for your role in guiding these improvements. We hope our new results and clarifications address your concerns and would sincerely appreciate it if you could reconsider your assessment.

Comments on the paper

1 comment
AuthorsOpenReview ↗

A Summary of Our Interactions with Reviewers

We sincerely thank all Reviewers, ACs, and SACs for their thoughtful and constructive feedback. We are grateful for the reviewers' recognition of DFC as an explicit and interpretable frequency-quantified degradation representation, the strong restoration performance of DFC-IR across standard and composite degradations, and its promising generalization to unseen and real-world degradations. Below is a concise summary of our interactions.


Reviewer LYGP

Main concerns: (1) independent contribution of DFC guidance, (2) mechanism of band-wise prior reuse, (3) DFC estimation accuracy and its relationship with restoration performance, and (4) how degradation severity enters spectral tokens.

Our response:

(1) We clarified the experimental settings for different degradation representations and compared reference, estimated, and shuffled DFCs to evaluate the contribution of DFC guidance (Sec. 5.4; Appendix A.3).

(2) We clarified that band-wise prior reuse refers to reusing restoration knowledge learned across degradations by MMD's band-wise branches. We explained how fixed token-to-branch correspondence enables this reuse and validated this mechanism through random-routing experiments (Sec. 3.3; Appendix F.2).

(3) We provided scale-wise DFC estimation errors and examined their relationship with restoration performance on unseen degradations (Tab. R1-4).

(4) We clarified how degradation severity is conveyed through relative DFC profiles and image features (Appendix B.3).


Reviewer xnds

Main concerns: (1) contribution of DFC guidance across the main benchmarks, (2) reliability of DFC estimation across scales and challenging degradations, (3) incorporation of degradation severity information, and (4) evidence for band-wise prior reuse.

Our response:

(1) We extended the controlled comparison between DFC-guided Tokens and Uniform Tokens to the main benchmarks (Appendix F.1).

(2) We provided scale-wise DFC estimation errors and extended the evaluation to unseen and real-world degradations (Table 9; Appendix D.3).

(3) We clarified how degradation severity is conveyed through relative DFC profiles and image features (Appendix B.3).

(4) We clarified the mechanism of band-wise prior reuse and validated it through random-routing experiments (Sec. 3.3; Appendix F.2).


Reviewer wxAx

Main concerns: (1) reliability of estimated DFCs under challenging conditions, (2) independent contribution of the DFC ratio formulation, (3) task-wise performance under comparable parameter budgets, and (4) sensitivity to frequency design choices and effectiveness under spatially varying or directional degradations.

Our response:

(1) We provided scale-wise DFC estimation errors and extended the evaluation to unseen and real-world degradations (Table 9; Appendix D.3).

(2) We compared DFC with degraded-image and residual spectra as restoration guidance under identical architecture and supervision (Appendix B.4).

(3) We introduced DFC-IR (Light) for comparison under similar parameter budgets and further clarified the task-wise performance differences (Appendix E).

(4) We evaluated the sensitivity of DFC to measurement band counts and bandwidth schedules and discussed its effectiveness on spatially varying and directional degradations (Table 8; Appendix D.2).

Reviewer's follow-up: The reviewer raised concerns about task preference due to the different task-wise advantages of DFC-IR (Light) and MoCE-IR. We acknowledged these differences and distinguished the comparison between complete models from the contribution of DFC guidance. Table 24 (Appendix F.1) shows that DFC guidance improves all five tasks in a controlled ablation. The reviewer subsequently confirmed that most of their concerns had been addressed and stated: "I will update my score to 'Weak Acceptance'."


Reviewer obes

Main concerns: (1) impact of early-stage DFC estimation errors on subsequent restoration, (2) generality of DFC properties across additional degradation types, and (3) robustness to imperfect ground truth.

Our response:

(1) We conducted controlled experiments on UIEB and CSD to examine how early-stage DFC estimation errors affect subsequent estimation and restoration performance (Appendix D.4).

(2) We extended DFC property analysis to seven degradation types and quantitatively evaluated its content robustness and degradation sensitivity (Appendix B.5).

(3) We clarified that DFC guidance is estimated without directly using ground truth and discussed the impact of imperfect ground truth on training and zero-shot evaluation.

Reviewer's follow-up: The reviewer explicitly confirmed that our response had addressed their concerns: "The authors' response has addressed my concerns."


All updates are highlighted in blue in the revised manuscript for easy reference.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.