acceptodds
Under review as a conference paper at ICLR 2027

DeMemSR: Transferable Degradation Representation Learning and Stable–Dynamic LR Memory Fusion for Diffusion-Based Real-World Image Super-Resolution

Abstract

Diffusion-based real-world image super-resolution can benefit from both degradation-aware conditioning and adaptive interaction between low-resolution (LR) observations and evolving noisy latents. However, contrastive degradation learning may distinguish different degradation configurations using only a subset of degradation cues, without adequately retaining the degradation information relevant to restoration. Moreover, although bidirectional LR–latent interaction allows LR guidance to adapt to the evolving denoising state, repeatedly updating the LR stream may cause it to drift from the original observation, weakening its role as a faithful structural anchor. To address these issues, we present DeMemSR, a degradation-aware diffusion Transformer framework that combines Transferable Degradation Representation Learning (TDRL) with Stable–Dynamic LR Memory Fusion (SD-LRMF). TDRL combines contrastive degradation learning with cross-content re-degradation. While contrastive learning encourages the projected features to distinguish degradation configurations, re-degradation encourages the encoder to retain degradation information beyond what is needed for discrimination. A degradation-aware conditioner then uses the projected features to provide layer-specific guidance to the diffusion Transformer. SD-LRMF explicitly separates LR guidance into an immutable Stable LR Memory derived from the initial observation and an interaction-updated Dynamic LR State. Learnable layer-wise fusion re-anchors the dynamic state to the stable memory across network depth, preserving observation fidelity while retaining the flexibility of bidirectional interaction. Experiments on four real-world benchmarks show that DeMemSR achieves leading no-reference perceptual quality, including the highest MUSIQ and LIQE scores among the evaluated methods on all four datasets.

Primary area:
generative models
Keywords:
Real-world image super-resolution, Diffusion Transformer, Transferable degradation representation learning, Degradation-aware conditioning, Stable–dynamic LR memory fusion

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_Kvw4OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The author proposes DeMemSR, a diffusion Transformer framework of degraded perception, and the experiments on four real-world benchmark data sets show that DeMemSR has achieved leading performance in the quality of non-reference perception.

Critical strengths
  1. The author makes a comparative experiment on four data sets, and the experiment is sufficient and convincing.

  2. The author's overall writing is good, and the flow chart looks clear.

  3. The visualization effect of the author in Figure 2 is better, and the details are more sufficient, and the visualization results are more pleasing to the eye.

  4. The author also compares a large number of the latest comparison methods, which achieve good performance.

Critical weaknesses
  1. In the initial submission of PSNR and SSIM, the author gave the corresponding results in my response.

  2. I think the related work on super-resolution should provide more visual results. I have only seen the results in Figure 2, and I feel that the visualization of the three images is not sufficient.

  3. Why are many indicators in Table 6 and Table 5-? Why is this? Why are these indicators not provided by the author? I hope the author will give an explanation.

4.It seems that SeeSR's performance is superior on RealSR dataset.

Questions for authors

I found that PSNR and SSIM were not used in Table 1, but they were tested in many ablation experiments. I was puzzled why the main evaluation indexes of the experiments were different. Of course, I saw that the reply given by the author supplemented the indexes.

1 comment

W1 & Q1: Full-Reference Metrics

Response:

Diffusion-based super-resolution methods generally place greater emphasis on perceptual quality. Due to space limitations, we focused on no-reference perceptual quality metrics in the original manuscript.

Following the reviewer's suggestion, we have added full-reference metrics for DeMemSR and its baseline model, DiT4SR, on DRealSR and RealSR in Table 2 (page 8) of Section 4.2 of the revised manuscript.

DatasetMethodPSNR ↑SSIM ↑LPIPS ↓MUSIQ ↑Q-Align ↑
DRealSRDiT4SR25.33910.67430.382064.66493.4304
DeMemSR26.16370.68930.388767.25613.6392
RealSRDiT4SR22.98660.65110.342267.87133.5197
DeMemSR23.23960.64310.350270.27973.7105

Compared with our baseline model, DiT4SR, DeMemSR also achieves improvements in full-reference metrics.

W2: Comparisons with Simpler Alternatives for SD-LRMF

Response:

Following the reviewer's suggestion, we have added two baseline configurations:

  1. Fixed-weight Fusion: The Stable LR Memory and Dynamic LR State are fused using fixed weights of 0.5/0.5.
  2. Standard Residual Connection: The Stable LR Memory is fused with the Dynamic LR State through an additive residual connection.

The corresponding experimental results have been added to Table 5 (page 10) of Section 4.4 in the revised manuscript.

DatasetMethodPSNR ↑SSIM ↑LPIPS ↓MUSIQ ↑Q-Align ↑
DRealSRFixed25.27560.63700.430067.78803.6136
Residual25.89110.66770.401567.57983.5515
SD-LRMF (Ours)26.16370.68930.388767.25613.6392
RealLR200Fixed———71.57764.3282
Residual———71.06644.3596
SD-LRMF (Ours)———71.67384.3857

SD-LRMF outperforms both simpler alternatives in nearly all metrics.

W3: More Fine-Grained Ablation Studies of Key Components and Design Choices

Response:

We have conducted additional ablation studies on SD-LRMF fusion strategies and auxiliary re-degradation supervision in TDRL.

(1) Ablation Study on Auxiliary Re-degradation Supervision in TDRL

To further evaluate the contribution of re-degradation supervision to downstream super-resolution performance, we compare two training strategies:

  • Contrastive Only: Uses cross-content contrastive learning alone.
  • Contrastive + Re-Deg: Adds auxiliary re-degradation supervision to cross-content contrastive learning.

Both strategies are evaluated using the same DAC conditioning mechanism to assess their effects on final restoration performance.

The corresponding experimental results have been added to Table 6 (page 10) of Section 4.4 in the revised manuscript.

DatasetTraining ObjectivePSNR ↑SSIM ↑LPIPS ↓MANIQA ↑Q-Align ↑LIQE ↑
DRealSRContrastive Only26.14470.69040.38970.48763.62514.2811
Contrastive + Re-Deg26.16370.68930.38870.48963.63924.2813
RealLR200Contrastive Only———0.48294.37284.4319
Contrastive + Re-Deg———0.49454.38574.4749

These results demonstrate that auxiliary re-degradation supervision improves downstream super-resolution performance across nearly all reported metrics, supporting its effectiveness in TDRL.

(2) Fine-Grained Ablation Study on SD-LRMF Fusion Strategies

For SD-LRMF, we have added two simpler alternatives, fixed-weight fusion and standard residual connections, and compared them with the complete learnable layer-wise fusion strategy.

The corresponding experimental results have been discussed in detail in our response to W2.

Reviewer_Dpv4OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents DeMemSR, a diffusion Transformer framework for real-world image super-resolution built on DiT4SR. Its two main components are Transferable Degradation Representation Learning (TDRL), which combines cross-content contrastive learning with auxiliary re-degradation supervision, and Stable–Dynamic LR Memory Fusion (SD-LRMF), which blends evolving LR features with an immutable initial LR representation through learnable layer-wise weights. The learned degradation representations provide layer-specific conditioning, while memory fusion aims to maintain reliable structural guidance across network depth. Experiments on four real-world benchmarks report competitive no-reference perceptual quality, including the highest MUSIQ and LIQE scores among the evaluated methods on all four datasets.

Critical strengths
  1. Degradation representations and stable–dynamic LR fusion provide complementary restoration guidance, while SD-LRMF introduces only 23 learnable scalar parameters.
  2. The method achieves the highest MUSIQ and LIQE scores among the evaluated methods on all four real-world benchmarks, with component ablations also showing performance gains.
Critical weaknesses
  1. The introduction does not clearly explain the problems with existing methods and lacks illustrative figures and concrete evidence to make these problems apparent.
  2. The experimental evaluation is limited, and the ablation studies are insufficient, particularly for TDRL. Although the ablations demonstrate performance gains from SD-LRMF, its proposed mechanism of mitigating feature drift and preserving observation fidelity remains insufficiently validated.
  3. The evaluation relies exclusively on no-reference metrics, which can assign high scores even to images with substantial artifacts. Reference-based metrics should also be included to assess restoration fidelity.
Questions for authors
  1. Can you provide concrete evidence of the degradation information missed by contrastive learning and demonstrate when LR feature drift harms restoration?
  2. How does TDRL compare with contrastive-only training in downstream SR performance under matched architectures and training budgets?
4 comments

W1 & Q1 & W2 & Q2: Limitations of Contrastive Degradation Learning And Ablation Study on SD-LRMF

Response:

Traditional contrastive degradation learning mainly encourages the encoder to learn discriminative degradation representations by pulling together features with the same degradation configuration and pushing apart features with different degradation configurations.

However, degradation discriminability does not necessarily imply sufficient preservation of degradation information. When certain degradation features are already sufficient to distinguish training samples, the contrastive learning objective does not require the encoder to further retain other degradation information that may be useful for image restoration.

Clearly, contrastive learning-based degradation representations are not theoretically optimized for strong representational capacity. To further demonstrate that our re-degradation supervision produces more effective degradation representations, we compare the super-resolution results guided by representations learned with and without re-degradation supervision. As shown in the table, incorporating re-degradation supervision improves our model's performance across nearly all evaluation metrics.

DatasetTraining ObjectivePSNR ↑SSIM ↑LPIPS ↓MANIQA ↑Q-Align ↑LIQE ↑
DRealSRContrastive Only26.14470.69040.38970.48763.62514.2811
Contrastive + Re-Deg26.16370.68930.38870.48963.63924.2813
RealLR200Contrastive Only———0.48294.37284.4319
Contrastive + Re-Deg———0.49454.38574.4749

W1 & W2 & Q1: LR Feature Drift

Response:

To investigate LR feature drift caused by the noise stream, we compare the relative changes in LR features under different noise levels. Specifically, we evaluate the effects of noise levels scaled by 0.5× and 1.5×.

DatasetBlockDiT4SR 0.5×DeMemSR 0.5×DiT4SR 1.5×DeMemSR 1.5×
DRealSR38.94%4.00%9.89%8.46%
99.45%6.21%17.09%15.49%
1513.57%9.38%24.68%21.33%
2115.94%11.67%28.00%23.67%
RealLR20039.28%3.90%10.34%7.36%
98.83%5.13%15.07%12.07%
1511.05%7.30%18.68%16.15%
2113.04%10.81%22.01%20.66%

We report the relative changes in LR features at Transformer blocks 3, 9, 25, and 21. As shown in the table, DeMemSR exhibits lower relative LR feature changes than DiT4SR at all four evaluated blocks and under both scaling settings on DRealSR and RealLR200. These results support the role of SD-LRMF in maintaining more stable LR representations during repeated feature interactions.

W2: Ablation Study on SD-LRMF

Response:

We have added two baseline configurations:

  1. Fixed-weight Fusion: The Stable LR and Dynamic LR are fused using fixed weights of 0.5/0.5.
  2. Standard Residual Connection: The Stable LR are fused with the Dynamic LR through an additive residual connection.

The corresponding experimental results are reported in Table 5 (page 10) of Section 4.4 in the revised manuscript.

DatasetMethodPSNR ↑SSIM ↑LPIPS ↓MUSIQ ↑Q-Align ↑
DRealSRFixed25.27560.63700.430067.78803.6136
Residual25.89110.66770.401567.57983.5515
SD-LRMF26.16370.68930.388767.25613.6392
RealLR200Fixed———71.57764.3282
Residual———71.06644.3596
SD-LRMF———71.67384.3857

SD-LRMF outperforms both simpler alternatives in nearly all metrics.

W3: Full-Reference Metrics

Response:

Diffusion-based super-resolution methods generally place greater emphasis on perceptual quality. Due to space limitations, we focused on no-reference perceptual quality metrics in the original manuscript.

We have added full-reference metrics for DeMemSR and its baseline model, DiT4SR, on DRealSR and RealSR.

DatasetMethodPSNR ↑SSIM ↑LPIPS ↓MUSIQ ↑Q-Align ↑
DRealSRDiT4SR25.33910.67430.382064.66493.4304
DeMemSR26.16370.68930.388767.25613.6392
RealSRDiT4SR22.98660.65110.342267.87133.5197
DeMemSR23.23960.64310.350270.27973.7105

Compared with our baseline model, DiT4SR, DeMemSR also achieves improvements in full-reference metrics.

Thank you for the detailed rebuttal and the additional experiments, which I appreciate. However, the motivation in the introduction still lacks concrete evidence (W1), and the new ablations don't yet clearly show TDRL's benefit or confirm how SD-LRMF works (W2). The full-reference results are also mixed relative to DiT4SR (W3). I will therefore keep my original score.

Thank you for the reviewer's further feedback.

Regarding W1, the evidence supporting the motivation for degradation representation learning.

First, we emphasize that, theoretically, contrastive learning alone primarily encourages the learning of discriminative degradation representations, rather than representations capable of capturing detailed degradation characteristics. In other words, degradation representations learned solely through contrastive learning can distinguish which types of degradation affect an image, but may fail to adequately represent the detailed characteristics of those degradations. Such discriminative representations mainly provide category-level semantic information.

Second, prior work such as DRAT notes that contrastive degradation features may contain content-related redundancy. While DRAT aims to remove such redundancy, this assumes sufficient degradation details are already encoded. In contrast, our cross-content re-degradation supervision encourages richer and more accurate degradation representations.

Third, Table 11 in Appendix D reports the similarity s−s_{-} between representations extracted from the same image content under different degradation configurations. For an effective degradation representation, the similarity s−s_{-} between different degradations should ideally be close to zero. However, as shown in Table 11, the similarities s−s_{-} of representations learned solely through contrastive learning are 0.9601 and 0.9627 on the two datasets, respectively. These high similarities suggest that the representations may be influenced by image content and other factors, preventing them from effectively characterizing degradation information. In contrast, our re-degradation-supervised representations achieve s−s_{-} values closer to zero, indicating more effective degradation representation learning.

Fourth, as demonstrated by the ablation experiments provided in our previous response, degradation representations learned with re-degradation supervision lead to better restoration performance than those learned using contrastive learning alone.

Regarding the motivation for LR feature drift in W1 and how SD-LRMF works in W2.

The original DiT4SR paper points out, in both its Introduction and experimental analysis, that as the number of Transformer interaction layers increases, LR features may experience undesirable disturbances during continuous interactions with the Noise Stream, causing the effectiveness of LR guidance to gradually weaken. The authors illustrate this phenomenon in Figure 4 and introduce an LR Residual within the MM-DiT Block to alleviate the disturbances caused by interactions with the Noise Stream.

Based on this observation, while retaining the original LR Residual structure of DiT4SR, we further introduce the initial LR features as Stable LR Memory and fuse them with the dynamically updated LR features at each layer. This allows the initial LR information to continuously provide stable spatial guidance throughout the network depth, thereby further mitigating LR feature drift caused by repeated feature interactions.

Furthermore, the ablation results in Table 5 of the revised manuscript show that SD-LRMF achieves better performance across multiple image restoration metrics than Fixed-weight Fusion and Residual Connection, further validating the effectiveness of the proposed layer-wise adaptive fusion mechanism.

Regarding the effectiveness of TDRL in W2, Tables 4 and 6 of the revised manuscript provide targeted ablation evidence from two different perspectives. First, the ablation study in Table 4 shows that incorporating the learned degradation representations into the restoration process improves final restoration performance, supporting the effectiveness of degradation information in guiding super-resolution restoration. Second, Table 6 further compares two training strategies: Contrastive Only and Contrastive + Re-Deg. Under the same downstream DAC conditioning mechanism, introducing re-degradation supervision achieves improvements on the majority of evaluation metrics across different datasets.

Regarding the full-reference evaluation metrics in W3, we would like to further emphasize that, compared with DiT4SR, our method improves PSNR on both DRealSR and RealSR, with a particularly notable improvement on DRealSR. SSIM is also improved on DRealSR.

Furthermore, we believe that the comprehensive improvements in no-reference perceptual quality metrics should not be overlooked. DeMemSR outperforms DiT4SR on all five no-reference evaluation metrics across four real-world benchmarks. These results demonstrate that our method does not simply trade reconstruction fidelity for perceptual quality. Instead, it achieves more consistent improvements in perceptual quality while maintaining competitive reconstruction fidelity and further improving selected full-reference metrics.

Dear Reviewer,

Thank you again for your valuable feedback. We have carefully considered your latest comments and provided further explanations and experimental evidence addressing each of the three concerns (W1–W3) in our latest response.

In particular, we have further clarified the motivation for degradation representation learning with theoretical explanations and quantitative evidence, provided additional explanations and ablation results to validate the effectiveness of TDRL and SD-LRMF, and clarified the performance comparison with DiT4SR in terms of both full-reference and no-reference metrics.

We believe that our latest response has substantially addressed the concerns you raised. As the rebuttal period is approaching its deadline, we would sincerely appreciate it if you could review our latest response when possible and reconsider your assessment. If any concerns remain, we would be more than happy to provide further clarification.

Thank you very much for your time and consideration.

Reviewer_K69eOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes DeMemSR, a diffusion Transformer framework for real world image super resolution. It introduces a transferable degradation representation learning module that combines contrastive learning with cross content redegradation. A stable and dynamic memory fusion mechanism is also used to preserve initial low resolution features throughout the network. Experiments on four real world benchmarks show competitive perceptual quality on no reference metrics.

Critical strengths

1.The paper addresses two practical issues in diffusion super resolution. It learns degradation representations that go beyond simple discrimination, and it maintains reliable low resolution guidance during feature interaction.

2.The overall framework is clear and easy to follow. Experiments show competitive results across multiple real world benchmarks, and the ablations confirm a clear gain from the SDLRMF module.

Critical weaknesses

1.While the stable LR memory aims to keep structural fidelity, the paper omits full reference metrics on DRealSR and RealSR. Since both datasets have HR references, missing these metrics makes it hard to tell if the method truly preserves image content.

2.While transferable degradation learning is a main claim, TDRL is only trained under one synthetic degradation pipeline. The experiments mainly show that the features can discriminate known degradations rather than generalize to unseen ones. The impact of redegradation supervision on final restoration quality is also unclear. Overall, the claimed transferability is not convincingly demonstrated.

Questions for authors

1.Since SDLRMF is designed to preserve structural fidelity from the LR input, please report full reference metrics on DRealSR and RealSR to verify how well the image content is actually preserved.

2.Is there any direct evidence showing that LR features really drift during repeated interactions? Also, how does SDLRMF compare against simple skip connections or residual links from the initial LR features?

3.The representation analysis only tests synthetic degradations from the same training pipeline. Do the learned features actually generalize to unseen degradation types or real world capture conditions?

4 comments

W1 & Q1: Full-Reference Metrics

Response:

Diffusion-based super-resolution methods generally place greater emphasis on perceptual quality. Due to space limitations, we focused on no-reference perceptual quality metrics in the original manuscript.

We have added full-reference metrics for DeMemSR and its baseline model, DiT4SR, on DRealSR and RealSR in Table 2 (page 8) of Section 4.2 of the revised manuscript.

DatasetMethodPSNR ↑SSIM ↑LPIPS ↓MUSIQ ↑Q-Align ↑
DRealSRDiT4SR25.33910.67430.382064.66493.4304
DeMemSR26.16370.68930.388767.25613.6392
RealSRDiT4SR22.98660.65110.342267.87133.5197
DeMemSR23.23960.64310.350270.27973.7105

As shown, compared with our baseline model, DiT4SR, our DeMemSR also achieves improvements in full-reference evaluation metrics.

W2 & Q3: Test on Unseen Degradations

Response:

We construct our training dataset using the widely adopted Real-ESRGAN synthetic degradation pipeline. However, we evaluate our method on four real-world benchmarks with unseen degradation conditions: DRealSR, RealSR, RealLR200, and RealLQ250.

As shown in Tables 1 and 2, As shown in Tables 1 and 2, DeMemSR achieves overall superior perceptual quality compared with the evaluated SOTA methods across all four real-world benchmarks. In particular, DeMemSR outperforms its baseline model, DiT4SR, on the majority of the reported evaluation metrics.

These results provide experimental evidence that our method can generalize beyond the synthetic training degradations and achieve effective restoration under diverse real-world degradation conditions.

W2: Impact of Re-degradation Supervision

Response:

To further evaluate the impact of re-degradation supervision on downstream super-resolution performance, we compare two training strategies:

  • Contrastive Only: Uses contrastive learning alone.
  • Contrastive + Re-Deg: Adds auxiliary re-degradation supervision to contrastive learning.

The corresponding experimental results have been added to Table 6 (page 10) of Section 4.4 in the revised manuscript.

DatasetTraining ObjectivePSNR ↑SSIM ↑LPIPS ↓MANIQA ↑Q-Align ↑LIQE ↑
DRealSRContrastive Only26.14470.69040.38970.48763.62514.2811
Contrastive + Re-Deg26.16370.68930.38870.48963.63924.2813
RealLR200Contrastive Only———0.48294.37284.4319
Contrastive + Re-Deg———0.49454.38574.4749

After adding auxiliary re-degradation supervision, nearly all metrics show improvements.

Q2: Evidence of LR Feature Drift

Response:

To investigate LR feature drift caused by the noise stream, we compare the relative changes in LR features under different noise levels. Specifically, we evaluate the effects of noise levels scaled by 0.5× and 1.5×.

DatasetBlockDiT4SR 0.5×DeMemSR 0.5×DiT4SR 1.5×DeMemSR 1.5×
DRealSR38.94%4.00%9.89%8.46%
99.45%6.21%17.09%15.49%
1513.57%9.38%24.68%21.33%
2115.94%11.67%28.00%23.67%
RealLR20039.28%3.90%10.34%7.36%
98.83%5.13%15.07%12.07%
1511.05%7.30%18.68%16.15%
2113.04%10.81%22.01%20.66%

We report the relative changes in LR features at Transformer blocks 3, 9, 25, and 21. As shown in the table, DeMemSR exhibits lower relative LR feature changes than DiT4SR at all four evaluated blocks and under both scaling settings on DRealSR and RealLR200. These results support the role of SD-LRMF in maintaining more stable LR representations during repeated feature interactions.

Q2: Comparisons with Simpler Alternatives for SD-LRMF

Response:

We have added two baseline configurations:

  1. Fixed-weight Fusion: The Stable LR and Dynamic LR are fused using fixed weights of 0.5/0.5.
  2. Standard Residual Connection: The Stable LR are fused with the Dynamic LR through an additive residual connection.

The corresponding experimental results are reported in Table 5 (page 10) of Section 4.4 in the revised manuscript.

DatasetMethodPSNR ↑SSIM ↑LPIPS ↓MUSIQ ↑Q-Align ↑
DRealSRFixed25.27560.63700.430067.78803.6136
Residual25.89110.66770.401567.57983.5515
SD-LRMF26.16370.68930.388767.25613.6392
RealLR200Fixed———71.57764.3282
Residual———71.06644.3596
SD-LRMF———71.67384.3857

SD-LRMF outperforms both simpler alternatives in nearly all metrics.

Thank you for the detailed rebuttal and additional experiments, which address several of my concerns; however, the transferability of the learned degradation representations remains insufficiently demonstrated, so I will maintain my original rating.

Thank you for the reviewer's further feedback. Regarding the transferability of degradation representations, the reviewer's comments mainly focus on whether the learned degradation representations can generalize to unseen degradation conditions and whether re-degradation supervision can effectively improve final image restoration quality. For these two concerns, we have supplemented relevant experiments and analyses in our previous response and the revised manuscript.

First, regarding generalization to unseen degradation conditions, we have clearly stated in our previous response that the model is trained using the Real-ESRGAN synthetic degradation pipeline, while all four test datasets, DRealSR, RealSR, RealLR200, and RealLQ250, contain real-world degradations whose degradation conditions were not directly encountered during training. The experimental results show that our method achieves good restoration performance under these unseen real-world degradation conditions, providing experimental evidence for the model's generalization ability from synthetic degradations to real-world scenarios.

Second, regarding the impact of re-degradation supervision on final restoration quality, the DAC ablation study in Table 4 of the revised manuscript shows that introducing the learned degradation representations into the restoration process brings performance improvements. In addition, in our previous response, we further supplemented an ablation experiment comparing contrastive learning alone (Contrastive Only) with contrastive learning combined with re-degradation supervision (Contrastive + Re-Deg). As shown in Table 6, introducing re-degradation supervision improves PSNR, LPIPS, MANIQA, and Q-Align on DRealSR, as well as MANIQA and Q-Align on RealLR200. These results further indicate that re-degradation supervision can improve degradation representation learning and positively contribute to final image restoration quality.

Therefore, we believe that our existing experiments have addressed the reviewer's concerns from three aspects: generalization ability to real-world degradation conditions, the guiding role of degradation representations in the restoration process, and the contribution of re-degradation supervision to final restoration quality.

However, we note that the reviewer still considers the transferability of degradation representations insufficiently demonstrated. We take this concern very seriously and would also like to further understand which aspects of the validation requirements have not yet been satisfied by our current experimental results. We kindly ask the reviewer to further clarify what experimental settings, evaluation metrics, or specific data we are expected to supplement, so that we can improve the relevant validation in a more targeted manner.

Thank you again for the reviewer's valuable comments.

Dear Reviewer,

We sincerely appreciate your further feedback. We would like to kindly draw your attention to our latest response regarding the transferability of degradation representations.

We believe that the additional experiments and analyses provided in our rebuttal and revised manuscript have substantially addressed your concerns, particularly regarding generalization to unseen real-world degradation conditions and the contribution of re-degradation supervision to restoration performance.

As the rebuttal period is approaching its deadline, we would greatly appreciate it if you could take another look at our response when time permits. If you still find any aspects insufficiently addressed, we would be more than happy to provide further clarification or additional evidence to resolve your concerns.

Thank you again for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.