Efficient RAW Image Deblurring with Adaptive Frequency Modulation

We are deeply grateful to the reviewer for their positive assessment and valuable, constructive feedback. We appreciate the recognition of our AFPM module's novelty, our model's efficiency, and its interpretability. We have addressed the suggestions below and will incorporate these new results and discussions into our final manuscript.

---

Response to Weakness 1 & Question 1: Ablation on Frequency Domain Design

We thank the reviewer for this excellent suggestion to ensure our performance gains stem from structural innovation rather than model capacity. We have conducted a crucial control experiment to explicitly assess the contribution of our frequency domain designs.

1. Performance of a Purely Spatial Version: We created a "pure spatial-domain" version of our model, named FrENetSpatial, by completely removing all frequency domain operations.

2. On Comparing with an Iso-Parametric NAFNet: Regarding the suggestion to compare with an iso-parametric NAFNet, we believe our existing results already provide a clear conclusion. As shown in our main paper (Table 1), our FrENet (19.76M Params, 2.22G MACs) achieves a PSNR of 44.73 dB, while NAFNet64 (67.79M Params, 3.96G MACs) achieves 40.35 dB. Our model obtains better performance with fewer parameters and MACs. This result indicates that our performance gains are driven by architectural efficiency rather than increased capacity.

The results of our new ablation study on the Deblur-RAW dataset are as follows:

Table: Ablation Study on Frequency-Domain Design.

This new result provides the final piece of evidence. Comparing FrENet with FrENetSpatial, our frequency domain design provides a performance gain of +2.84 dB. This, combined with our results against NAFNet64, robustly demonstrates that our model's performance is fundamentally driven by our novel and efficient frequency-domain architecture.

---

Response to Weakness 2 & Question 3: Quantifying Runtime Cost

We agree that quantifying the actual runtime cost is essential. While MACs provide a hardware-agnostic measure, real-world inference speed is crucial. We have benchmarked the end-to-end inference time on full-resolution RAW images on a single NVIDIA RTX 5880 Ada GPU.

Table: Efficiency Comparison on Full-Resolution RAW Images.

The results clearly show that despite its reliance on FFT/IFFT, FrENet's highly efficient architecture makes it significantly faster (1.36x to 3x) than other high-performing methods in a practical, high-resolution scenario. This confirms that the FFT/IFFT operations do not create a performance bottleneck in our framework.

---

Response to Question 2: Generalization to Other RAW Restoration Tasks

We thank the reviewer for this valuable suggestion. We agree that the design of AFPM is indeed versatile and that its core principle—modulating frequencies based on their position—is applicable to other RAW-to-RAW restoration tasks where degradations have distinct frequency-domain patterns (e.g., high-frequency noise or demosaicing artifacts). During the rebuttal phase, we have tried to applied our method to RAW-to-RAW denoising, and the training has not completed during the limited rebuttal time. As for demosaicing, current methods mainly focus on RAW-to-sRGB task, where our FrENet has the potential to be effective in RAW processing, and the subsequent RAW-to-sRGB stages may require more specialized design.

---

Once again, we express our sincere gratitude for the reviewer's insightful feedback, which has been instrumental in enhancing the clarity and impact of our manuscript.

We sincerely thank the reviewer for their positive and encouraging evaluation. We are grateful for their recognition of our method's novelty, efficiency, and strong performance. The reviewer's insightful questions are highly valuable, and we appreciate the opportunity to clarify these aspects. We will incorporate these clarifications into our final manuscript.

---

Response to Weakness 1: The content-agnostic AFPM mechanism

We thank the reviewer for this excellent question. Our design choice is rooted in decoupling the general blur restoration strategy from its application to specific image content.

1. Blur is Position-Dependent in the Frequency Domain: Fundamentally, blur acts as a low-pass filter, attenuating high-frequency components more severely than low-frequency ones. Since spectral position is a direct proxy for frequency bands (low frequencies at the center, high frequencies at the periphery), learning a restoration strategy based on position is equivalent to learning how to differentially treat various frequency bands to counteract blur. For instance, the model learns a general rule: "components at the periphery require stronger amplification than those at the center."

2. Decoupling Strategy from Content: Our AFPM module generates modulation kernels ($w_{ij}$, $b_{ij}$) using only positional information. This kernel represents a learned, universal restoration strategy for a specific frequency location. The final modulation, however, is content-dependent, as this strategy is applied to the actual content features $f_{ij}$. This decoupling allows the model to learn a robust and general restoration map without being misled by specific content in any given patch.

3. Ablation Study on Content-Dependent Kernels: To validate this design, we performed a new experiment where the kernel generation was made purely content-dependent (by feeding $f_{ij}$ into the kernel generator). This variant yielded an inferior performance (44.59 dB PSNR vs. our 44.73 dB). This suggests that our decoupled approach, where a content-agnostic strategy is applied to content-specific features, is a more effective and robust design for the deblurring task. We will add a discussion of this insightful ablation study to the paper to further validate our design.

---

Response to Weakness 2: The use of SCA in the frequency domain

We thank the reviewer for this insightful point and will expand on this in our revision.

While originally for spatial data, we reinterpret SCA's operations for the frequency domain. In this context, Global Average Pooling does not average spatial pixels but computes a "global spectral descriptor" for each channel. This single value summarizes the overall energy distribution across all frequencies for that feature.

This allows the network to perform content-adaptive channel recalibration based on the image's global spectral properties. For example, the model can learn to globally suppress a channel that represents high-frequency noise if the input image's overall spectral descriptor indicates a high noise level. It acts as a complementary mechanism to AFPM: while AFPM provides local, position-aware control within the frequency map, SCA provides global, channel-wise control based on the image's overall spectral characteristics.

---

Response to Weakness 3: Potential boundary artifacts from hard partitioning

This is an important and valid concern for any patch-based method. We argue this is not a significant issue in our framework for two key reasons.

1. The Global Nature of IFFT: The modulation occurs in the frequency domain, but artifacts are perceived in the spatial domain. The final IFFT is a global integration operation. Every pixel in the output spatial feature map is an integral of all components across the entire frequency map. This inherent integration process naturally smooths out potential sharp discontinuities between adjacent frequency patches, preventing visible blocking artifacts.

2. Experimental Validation and Trade-off Analysis: Empirically, our extensive results show no such artifacts. To investigate this further, we conducted a new experiment using overlapping patches (50% overlap). While this slightly improved PSNR on Deblur-RAW from 44.73 dB to 44.87 dB (0.14dB higher), it significantly increased inference time from 75.36ms to 93.46ms (~24% slower). This reveals a clear trade-off between a marginal performance gain and a substantial loss in efficiency. Our non-overlapping design achieves a superior balance, which aligns with the core "Efficient" goal of our work. We believe a discussion of this trade-off analysis will be a valuable addition to our paper, and we plan to incorporate it to justify our design choice.

---

Once again, we thank the reviewer for their time and constructive feedback, which have been instrumental in improving our work.

We sincerely thank you for your time and for providing such detailed and constructive feedback. Your insightful comments have been invaluable in helping us identify the weaknesses of our work and have provided clear directions for improvement. We appreciate that you recognized the importance of RAW image deblurring and the novelty of our proposed direction.

We have carefully considered all your concerns. Below, we address each point in detail and outline the concrete revisions and new experiments we will incorporate into the final manuscript to resolve these issues.

---

Response to Weakness 1: Lack of RAW-Specific Design Justification & Novelty

We thank the reviewer for this critical question, which pushes us to better articulate our core contributions.

1. On RAW-Specific Design:

We agree that a clearer justification is needed. The primary advantage of our frequency-domain approach on RAW data stems from a fundamental physical principle: the linearity of RAW data.

• Physical Fidelity: Image blur is physically a convolution operation. In the frequency domain, this corresponds exactly to an element-wise multiplication for linear signals. RAW data, being a direct representation of sensor captures, is linear. Therefore, processing RAW data in the frequency domain allows our model to learn to reverse the blur by learning a physically meaningful multiplicative modulation. In contrast, sRGB images undergo a non-linear gamma correction. Applying frequency-domain deblurring to sRGB data is merely an approximation, as the strict convolution-multiplication relationship is broken by the non-linearity. Our AFPM module is precisely designed to learn this fine-grained, position-aware multiplicative modulation, making it fundamentally better suited for the linear RAW space where the physical model holds true.

• Noise Characteristics: Furthermore, RAW images have simpler, more predictable noise patterns (e.g., shot and read-out noise) compared to the complex, spatially-variant, and often correlated noise in sRGB images post-ISP. The AFPM's ability to adaptively treat different frequency bands allows it to implicitly model and handle RAW-specific noise characteristics more effectively than methods designed for the complex noise distributions in sRGB.

2. On Architectural Difference from LoFormer:

We acknowledge that we build upon a U-Net backbone, which is a common and effective practice in image restoration. However, we must clarify a key architectural distinction: our FrENet is fundamentally a CNN-based architecture, whereas LoFormer is a Transformer-based method.

This is a deliberate design choice motivated by the practical demands of RAW processing. RAW images from modern cameras often have very high resolutions. Transformer-based models, with their quadratic complexity with respect to spatial dimensions, face significant challenges in terms of memory and computational cost when applied to such high-resolution inputs. Our CNN-based design offers superior scalability and efficiency, making it a more practical solution for real-world RAW deblurring applications.

To better highlight our core motivation, we will revise our introduction and methodology sections to more prominently and explicitly articulate these points. This will ensure the critical link between RAW data's linearity and our frequency-domain design is unmistakable for the reader, directly addressing the need for clearer justification you have pointed out.

---

Response to Weakness 2: Limited Performance Gains & Efficiency

1. On Performance Gains:

We appreciate the reviewer's analysis and wish to contextualize our performance gains. Our model's frequency-dependent modulation is particularly effective for the complex, spatially-varying blurs found in real-world scenarios, which is proven by the significant +0.99 dB PSNR gain on the challenging RealBlur-J benchmark (33.87 vs. 32.88). We believe this demonstrates our method's practical value. The marginal gains on synthetic datasets like GoPro and HIDE can be attributed to their simpler blur synthesis, which do not fully reflect real-world degradation. The strong performance on RealBlur-J thus confirms our approach is well-suited for real-world applications.

2. On Efficiency (Parameters vs. Computational Cost):

We agree that the parameter counts are similar. To provide a complete efficiency analysis, we also benchmarked MACs and runtime against LoFormer-L. The results below are from a head-to-head comparison on $256 \times256$ sRGB patches.

Table: Efficiency comparison on $256 \times256$ sRGB patches.

This comprehensive view shows that with a comparable model size, our FrENet+ requires ~20% fewer MACs and is over 2x faster than LoFormer-L. This demonstrates a more favorable balance between model size and computational efficiency, which we will highlight in the paper.

---

Response to Weakness 3: Unfair Experimental Comparison

This is an excellent point, and we agree that the suggested experimental setup provides the most rigorous and fair evaluation. While our original experiments retrained all baselines on RAW data from scratch (by changing input/output channels), which provides a degree of fairness, we acknowledge the reviewer's proposed pipeline is superior.

We have conducted new experiments during this rebuttal period to address these concerns directly. We will add the following table.

We converted the Deblur-RAW dataset to sRGB using the standard LibRaw [r2] processing pipeline and re-trained the LoFormer on this new sRGB dataset. The results are tested on the sRGB images converted from the original RAW test set.

Table: Comparison on sRGB images converted from Deblur-RAW.

---

Response to Weakness 4: Comparing Deblurring Pipelines

We thank the reviewer for this excellent and critical suggestion. To validate the superiority of deblurring in the RAW domain, we have conducted the requested comparison between the two key pipelines:

• (i) Deblur-then-ISP: Deblurring is performed on RAW data first, followed by LibRaw [r2] processing pipeline.
• (ii) ISP-then-Deblur: The blurred RAW image is first converted to sRGB via the LibRaw [r2] pipeline, and then deblurring is applied.

We evaluated both pipelines using two different network architectures, our FrENet and a strong baseline LoFormer-L [26], to ensure the conclusion is generalizable. For a fair comparison, both models were trained separately for each pipeline on the corresponding data (RAW or sRGB). The results are evaluated on the sRGB images from the Deblur-RAW test set.

Table: Pipeline Comparison: (i) Deblur-then-ISP vs. (ii) ISP-then-Deblur.

These new results will empirically validate our core claim: deblurring in the linear RAW domain before applying the ISP yields superior results due to the preservation of critical information.

[r2] LibRaw: Open-source library for processing RAW files.

---

Response to Weakness 5: Incomplete Ablation Study

We appreciate the suggestion for a more detailed and clearer ablation study.

1. Contribution of AFPM and SCA:

We thank the reviewer for this suggestion. We would like to clarify that this analysis is provided in Table 4. In that table, "Local Branch" corresponds to using only the AFPM module, while "Global Branch" uses only the SCA module. The results therefore demonstrate their individual contributions and synergy. We will revise the paper to make this correspondence unambiguous.

2. Module-level Efficiency Analysis:

We agree that a cost breakdown is essential for understanding the efficiency impact. We will add the following table to the Appendix to detail the parameter distribution of the key components within one FrE-Block. The results, presented in the table below, demonstrate that our key components, AFPM and SCA, are highly parameter-efficient. Collectively, they constitute only 28.1% of the model's total parameters. The majority of parameters reside in standard components like the Feed-Forward Network (FFN) and other foundational elements. This highlights that our performance gains are achieved through targeted and efficient architectural contributions, rather than by simply increasing the overall model size.

Table: Per-Module Cost Analysis of FrENet on $128 \times 128$ size patches.

---

Once again, we thank you for your thorough review. We are confident that by incorporating these extensive revisions, new experiments, and detailed analyses, we can address all the concerns raised and significantly strengthen the quality and impact of our paper. We believe the revised manuscript will present a more compelling case for the effectiveness and efficiency of our proposed method for RAW image deblurring.

Dear Reviewer,

Thank you once again for your insightful comments and time spent reviewing our work. As we are approaching the conclusion of the discussion period for our submission, we wished to inquire whether the clarifications and responses in our rebuttal have resolved your concerns.

We deeply appreciate your valuable input.

Thank you very much for your thoughtful feedback. We are glad that our rebuttal has addressed your concerns.

The substantial performance gap you highlighted is indeed a key finding of our work. This improvement stems from the fundamental advantages of processing images in the RAW domain, before the irreversible information loss caused by the non-linear ISP pipeline (e.g., dynamic range compression, gamma correction). As you correctly observed, this is not just a unique benefit of our FrENet architecture. Our experiments show that even LoFormer-L, when adapted to the RAW domain, achieves a significant performance gain compared to its sRGB-domain counterpart. This strongly suggests that deblurring in the linear RAW domain is a highly promising direction with great potential for image restoration research.

To ensure the reproducibility of all our findings, including these pipeline comparisons, we commit to publicly releasing our source code. This will include scripts for:

1. Training and testing models in both the RAW domain and the sRGB domain.
2. The ISP processing pipeline (using LibRaw) to convert RAW data to the sRGB domain for training and evaluation.

This will allow for the complete and transparent reproduction of all results presented in our paper.

Thank you again for your valuable and constructive feedback throughout the review process.

We sincerely thank the reviewer for the detailed and constructive feedback. Your insightful questions have helped us improve the clarity and completeness of our work. We address each of your points below and will incorporate these clarifications into the final manuscript.

---

Response to Weakness 1 & Question 2: AFPM's Positional Encoding

We appreciate the reviewer's excellent point regarding the directional nature of motion blur. This is a crucial aspect we considered during our model's development. Our final design, which uses only Euclidean distance as the positional encoding, was chosen after exploring more complex, direction-aware alternatives, including the 2D sinusoidal coordinates suggested by the reviewer.

Specifically, we conducted experiments to compare our simple distance-based encoding with two direction-aware alternatives: one using Polar coordinates (radius $\gamma$, angle $\theta$) and another using 2D sinusoidal coordinates. The results on the Deblur-RAW dataset are as follows:

Table: Comparison of Positional Encodings in AFPM on Deblur-RAW.

The results are very insightful. Our experiments show that both direction-aware encodings (Polar and Sinusoidal) achieve slightly lower performance in both PSNR and SSIM compared to our simple distance-based method.

This suggests that while providing explicit directional information is a theoretically sound idea, it may introduce an overly complex learning objective for the network. The model might struggle to effectively utilize the directional signals, potentially leading to overfitting on directional artifacts present in the training data, which ultimately compromises both pixel-level accuracy and structural similarity.

Given that our simpler, distance-based approach consistently yields superior results across both primary metrics, we concluded that it offers the most robust and effective solution. It guides the model to focus on inverting the primary low-pass filtering effect of blur—the most critical aspect for restoration—which proves to be a more direct and successful strategy. This empirical finding is why we adopted the distance-only approach in our final model.

---

Response to Weakness 2, 4 & Question 1: Inference Strategy, Efficiency, and Hardware Information

We would like to clarify our inference strategy and provide the requested practical efficiency metrics. We follow the standard practice of using a sliding-window approach, which is also adopted by competing methods like LoFormer [26] and DeepRFT [24], to ensure a fair comparison on high-resolution test images.

A key concern with this approach is often the runtime overhead. However, we want to emphasize that our model's highly efficient architecture translates to a significant real-world performance advantage. We have benchmarked the end-to-end inference time and GPU memory usage on full-resolution RAW images on a single NVIDIA RTX 5880 Ada GPU.

Table: Efficiency Comparison on Full-Resolution RAW Images.

The results clearly show that FrENet is significantly faster (1.36$\times$ to 3$\times$) and more memory-efficient than these powerful Transformer-based baselines in a practical, end-to-end scenario. To clarify hardware information for trainining and testing, we will add the following sentences to Implementation Details: All experiments are conducted on a single NVIDIA RTX 5880 Ada GPU and our model requires approximately 20 hours of training.'

---

Response to Weakness 3 & Question 4: Experimental Comparisons

We appreciate the suggestion to broaden our comparative evaluation; benchmarking against a wider range of models is indeed essential.

We have now included StripFormer [35] in our main comparison table to provide a more comprehensive view against recent efficient models. For RawIR [10] and ELMFormer [23], we were unable to include their results in our direct benchmark at this time, as their official implementations were not available for us to retrain on the Deblur-RAW dataset.

Table: Extended Comparison on Deblur-RAW. MACs and Params are calculated on $128 \times 128 \times 1$ patches.

The expanded comparison shows that our FrENet continues to achieve state-of-the-art PSNR, outperforming even recent efficient models like StripFormer by a significant margin (+1.76 dB). As for Hybrid Attention Transformer, it is not included as its design is primarily focused on super-resolution.

Finally, to reiterate for clarity, all baselines presented in Table 1 of our paper were retrained from scratch by us on the Deblur-RAW dataset. We will ensure this point is explicitly clarified in the paper.

---

Response to Question 3: Processing Fourier Features

Thank you for this insightful question. Our approach of concatenating the real and imaginary parts of the Fourier features is a deliberate and principled design choice, grounded in both theory and empirical validation.

Fundamentally, the real and imaginary parts are the orthogonal components that completely define a signal in the frequency domain. Concatenating them is a direct and effective way to feed this complete information into standard, highly-optimized 2D convolutions. This method aligns with the proven techniques used in other methods like DeepRFT [24] and Fast Fourier Convolution [r1], demonstrating its robustness.

To explore alternatives, we also empirically investigated magnitude-phase concatenation. Our experiments revealed that models using this representation failed to converge during training. We attribute this instability primarily to the challenges of processing the phase component. The phase periodicity creates a highly discontinuous space that is challenging for standard loss functions and gradient-based optimizers, leading to unstable gradients.

In contrast, the real-imaginary representation is continuous and well-behaved, providing a stable learning landscape. Therefore, our findings suggest that our chosen approach offers a more robust and practical path to achieving high-performance frequency-domain processing within modern deep learning frameworks.

[r1] Chi, L., Jiang, B., and Mu, Y. (2020). Fast fourier convolution. Advances in Neural Information Processing Systems, 33, 4479-4488.

---

Response to Question 5: Failure Cases and Limitations

Thank you for this suggestion. Like most deblurring methods, our model has limitations. It can face challenges and produce suboptimal results in scenes with extremely severe and complex blur, for instance, where the motion is exceptionally large or highly non-uniform. In such cases, some residual artifacts may remain. Acknowledging this provides a more balanced view of our method's capabilities. More failure cases will be given in the revised manuscript along with disscussions.

---

Response to Question 6: Ablating Spatial Skip Connections

To provide a more thorough ablation on our network's components, we have updated our results for the skip connections. The table below confirms that both spatial and frequency skip connections are vital, and their synergy is key to our model's strong performance.

Table: Ablation of Skip Connections.

We hope these clarifications and new results comprehensively address the reviewer's concerns.

Thank you for the detailed rebuttal. It does address several of my concerns.

In the efficiency table provided by above, the authors report the end-to-end inference time on full-resolution RAW images. I suppose it includes the total time of tiling of the input image, forward pass of teach tile through the model, and stitching the deblurred tiles back. It would be useful to include the individual components such as "time for model forward pass for 1 tile". i.e., comparison of inference time of all competing models on 1 single tile without any pre- or post-processing.

Regarding the challenges of processing the phase component, I am not particularly convinced by just the hypothesis of "unstable gradients". It would be more convincing to have a valid citation or reference or an experimental result of training a design that processes magnitude and phase component (while stabilizing the training using common engineering techniques). Nonetheless, since it is not the most critical part of the paper, and due to the lack of time, I am fine with author's current response on it.