acceptodds
Under review as a conference paper at ICLR 2027

Improving Complex Moiré Removal with Generative Supervision

13 views

Abstract

The availability of high-quality paired data is essential for training learning-based image demoiréing models. However, it remains challenging for existing datasets to encompass the complex moiré patterns captured in uncontrolled real-world scenarios. Such degradations typically manifest as large-scale, multicolored moiré patterns. Moreover, these patterns frequently occur in images for which clean counterparts are difficult to obtain, such as photographs acquired from public displays and existing online resources. In this work, we propose a novel data engine designed to improve the removal of these challenging moiré patterns by generating training supervision. Specifically, we initially collect real-world images containing complex moiré patterns and localize the corresponding screen regions. Multiple image-conditioned generative foundation models are subsequently deployed to produce candidate references. To establish reliable supervision, these candidates are subjected to patch-level quality control to filter and select the optimal results. Based on this systematic paradigm, we construct the WildMoiré dataset, which contains 6.8K moiré-GT training pairs. For evaluation, we additionally build an independent test set comprising 250 pairs with captured clean ground truth. Extensive experiments on ESDNet, SDXL, and Qwen-Image-Edit demonstrate that the proposed generative supervision consistently improves the performance of complex moiré removal. The code and dataset will be made publicly available.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Low-level Vision, Image Demoiréing, Generative Supervision

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_TdcmOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the lack of clean paired supervision for complex moiré patterns. It uses multiple generative models to produce candidate references and constructs the WildMoiré dataset through alignment, local filtering, and candidate selection. Training on a mixture of this dataset and existing paired data improves the performance of different restoration models.

Critical strengths

Clean source content is often unavailable for public displays and online images, and simply expanding controlled paired data collection is unlikely to cover these degradations. The paper retains real moiré inputs, exploits the complementary behavior of different generators on the candidate-reference side, and combines generative supervision with reliable paired data. This provides a practical way to use real samples for which clean pairs are difficult to obtain. The resulting data can benefit existing restoration models without adding generative computation at deployment.

Critical weaknesses

Generative pseudo-supervision and quality filtering have been studied in related work, so the contribution here needs to be supported primarily by the complex moiré data and the quality of its supervision. However, A-FINE is used for both candidate selection and post-filtering, and its quality preferences do not guarantee that text, human details, and local structures remain unaltered. The current experiments also lack an independent content-fidelity check of the final references. Improvements in downstream restoration show that the data are useful, but do not by themselves establish the reliability of the filtering process.

Questions for authors
  1. How were the post-filtering thresholds and training configurations determined before inspecting the final test results, and which independent validation data were used?
  2. It remains unclear whether the training and test sets were deduplicated by original display content and scene, and whether different captures or crops of the same content could appear in different splits.
  3. Could the authors independently check the content fidelity of the retained references and compare them with candidates from a single generator for the same inputs, to determine whether local selection reduces text alterations, structural changes, and residual moiré?
  4. Do evaluations on complete existing benchmarks, or evaluations with devices and screen content held out, still support the reported restoration improvements?
4 comments
AuthorsOpenReview ↗

Response to Reviewer Tdcm (Part 1/3)

We thank the reviewer for the careful assessment and constructive feedback. We will respond to each point one by one.

1. Validation and configuration (Question 1)

Post-filtering threshold. The post-filtering threshold τpost=15\tau_{post} = 15 was not determined based on test-set performance, instead, it was established via human assessment on a small subset during the early phase of dataset construction. The ablation study in Table C of Appendix C.2 in the orginal main paper (now Table D of Appendix C.3 in the revised main paper) confirms its validity. Training configurations. Training configurations follow the official settings publicly released by each selected method or related work. For ESDNet, we use the optimization recipe of the original work. For SDXL, we adopt the InstructPix2Pix [1] training formulation and follow the configuration of the SDXL training example in its Hugging Face Diffusers. For Qwen-Image-Edit, we use the LoRA fine-tuning settings in DiffSynth-Studio [2], with a learning rate of 10−410^{-4} and LoRA rank 8. Detailed training settings are reported in Section 5.1 and Appendix B.3 of the main paper.

2. Content and split independence (Question 2)

Our captured test set has no overlap with WildMoiré in either original display content or scenes. Therefore, different captures or crops of the same content do not appear in different splits.

3. Independent content-fidelity check (Question 3)

We first randomly sampled 500 groups from WildMoiré. Then we compare the content fidelity of the retained references with candidates from each single generator for the same inputs, as shown in Table R1. We recorded four error types: Residual moiré refers to visible moiré patterns remaining in the reference, Text alterations refer to changes to the original text or distortions of character shapes, Changes in human details mainly refer to visible changes in facial features, Structural changes refer to alterations or loss of object structures in the image. Since each generator has different strengths, our selection strategy (choosing the best out of five candidates) can effectively leverage their complementary advantages to reduce these errors in the final supervision.

Table R1. Content-fidelity comparison on the same 500 groups.

{\let\T\text\def\B#1{\rlap{\kern.03em#1}#1}\small\begin{array}{lcccccc}\hline\T{Error type}&\T{GPT-Image-2}&\T{NanoBanana-2}&\T{SDXL}&\T{FLUX.2}&\T{Qwen-Image-Edit}&\T{Ours}\\\\\hline\T{Residual moir}\acute{\rm e}&2.6\\%&14.2\\%&43.2\\%&30.4\\%&28.2\\%&\B{0.6\\%}\\\\\T{Text alterations}&7.8\\%&3.8\\%&4.8\\%&4\\%&3\\%&\B{1.6\\%}\\\\\T{Changes in human details}&5.4\\%&0.8\\%&2\\%&2\\%&1.6\\%&\B{0.4\\%}\\\\\T{Structural changes}&6.2\\%&1.2\\%&1.4\\%&2.2\\%&3.4\\%&\B{0.6\\%}\\\\\hline\end{array}}

AuthorsOpenReview ↗

Response to Reviewer Tdcm (Part 2/3)

4. Complete and held-out evaluation (Question 4)

Evaluations on complete existing benchmarks. Table R2 and Table R3 report results on the complete UHDM and DCID test sets separately. The results show that our method also improving overall performance on the complete test sets. Evaluations with devices and screen content held out. When constructing our own test set, we deliberately selected screens and scenes strictly distinct from those in WildMoiré, with only the capturing device kept the same. Therefore, to further evaluate generalization with both devices and screen content strictly held out, we evaluate the models on FHDMi [3], an external benchmark collected with completely different capturing devices as well as unseen screens and scenes. As shown in Table R4, the results are highly consistent with our findings in the main paper, further confirming that WildMoiré can improve model generalization capability.

Table R2. Results on the complete UHDM test set.

ModelTraining dataPSNR↑SSIM↑LPIPS↓MUSIQ↑TOPIQ↑Q-Align↑ESDNetUHDM + DCID23.52170.78810.262033.17160.30414.0421ESDNetUHDM + DCID + WildMoireˊ23.651023.65100.79600.79600.24620.246233.254333.25430.30670.30674.07394.0739SDXLUHDM + DCID23.88110.78170.255035.50170.31690.31694.1310SDXLUHDM + DCID + WildMoireˊ24.013124.01310.78320.78320.24730.247335.821935.82190.31574.16724.1672Qwen-Image-EditUHDM + DCID23.86060.79030.26110.261143.66310.40924.2529Qwen-Image-EditUHDM + DCID + WildMoireˊ23.941323.94130.80260.80260.261743.821143.82110.41500.41504.27914.2791{\let\T\text\def\B#1{\rlap{\kern.03em#1}#1}\def\W{\T{UHDM + DCID + WildMoir}\acute{\rm e}}\def\D{\T{UHDM + DCID}}\def\Q{\T{Qwen-Image-Edit}}\small\begin{array}{llrrrrrr}\hline\T{Model}&\T{Training data}&\T{PSNR}↑&\T{SSIM}↑&\T{LPIPS}↓&\T{MUSIQ}↑&\T{TOPIQ}↑&\T{Q-Align}↑\\\\\hline\T{ESDNet}&\D&23.5217&0.7881&0.2620&33.1716&0.3041&4.0421\\\\\T{ESDNet}&\W&\B{23.6510}&\B{0.7960}&\B{0.2462}&\B{33.2543}&\B{0.3067}&\B{4.0739}\\\\[3pt]\T{SDXL}&\D&23.8811&0.7817&0.2550&35.5017&\B{0.3169}&4.1310\\\\\T{SDXL}&\W&\B{24.0131}&\B{0.7832}&\B{0.2473}&\B{35.8219}&0.3157&\B{4.1672}\\\\[3pt]\Q&\D&23.8606&0.7903&\B{0.2611}&43.6631&0.4092&4.2529\\\\\Q&\W&\B{23.9413}&\B{0.8026}&0.2617&\B{43.8211}&\B{0.4150}&\B{4.2791}\\\\\hline\end{array}}

Table R3. Results on the complete DCID test set.

ModelTraining dataPSNR↑SSIM↑LPIPS↓MUSIQ↑TOPIQ↑Q-Align↑ESDNetUHDM + DCID27.08060.88010.233135.57130.31674.0836ESDNetUHDM + DCID + WildMoireˊ27.301227.30120.88790.88790.22820.228235.730135.73010.31820.31824.08614.0861SDXLUHDM + DCID27.10100.88920.235637.58120.32474.1727SDXLUHDM + DCID + WildMoireˊ27.356427.35640.89710.89710.22930.229337.653337.65330.32690.32694.19034.1903Qwen-Image-EditUHDM + DCID27.13490.88520.242146.00190.42674.3076Qwen-Image-EditUHDM + DCID + WildMoireˊ27.348727.34870.90610.90610.24040.240446.210146.21010.42880.42884.35534.3553{\let\T\text\def\B#1{\rlap{\kern.03em#1}#1}\def\W{\T{UHDM + DCID + WildMoir}\acute{\rm e}}\def\D{\T{UHDM + DCID}}\def\Q{\T{Qwen-Image-Edit}}\small\begin{array}{llrrrrrr}\hline\T{Model}&\T{Training data}&\T{PSNR}↑&\T{SSIM}↑&\T{LPIPS}↓&\T{MUSIQ}↑&\T{TOPIQ}↑&\T{Q-Align}↑\\\\\hline\T{ESDNet}&\D&27.0806&0.8801&0.2331&35.5713&0.3167&4.0836\\\\\T{ESDNet}&\W&\B{27.3012}&\B{0.8879}&\B{0.2282}&\B{35.7301}&\B{0.3182}&\B{4.0861}\\\\[3pt]\T{SDXL}&\D&27.1010&0.8892&0.2356&37.5812&0.3247&4.1727\\\\\T{SDXL}&\W&\B{27.3564}&\B{0.8971}&\B{0.2293}&\B{37.6533}&\B{0.3269}&\B{4.1903}\\\\[3pt]\Q&\D&27.1349&0.8852&0.2421&46.0019&0.4267&4.3076\\\\\Q&\W&\B{27.3487}&\B{0.9061}&\B{0.2404}&\B{46.2101}&\B{0.4288}&\B{4.3553}\\\\\hline\end{array}}

Table R4. Results on FHDMi test set.

ModelTraining dataPSNR↑SSIM↑LPIPS↓MUSIQ↑TOPIQ↑Q-Align↑ESDNetUHDM + DCID25.07100.81910.150934.22300.29453.9014ESDNetUHDM + DCID + WildMoireˊ25.765325.76530.84070.84070.14170.141734.917734.91770.30370.30373.97623.9762SDXLUHDM + DCID25.26020.82420.156235.16240.30523.9676SDXLUHDM + DCID + WildMoireˊ25.787025.78700.84010.84010.15330.153336.070136.07010.31090.31094.01104.0110Qwen-Image-EditUHDM + DCID25.32140.83030.158443.28060.37764.1026Qwen-Image-EditUHDM + DCID + WildMoireˊ25.900225.90020.85750.85750.14810.148143.923143.92310.38480.38484.20794.2079{\let\T\text\def\B#1{\rlap{\kern.03em#1}#1}\def\W{\T{UHDM + DCID + WildMoir}\acute{\rm e}}\def\D{\T{UHDM + DCID}}\def\Q{\T{Qwen-Image-Edit}}\small\begin{array}{llrrrrrr}\hline\T{Model}&\T{Training data}&\T{PSNR}↑&\T{SSIM}↑&\T{LPIPS}↓&\T{MUSIQ}↑&\T{TOPIQ}↑&\T{Q-Align}↑\\\\\hline\T{ESDNet}&\D&25.0710&0.8191&0.1509&34.2230&0.2945&3.9014\\\\\T{ESDNet}&\W&\B{25.7653}&\B{0.8407}&\B{0.1417}&\B{34.9177}&\B{0.3037}&\B{3.9762}\\\\[3pt]\T{SDXL}&\D&25.2602&0.8242&0.1562&35.1624&0.3052&3.9676\\\\\T{SDXL}&\W&\B{25.7870}&\B{0.8401}&\B{0.1533}&\B{36.0701}&\B{0.3109}&\B{4.0110}\\\\[3pt]\Q&\D&25.3214&0.8303&0.1584&43.2806&0.3776&4.1026\\\\\Q&\W&\B{25.9002}&\B{0.8575}&\B{0.1481}&\B{43.9231}&\B{0.3848}&\B{4.2079}\\\\\hline\end{array}}

5. Reliability of retained references (Weakness 1)

Kindly refer to your Q3, in which we have checked the content-fidelity of the final references independently. Furthermore, we compare A-FINE with CKDN [4], DRGQA [5], and Q-Insight [6], which are applicable to imperfect references. Specifically, we first randomly sample 1,000 groups from WildMoiré, each containing one moiré input and five candidate references. Then, we calculate the overlap rate between the candidates selected by each criterion and those selected by humans, as shown in Table R5. A-FINE achieves a significantly higher alignment with human judgments than other criteria, confirming that it is the most reliable selection criterion for the demoiréing task. Together with the independent content-fidelity check of the final references, these results provide sufficient evidence to validate the reliability of our filtering process.

Table R5. Overlap rate with human selection.

{\let\T\text\def\B#1{\rlap{\kern.03em#1}#1}\small\begin{array}{lcccc}\hline\T{Selection criteria}&\T{CKDN}&\T{DRGQA}&\T{Q-Insight}&\T{A-FINE}\\\\\hline\T{Overlap rate (\\%)}&71.2&80.4&85.2&\B{96.9}\\\\\hline\end{array}}

AuthorsOpenReview ↗

Response to Reviewer Tdcm (Part 3/3)

References

[1] Tim Brooks, Aleksander Holynski, and Alexei A. Efros. InstructPix2Pix: Learning to Follow Image Editing Instructions. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 18392–18402, 2023.

[2] ModelScope Team. DiffSynth-Studio: An open-source diffusion model engine. https://github.com/modelscope/DiffSynth-Studio, 2024.

[3] Bin He, Ce Wang, Boxin Shi, and Ling-Yu Duan. FHDe$^{2}$Net: Full High Definition Demoireing Network. European Conference on Computer Vision (ECCV), pp. 713–729, 2020.

[4] Heliang Zheng, Huan Yang, Jianlong Fu, Zheng-Jun Zha, and Jiebo Luo. Learning Conditional Knowledge Distillation for Degraded-Reference Image Quality Assessment. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp. 10242–10251, 2021.

[5] Shishun Tian, Tiantian Zeng, Zhengyu Zhang, Wenbin Zou, and Xia Li. Dual Residual-Guided Interactive Learning for the Quality Assessment of Enhanced Images. IEEE Transactions on Multimedia, 27, pp. 1637–1651, 2025.

[6] Weiqi Li, Xuanyu Zhang, Shijie Zhao, Yabin Zhang, Junlin Li, Li Zhang, and Jian Zhang. Q-Insight: Understanding Image Quality via Visual Reinforcement Learning. In Advances in Neural Information Processing Systems, vol. 38, pp. 41337–41362, 2025.

AuthorsOpenReview ↗

Thanks & Friendly Reminder

We sincerely appreciate your thoughtful review and constructive feedback. As the discussion period ends tomorrow, we would be grateful to know whether our rebuttal has addressed your concerns. If anything remains unclear, we would be happy to provide further clarification.

Reviewer_TsZSOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces WildMoire, a dataset which contains 6.8K moire-GT training pairs. The authors collected complex real-world moire images that have no corresponding GTs using mobile phones and from the web. Hence, a pipeline is developed to reliably use generative supervision. The pipeline includes candidate GT generation, data processing and GT selection. The models used for candidate GT generation are GPT-Image-2, Nano-Banana-2, SDXL, FLUX-2 and Qwen-Image-Edit. A-FINE is incorporated to select the best GT candidate independently for each input patch. Experiments are conducted to compare training 3 models from different families on existing demoiring datasets with and without WildMoire. All models are evaluated on a test set of collected 247 additional moire images whose ground-truths are obtained through camera capture rather than generation. In addition, a scale transformation data augmentation is proposed and evaluated to improve robustness to moire patterns at different scales

Critical strengths
  1. The 247-pair test set uses captured clean ground truths rather than generations to simulate real-world labels and strengthen results credibility.
  2. The use of A-FINE as the selection criterion is well motivated. It avoids both the perfect reference assumption of full-reference metrics and the content inconsistency of no-reference metrics.
  3. Experiments (Table 3) show that using WildMoire in the training data improves the quantitative results significantly on the collected test set across all reported metrics and model families. Using WildMoire also shows clear qualitative improvements in the demoiring outputs of all evaluated models (Figure 3).
  4. Challenging cases from the UHDM and DCID test sets are used as additional evaluation beyond the authors’ collected test set which also shows gains from using WildMoire as part of the training data (Table 4 and Figure B in Appendix C.3).
  5. The writing is clear and the paper reports exact prompts (Table A in Appendix B.2), thresholds, training hyperparameters, and fine-tuning details.
Critical weaknesses
  1. The introduction defines the core problem as models yielding satisfactory results “only within the distribution of the training data, exhibiting limited generalizability across different screens and complex scenes”. The paper then claims that WildMoire improves “generalization ability” (Section 4.2). However, all evaluations are on data that is in-distribution for the trained models: the authors’ own test set (whose variations from WildMoire in screens, scenes, and devices is not stated) and the UHDM/DCID challenging test cases. Evaluation on an external benchmark not used in training is needed to support the generalization claim.
  2. The paper only evaluates WildMoire mixed with UHDM and DCID, where real clean ground-truths can mask noise in the generated labels. An ablation training on WildMoire alone is needed to show what errors the pseudo-labels caused and how WildMoire affects both simple and complex moire.
  3. The paper does not explain how the pre-filtering thresholds were selected, and unlike the post-filtering threshold, they are not ablated. Furthermore, the post-filtering threshold value is selected using the same 247-pair test set used for reporting results (Table C in Appendix C.2). Thresholds should be chosen on a separate validation split and the test set should be used only for final evaluation. 4.Supervision quality is controlled only by automated, relative criteria, with no human verification. The final quality stage (post-filtering) only requires the selected candidate to improve on the input’s A-FINE score by at least a threshold. Since A-FINE selects the best of five candidates, which may all be flawed, and a severely degraded input has a very poor self-score, a candidate that retains residual moire or alters content may still be retained. No safeguard specific to severe cases is provided, despite the dataset’s emphasis on complex moire. The limitations section acknowledges that residual moire, text modifications, and color inconsistencies may remain, but the paper gives no estimate of how often this occurs.
  4. Quantitative improvements on UHDM and DCID challenging cases (Table 4) are much smaller than on the authors’ own test set (Table 3). This weakens the claim that WildMoire substantially improves robustness on challenging moire.
  5. The contribution is mainly the dataset and an engineering pipeline, so the novelty is limited.
Questions for authors

Even when high post-filtering thresholds are used and only 2.8K samples are retained from WildMore, results across most metrics is comparable with the optimal case of using 6.8K samples (Table C in Appendix C.2). Does this indicate that a small high-quality subset is as useful as a larger, lower-quality one, or that returns from dataset size saturate quickly?

3 comments
AuthorsOpenReview ↗

Response to Reviewer TsZS (Part 1/2)

We thank the reviewer for the careful assessment and constructive feedback. We address the comments below.

1. Data quantity and quality (Question 1)

Ideally, a large-scale and high-quality dataset would be optimal. For more general tasks, we agree that a small high-quality subset might be as useful as a larger, lower-quality one. However, in this work, we aimed to increase the amount of data as much as possible while maintaining data quality during dataset construction. Although τpost=15\tau_{post} = 15 only showed a slight improvement in PSNR (0.07 dB) compared to τpost=25\tau_{post} = 25, the LPIPS showed a significant improvement (0.2925 vs 0.2614). Overall, we believe that increasing the amount of data is beneficial while maintaining data quality.

2. External generalization (Weakness 1)

When constructing our own test set, we deliberately selected screens and scenes strictly distinct from those in WildMoiré, with only the capturing device kept the same. Therefore, the performance gains brought by WildMoiré in Table 3 and Table 4 of the main paper can demonstrate its effectiveness in improving the generalization ability of the demoiréing models. Furthermore, we evaluate these models on FHDMi [1], an external benchmark collected with completely different capturing devices as well as unseen screens and scenes. As shown in the Table R1, the results are highly consistent with our findings in the main paper, further confirming that WildMoiré can improve their generalization capability.

Table R1. Results on FHDMi test set.

ModelTraining dataPSNR↑SSIM↑LPIPS↓MUSIQ↑TOPIQ↑Q-Align↑ESDNetUHDM + DCID25.07100.81910.150934.22300.29453.9014ESDNetUHDM + DCID + WildMoireˊ25.765325.76530.84070.84070.14170.141734.917734.91770.30370.30373.97623.9762SDXLUHDM + DCID25.26020.82420.156235.16240.30523.9676SDXLUHDM + DCID + WildMoireˊ25.787025.78700.84010.84010.15330.153336.070136.07010.31090.31094.01104.0110Qwen-Image-EditUHDM + DCID25.32140.83030.158443.28060.37764.1026Qwen-Image-EditUHDM + DCID + WildMoireˊ25.900225.90020.85750.85750.14810.148143.923143.92310.38480.38484.20794.2079{\let\T\text\def\B#1{\rlap{\kern.03em#1}#1}\def\W{\T{UHDM + DCID + WildMoir}\acute{\rm e}}\def\D{\T{UHDM + DCID}}\def\Q{\T{Qwen-Image-Edit}}\small\begin{array}{llrrrrrr}\hline\T{Model}&\T{Training data}&\T{PSNR}↑&\T{SSIM}↑&\T{LPIPS}↓&\T{MUSIQ}↑&\T{TOPIQ}↑&\T{Q-Align}↑\\\\\hline\T{ESDNet}&\D&25.0710&0.8191&0.1509&34.2230&0.2945&3.9014\\\\\T{ESDNet}&\W&\B{25.7653}&\B{0.8407}&\B{0.1417}&\B{34.9177}&\B{0.3037}&\B{3.9762}\\\\[3pt]\T{SDXL}&\D&25.2602&0.8242&0.1562&35.1624&0.3052&3.9676\\\\\T{SDXL}&\W&\B{25.7870}&\B{0.8401}&\B{0.1533}&\B{36.0701}&\B{0.3109}&\B{4.0110}\\\\[3pt]\Q&\D&25.3214&0.8303&0.1584&43.2806&0.3776&4.1026\\\\\Q&\W&\B{25.9002}&\B{0.8575}&\B{0.1481}&\B{43.9231}&\B{0.3848}&\B{4.2079}\\\\\hline\end{array}}

3. Training on WildMoiré alone (Weakness 2)

We supplement an ablation study with ESDNet trained solely on WildMoiré, as shown in Table R2, Table R3 and Table R4. On our captured test set, which consists of images with complex moiré patterns, the model trained on WildMoiré alone achieves performance comparable to that of the model jointly trained on WildMoiré, UHDM and DCID (Table R2). This is because the data in WildMoiré are as challenging as those in our captured test set and provide sufficiently reliable supervision, while UHDM and DCID mainly contain simple moiré patterns, thus offering only limited assistance. Conversely, on the UHDM and DCID test sets, the data predominantly consist of simple moiré patterns with only a small number of complex cases, so jointly training on all three datasets yields better performance (Table R3 and Table R4).

Table R2. ESDNet data ablation on our captured test set.

Training dataPSNR↑SSIM↑LPIPS↓MUSIQ↑TOPIQ↑Q-Align↑WildMoireˊ23.33210.78300.78300.261744.241244.24120.34093.96163.9616UHDM + DCID21.50020.73410.347138.19710.27703.6343UHDM + DCID + WildMoireˊ23.371423.37140.78210.26140.261444.21090.34130.34133.9607{\let\T\text\def\B#1{\rlap{\kern.03em#1}#1}\small\begin{array}{lrrrrrr}\hline\T{Training data}&\T{PSNR}↑&\T{SSIM}↑&\T{LPIPS}↓&\T{MUSIQ}↑&\T{TOPIQ}↑&\T{Q-Align}↑\\\\\hline\T{WildMoir}\acute{\rm e}&23.3321&\B{0.7830}&0.2617&\B{44.2412}&0.3409&\B{3.9616}\\\\\T{UHDM + DCID}&21.5002&0.7341&0.3471&38.1971&0.2770&3.6343\\\\\T{UHDM + DCID + WildMoir}\acute{\rm e}&\B{23.3714}&0.7821&\B{0.2614}&44.2109&\B{0.3413}&3.9607\\\\\hline\end{array}}

Table R3. ESDNet data ablation on the complete UHDM test set.

Training dataPSNR↑SSIM↑LPIPS↓MUSIQ↑TOPIQ↑Q-Align↑WildMoireˊ21.66300.74070.324530.11240.28633.5759UHDM + DCID23.52170.78810.262033.17160.30414.0421UHDM + DCID + WildMoireˊ23.651023.65100.79600.79600.24620.246233.254333.25430.30670.30674.07394.0739{\let\T\text\def\B#1{\rlap{\kern.03em#1}#1}\small\begin{array}{lrrrrrr}\hline\T{Training data}&\T{PSNR}↑&\T{SSIM}↑&\T{LPIPS}↓&\T{MUSIQ}↑&\T{TOPIQ}↑&\T{Q-Align}↑\\\\\hline\T{WildMoir}\acute{\rm e}&21.6630&0.7407&0.3245&30.1124&0.2863&3.5759\\\\\T{UHDM + DCID}&23.5217&0.7881&0.2620&33.1716&0.3041&4.0421\\\\\T{UHDM + DCID + WildMoir}\acute{\rm e}&\B{23.6510}&\B{0.7960}&\B{0.2462}&\B{33.2543}&\B{0.3067}&\B{4.0739}\\\\\hline\end{array}}

Table R4. ESDNet data ablation on the complete DCID test set.

Training dataPSNR↑SSIM↑LPIPS↓MUSIQ↑TOPIQ↑Q-Align↑WildMoireˊ23.79410.79760.296931.76600.29223.6882UHDM + DCID27.08060.88010.233135.57130.31674.0836UHDM + DCID + WildMoireˊ27.301227.30120.88790.88790.22820.228235.730135.73010.31820.31824.08614.0861{\let\T\text\def\B#1{\rlap{\kern.03em#1}#1}\small\begin{array}{lrrrrrr}\hline\T{Training data}&\T{PSNR}↑&\T{SSIM}↑&\T{LPIPS}↓&\T{MUSIQ}↑&\T{TOPIQ}↑&\T{Q-Align}↑\\\\\hline\T{WildMoir}\acute{\rm e}&23.7941&0.7976&0.2969&31.7660&0.2922&3.6882\\\\\T{UHDM + DCID}&27.0806&0.8801&0.2331&35.5713&0.3167&4.0836\\\\\T{UHDM + DCID + WildMoir}\acute{\rm e}&\B{27.3012}&\B{0.8879}&\B{0.2282}&\B{35.7301}&\B{0.3182}&\B{4.0861}\\\\\hline\end{array}}

AuthorsOpenReview ↗

Response to Reviewer TsZS (Part 2/2)

4. Threshold settings and supervision quality (Weakness 3)

Pre-filtering thresholds.

Pre-filtering aims to ensure that training inputs contain distinct moiré patterns. Therefore, we adopt a coarse-to-fine, two-stage pipeline to filter unqualified images. Stage 1: Following HGGT [2], we quickly filter out smooth images (since moiré images typically contain rich high-frequency textures). Since our goal aligns with HGGT, we directly adopt their threshold. Stage 2: We leverage the prior of demoiréing model to further eliminate images with rich textures but weak moiré patterns. Specifically, as detailed in the main paper, we process the images retained from the Stage 1 using ESDNet and compute the SSIM between its inputs and outputs. The images exceeding the SSIM threshold (τssim\tau_{ssim}) are discarded. τssim=0.90\tau_{ssim}=0.90 was initially determined via human assessment on a small subset during early phase of dataset construction, and our supplementary ablation study (Table R5) further confirms its validity.

Post-filtering threshold. The post-filtering threshold τpost=15\tau_{post} = 15 was not determined based on test-set performance, instead, it was established via human assessment on a small subset during the early phase of dataset construction. The ablation study of Table D in Appendix C.3 in revised main paper confirms its validity.

Supervision quality. After obtaining the five candidates, we first score each candidate using A-FINE and select the highest-scoring one as the ground truth (Optimal GT Selection in the main paper). To further ensure supervision quality, our post-filtering discards low-quality supervision, effectively acting as a safeguard. Table D in Appendix C.3 demonstrates that setting the post-filtering threshold too low (e.g., τpost=5\tau_{\text{post}} = 5) reduces supervision quality and consequently degrades the model's demoiréing performance.

Error frequencies. We randomly sample 500 image pairs from WildMoiré and manually inspect their ground-truths for residual moiré, text modifications, and color inconsistencies. Table R6 reports the frequency of each error type in this subset.

Table R5. Sensitivity to the SSIM pre-filtering threshold.

τssimPSNR↑SSIM↑LPIPS↓MUSIQ↑TOPIQ↑Q-Align↑0.8023.19810.77020.278742.79120.32713.81030.8523.27070.77670.272743.26790.33243.87050.9023.371423.37140.78210.78210.261444.210944.21090.34133.96073.96070.9523.37020.78190.26110.261144.20930.34170.34173.9596{\let\T\text\def\B#1{\rlap{\kern.03em#1}#1}\small\begin{array}{crrrrrr}\hline\tau_{\mathrm{ssim}}&\T{PSNR}↑&\T{SSIM}↑&\T{LPIPS}↓&\T{MUSIQ}↑&\T{TOPIQ}↑&\T{Q-Align}↑\\\\\hline0.80&23.1981&0.7702&0.2787&42.7912&0.3271&3.8103\\\\0.85&23.2707&0.7767&0.2727&43.2679&0.3324&3.8705\\\\0.90&\B{23.3714}&\B{0.7821}&0.2614&\B{44.2109}&0.3413&\B{3.9607}\\\\0.95&23.3702&0.7819&\B{0.2611}&44.2093&\B{0.3417}&3.9596\\\\\hline\end{array}}

Table R6. Error frequencies in 500 randomly sampled pairs from WildMoiré.

{\let\T\text\small\begin{array}{lccc}\hline\T{Error type}&\T{Residual moir}\acute{\rm e}&\T{Text modifications}&\T{Color inconsistencies}\\\\\hline\T{Frequency (\\%)}&1.6&0.8&1.4\\\\\hline\end{array}}

5. Gains on public subsets (Weakness 4)

In fact, even the challenging cases in UHDM and DCID are easier than those in our captured test set, as reflected in Tables 3 and 4 of the main paper. Specifically, Qwen-Image-Edit trained with WildMoiré, UHDM and DCID achieves the highest PSNR on both test sets. However, its PSNR on the UHDM and DCID challenging cases is approximately 3.4 dB higher than that on our captured test set (27.3071 dB vs 23.9313 dB), demonstrating that removing moiré from the UHDM and DCID challenging cases is comparatively easier. Therefore, performance on the UHDM and DCID challenging cases is more likely to saturate, leaving less room for further improvement. Consequently, the quantitative improvements on the challenging cases of UHDM and DCID are smaller than those on our captured test set.

6. Novelty (Weakness 5)

Existing demoiréing methods struggle to handle complex moiré images effectively, often suffering from low fidelity or residual moiré. Meanwhile, collecting real-world paired data with complex moiré is costly. To address this, we propose a generative supervision paradigm that selects the optimal outputs from multiple generative models and applies strict filtering to obtain high-quality supervision. Furthermore, including the data constructed via this pipeline (WildMoiré) into the training set significantly improves the generalization capability of demoiréing models. To the best of our knowledge, no similar approach has been explored in previous demoiréing methods, and the novelty is acceptable.

References

[1] Bin He, Ce Wang, Boxin Shi, and Ling-Yu Duan. FHDe$^{2}$Net: Full High Definition Demoireing Network. European Conference on Computer Vision (ECCV), pp. 713–729, 2020.

[2] Du Chen, Jie Liang, Xindong Zhang, Ming Liu, Hui Zeng, and Lei Zhang. Human Guided Ground-Truth Generation for Realistic Image Super-Resolution. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 14082–14091, 2023.

AuthorsOpenReview ↗

Thanks & Friendly Reminder

We sincerely appreciate your thoughtful review and constructive feedback. As the discussion period ends tomorrow, we would be grateful to know whether our rebuttal has addressed your concerns. If anything remains unclear, we would be happy to provide further clarification.

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

The paper proposes a data engine for complex real-world image demoiréing. Five image-conditioned generative models produce candidate clean references for unpaired moiré photographs; spatial and color alignment, patch filtering, and A-FINE-based local selection produce the 6,832-pair WildMoiré dataset. The generated supervision is mixed with UHDM and DCID to train ESDNet, SDXL, and Qwen-Image-Edit. Evaluation uses an independently captured 247-pair test set and selected challenging samples from the UHDM and DCID test sets. The results support the utility of the generated supervision, although the selection criterion, evaluation coverage, and validation protocol still require stronger evidence or clarification.

Critical strengths

The paper addresses the practical difficulty of obtaining paired supervision for complex moiré patterns in uncontrolled scenes. The data engine combines complementary generators with spatial and photometric alignment and local quality control, and WildMoiré improves three distinct model architectures. The additional UHDM/DCID difficult-subset evaluation provides evidence beyond the self-collected test set. Exact prompts, filtering statistics, a clearer capture protocol, and explicitly matched training sample counts improve transparency and experimental rigor. The generated data can be reused without increasing the deployed models’ inference cost.

Critical weaknesses

W1. The suitability of A-FINE for demoiréing supervision selection is insufficiently validated. Although the paper motivates A-FINE through its support for imperfect references, it does not establish whether its candidate rankings reliably reflect moiré removal and content preservation. Task-specific analysis and comparisons with alternative quality or fidelity criteria are missing, leaving the basis for this central selection mechanism insufficiently supported.

W2. Evaluation coverage remains incomplete. Although Table 4 adds public-benchmark evidence, it reports pooled results on only 11 UHDM and 378 DCID samples selected by the proposed pre-filter, leaving full-test-set performance and dataset-specific gains unclear.

W3. Comparisons with recent dedicated demoiréing methods are missing. The experiments mainly compare three selected architectures before and after adding WildMoiré, together with two commercial image editors. These results demonstrate a data benefit within the chosen architectures, but do not establish competitiveness against recent specialized demoiréing approaches.

Questions for authors

Q1. Task-specific validation of A-FINE. Can the authors compare A-FINE with alternative selection criteria applicable to imperfect references, using the same candidate pool? Evaluation on a small paired diagnostic set or through human assessment would help determine whether its preferred candidates remove moiré while preserving text, textures, and structures.

Q2. Public-benchmark coverage and test-set acquisition. Can the authors report UHDM and DCID results separately and, where feasible, evaluate their complete test sets?

Q3. Comparisons with specialized methods. Can recent dedicated demoiréing methods be evaluated under the same test protocol, with their training-data settings clearly stated? Comparisons on the captured test set and public subsets would clarify how the models trained with WildMoiré perform relative to recent specialized approaches.

4 comments
AuthorsOpenReview ↗

Response to Reviewer WXPG (Part 1/3)

We thank the reviewer for the careful assessment and constructive feedback. We address the comments below.

1. Task-specific validation of A-FINE (Weakness 1; Question 1)

We compare A-FINE with CKDN [1], DRGQA [2], and Q-Insight [3], which are applicable to imperfect references. Specifically, we first randomly sample 1,000 groups from WildMoiré, each containing one moiré input and five candidate references. Then, we calculate the overlap rate between the candidates selected by each criterion and those selected by humans, as shown in Table R1. A-FINE achieves a significantly higher alignment with human judgments than other criteria, confirming that it is the most reliable selection criterion for the demoiréing task.

Table R1. Overlap rate with human selection.

{\let\T\text\def\B#1{\rlap{\kern.03em#1}#1}\small\begin{array}{lcccc}\hline\T{Selection criteria}&\T{CKDN}&\T{DRGQA}&\T{Q-Insight}&\T{A-FINE}\\\\\hline\T{Overlap rate (\\%)}&71.2&80.4&85.2&\B{96.9}\\\\\hline\end{array}}

2. Complete public benchmarks (Weakness 2; Question 2)

Table R2 and Table R3 report results on the complete UHDM and DCID test sets separately. Together with Table 4 in the main paper, these results show that our method improves performance on challenging cases while also improving overall performance on the complete test sets.

Table R2. Results on the complete UHDM test set.

ModelTraining dataPSNR↑SSIM↑LPIPS↓MUSIQ↑TOPIQ↑Q-Align↑ESDNetUHDM + DCID23.52170.78810.262033.17160.30414.0421ESDNetUHDM + DCID + WildMoireˊ23.651023.65100.79600.79600.24620.246233.254333.25430.30670.30674.07394.0739SDXLUHDM + DCID23.88110.78170.255035.50170.31690.31694.1310SDXLUHDM + DCID + WildMoireˊ24.013124.01310.78320.78320.24730.247335.821935.82190.31574.16724.1672Qwen-Image-EditUHDM + DCID23.86060.79030.26110.261143.66310.40924.2529Qwen-Image-EditUHDM + DCID + WildMoireˊ23.941323.94130.80260.80260.261743.821143.82110.41500.41504.27914.2791{\let\T\text\def\B#1{\rlap{\kern.03em#1}#1}\def\W{\T{UHDM + DCID + WildMoir}\acute{\rm e}}\def\D{\T{UHDM + DCID}}\def\Q{\T{Qwen-Image-Edit}}\small\begin{array}{llrrrrrr}\hline\T{Model}&\T{Training data}&\T{PSNR}↑&\T{SSIM}↑&\T{LPIPS}↓&\T{MUSIQ}↑&\T{TOPIQ}↑&\T{Q-Align}↑\\\\\hline\T{ESDNet}&\D&23.5217&0.7881&0.2620&33.1716&0.3041&4.0421\\\\\T{ESDNet}&\W&\B{23.6510}&\B{0.7960}&\B{0.2462}&\B{33.2543}&\B{0.3067}&\B{4.0739}\\\\[3pt]\T{SDXL}&\D&23.8811&0.7817&0.2550&35.5017&\B{0.3169}&4.1310\\\\\T{SDXL}&\W&\B{24.0131}&\B{0.7832}&\B{0.2473}&\B{35.8219}&0.3157&\B{4.1672}\\\\[3pt]\Q&\D&23.8606&0.7903&\B{0.2611}&43.6631&0.4092&4.2529\\\\\Q&\W&\B{23.9413}&\B{0.8026}&0.2617&\B{43.8211}&\B{0.4150}&\B{4.2791}\\\\\hline\end{array}}

Table R3. Results on the complete DCID test set.

ModelTraining dataPSNR↑SSIM↑LPIPS↓MUSIQ↑TOPIQ↑Q-Align↑ESDNetUHDM + DCID27.08060.88010.233135.57130.31674.0836ESDNetUHDM + DCID + WildMoireˊ27.301227.30120.88790.88790.22820.228235.730135.73010.31820.31824.08614.0861SDXLUHDM + DCID27.10100.88920.235637.58120.32474.1727SDXLUHDM + DCID + WildMoireˊ27.356427.35640.89710.89710.22930.229337.653337.65330.32690.32694.19034.1903Qwen-Image-EditUHDM + DCID27.13490.88520.242146.00190.42674.3076Qwen-Image-EditUHDM + DCID + WildMoireˊ27.348727.34870.90610.90610.24040.240446.210146.21010.42880.42884.35534.3553{\let\T\text\def\B#1{\rlap{\kern.03em#1}#1}\def\W{\T{UHDM + DCID + WildMoir}\acute{\rm e}}\def\D{\T{UHDM + DCID}}\def\Q{\T{Qwen-Image-Edit}}\small\begin{array}{llrrrrrr}\hline\T{Model}&\T{Training data}&\T{PSNR}↑&\T{SSIM}↑&\T{LPIPS}↓&\T{MUSIQ}↑&\T{TOPIQ}↑&\T{Q-Align}↑\\\\\hline\T{ESDNet}&\D&27.0806&0.8801&0.2331&35.5713&0.3167&4.0836\\\\\T{ESDNet}&\W&\B{27.3012}&\B{0.8879}&\B{0.2282}&\B{35.7301}&\B{0.3182}&\B{4.0861}\\\\[3pt]\T{SDXL}&\D&27.1010&0.8892&0.2356&37.5812&0.3247&4.1727\\\\\T{SDXL}&\W&\B{27.3564}&\B{0.8971}&\B{0.2293}&\B{37.6533}&\B{0.3269}&\B{4.1903}\\\\[3pt]\Q&\D&27.1349&0.8852&0.2421&46.0019&0.4267&4.3076\\\\\Q&\W&\B{27.3487}&\B{0.9061}&\B{0.2404}&\B{46.2101}&\B{0.4288}&\B{4.3553}\\\\\hline\end{array}}

AuthorsOpenReview ↗

Response to Reviewer WXPG (Part 2/3)

3. Comparisons with specialized methods (Weakness 3; Question 3)

We retrained recent demoiréing methods (SADT [4] and FaDeNet [5]) on UHDM+DCID and UHDM+DCID+WildMoiré, comparing them with their publicly released checkpoints on our captured test set and public subsets. As shown in Table R4 and Table R5, both methods achieve the best results when trained on UHDM+DCID+WildMoiré, highlighting that our data improve model capability in handling complex moiré patterns. The publicly released checkpoints of SADT were trained separately on the TIP2018 [6], FHDMi [7], and LCDMoire [8]. Both TIP2018 and FHDMi were constructed by displaying images on screens and capturing them with smartphones. They are limited in scale and contain relatively simple moiré patterns. Meanwhile, LCDMoire is a synthetic dataset that exhibits a domain gap compared to real-world moiré images. The publicly released checkpoint of FaDeNet was trained on its PRISM dataset, which contains only synthetic fabric moiré images, thereby suffering from a domain gap with real-world data and limited scene diversity. In contrast, our dataset (UHDM+DCID+WildMoiré) includes not only simple and small-scale moiré patterns but also complex ones across diverse real-world scenes, which enhances the demoiréing performance of models.

Table R4. Dedicated methods on our captured test set.

ModelTraining dataPSNR↑SSIM↑LPIPS↓MUSIQ↑TOPIQ↑Q-Align↑SADTTIP201818.64310.69710.321732.34110.25163.4876SADTFHDMi20.30910.72020.348234.79870.26973.5780SADTLCDMoire20.02320.71180.339734.11030.27113.5212SADTUHDM + DCID21.82190.74040.325338.76240.28663.7919SADTUHDM + DCID + WildMoireˊ23.676423.67640.78710.78710.26030.260345.613245.61320.35900.35904.00434.0043FaDeNetPRISM20.66570.72790.327135.87920.27933.7514FaDeNetUHDM + DCID22.09070.73790.317238.10950.29013.8417FaDeNetUHDM + DCID + WildMoireˊ23.813523.81350.79400.79400.25340.253445.375145.37510.36260.36263.97013.9701{\let\T\text\def\B#1{\rlap{\kern.03em#1}#1}\def\W{\T{UHDM + DCID + WildMoir}\acute{\rm e}}\small\begin{array}{llrrrrrr}\hline\T{Model}&\T{Training data}&\T{PSNR}↑&\T{SSIM}↑&\T{LPIPS}↓&\T{MUSIQ}↑&\T{TOPIQ}↑&\T{Q-Align}↑\\\\\hline\T{SADT}&\T{TIP2018}&18.6431&0.6971&0.3217&32.3411&0.2516&3.4876\\\\\T{SADT}&\T{FHDMi}&20.3091&0.7202&0.3482&34.7987&0.2697&3.5780\\\\\T{SADT}&\T{LCDMoire}&20.0232&0.7118&0.3397&34.1103&0.2711&3.5212\\\\\T{SADT}&\T{UHDM + DCID}&21.8219&0.7404&0.3253&38.7624&0.2866&3.7919\\\\\T{SADT}&\W&\B{23.6764}&\B{0.7871}&\B{0.2603}&\B{45.6132}&\B{0.3590}&\B{4.0043}\\\\[3pt]\T{FaDeNet}&\T{PRISM}&20.6657&0.7279&0.3271&35.8792&0.2793&3.7514\\\\\T{FaDeNet}&\T{UHDM + DCID}&22.0907&0.7379&0.3172&38.1095&0.2901&3.8417\\\\\T{FaDeNet}&\W&\B{23.8135}&\B{0.7940}&\B{0.2534}&\B{45.3751}&\B{0.3626}&\B{3.9701}\\\\\hline\end{array}}

Table R5. Dedicated methods on the public challenging subsets.

ModelTraining dataPSNR↑SSIM↑LPIPS↓MUSIQ↑TOPIQ↑Q-Align↑SADTTIP201822.79960.76040.310830.10210.24123.5974SADTFHDMi24.17160.79310.291631.77640.26313.6671SADTLCDMoire23.32400.79020.295731.23160.26193.6211SADTUHDM + DCID27.18330.87620.247737.50050.32494.0713SADTUHDM + DCID + WildMoireˊ27.901427.90140.88420.88420.23100.231038.432138.43210.33310.33314.14514.1451FaDeNetPRISM23.86190.77840.298632.21430.26903.6851FaDeNetUHDM + DCID27.52120.88060.249238.10340.32274.0932FaDeNetUHDM + DCID + WildMoireˊ28.173028.17300.89410.89410.23550.235539.231439.23140.32960.32964.15074.1507{\let\T\text\def\B#1{\rlap{\kern.03em#1}#1}\def\W{\T{UHDM + DCID + WildMoir}\acute{\rm e}}\small\begin{array}{llrrrrrr}\hline\T{Model}&\T{Training data}&\T{PSNR}↑&\T{SSIM}↑&\T{LPIPS}↓&\T{MUSIQ}↑&\T{TOPIQ}↑&\T{Q-Align}↑\\\\\hline\T{SADT}&\T{TIP2018}&22.7996&0.7604&0.3108&30.1021&0.2412&3.5974\\\\\T{SADT}&\T{FHDMi}&24.1716&0.7931&0.2916&31.7764&0.2631&3.6671\\\\\T{SADT}&\T{LCDMoire}&23.3240&0.7902&0.2957&31.2316&0.2619&3.6211\\\\\T{SADT}&\T{UHDM + DCID}&27.1833&0.8762&0.2477&37.5005&0.3249&4.0713\\\\\T{SADT}&\W&\B{27.9014}&\B{0.8842}&\B{0.2310}&\B{38.4321}&\B{0.3331}&\B{4.1451}\\\\[3pt]\T{FaDeNet}&\T{PRISM}&23.8619&0.7784&0.2986&32.2143&0.2690&3.6851\\\\\T{FaDeNet}&\T{UHDM + DCID}&27.5212&0.8806&0.2492&38.1034&0.3227&4.0932\\\\\T{FaDeNet}&\W&\B{28.1730}&\B{0.8941}&\B{0.2355}&\B{39.2314}&\B{0.3296}&\B{4.1507}\\\\\hline\end{array}}

AuthorsOpenReview ↗

Response to Reviewer WXPG (Part 3/3)

References

[1] Heliang Zheng, Huan Yang, Jianlong Fu, Zheng-Jun Zha, and Jiebo Luo. Learning Conditional Knowledge Distillation for Degraded-Reference Image Quality Assessment. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp. 10242–10251, 2021.

[2] Shishun Tian, Tiantian Zeng, Zhengyu Zhang, Wenbin Zou, and Xia Li. Dual Residual-Guided Interactive Learning for the Quality Assessment of Enhanced Images. IEEE Transactions on Multimedia, 27, pp. 1637–1651, 2025.

[3] Weiqi Li, Xuanyu Zhang, Shijie Zhao, Yabin Zhang, Junlin Li, Li Zhang, and Jian Zhang. Q-Insight: Understanding Image Quality via Visual Reinforcement Learning. In Advances in Neural Information Processing Systems, vol. 38, pp. 41337–41362, 2025.

[4] Xuyi He, Yuhui Quan, Ruotao Xu, and Hui Ji. A Universal Scale-Adaptive Deformable Transformer for Image Restoration across Diverse Artifacts. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 12731–12741, 2025.

[5] Pengchao Wei and Xiaojie Guo. Fabric Image Demoiréing Benchmark from Synthesis to Restoration. In Computer Vision – ECCV 2026, pp. 528–545. Springer, 2026.

[6] Yujing Sun, Yizhou Yu, and Wenping Wang. Moiré Photo Restoration Using Multiresolution Convolutional Neural Networks. IEEE Transactions on Image Processing, 27(8), pp. 4160–4172, 2018.

[7] Bin He, Ce Wang, Boxin Shi, and Ling-Yu Duan. FHDe$^{2}$Net: Full High Definition Demoireing Network. European Conference on Computer Vision (ECCV), pp. 713–729, 2020.

[8] Shanxin Yuan, Radu Timofte, Gregory Slabaugh, and Aleš Leonardis. AIM 2019 Challenge on Image Demoireing: Dataset and Study. In 2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW), pp. 3526–3533, 2019.

AuthorsOpenReview ↗

Thanks & Friendly Reminder

We sincerely appreciate your thoughtful review and constructive feedback. As the discussion period ends tomorrow, we would be grateful to know whether our rebuttal has addressed your concerns. If anything remains unclear, we would be happy to provide further clarification.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Rebuttal Summary and Manuscript Update Summary

Rebuttal Summary

We thank all three reviewers for their helpful feedback. We have addressed the main concerns through additional experiments and clarifications:

  • Extended the evaluation to different architectures across CNNs (ESDNet, FaDeNet), Transformers (SADT), and diffusion models (SDXL, Qwen-Image-Edit), and compared against the publicly released checkpoints of SADT and FaDeNet.
  • Evaluated generalization on the external FHDMi benchmark and the complete UHDM and DCID test sets, and added an ESDNet baseline trained on WildMoiré alone to examine the effect of training data composition.
  • Validated A-FINE against CKDN, DRGQA, and Q-Insight on 1,000 sampled groups, and conducted independent human checks on 500 groups to measure residual moiré and content changes.
  • Added threshold ablations for τssim\tau_{\mathrm{ssim}} and τpost\tau_{\mathrm{post}}, filtering retention statistics, and clarifications on training parity, and test-set independence.

Update Summary

The revised manuscript (updates highlighted in blue) incorporates the following additions and clarifications:

  • Expanded quantitative evaluations to include FaDeNet and SADT (Abstract, Sec. 1, Sec. 2.1, Secs. 5.1–5.2, Sec. 6, Tables 3–5, Table J-K) and compared them with their publicly released checkpoints (Appendix C.4, Tables E–F).
  • Added evaluations on the external FHDMi test set (Sec. 5.2, Table 5), the complete UHDM and DCID test sets (Appendix D.3, Tables J–K), and WildMoiré-only vs. joint-training ablations on ESDNet (Appendix C.4, Table G).
  • Added human-alignment validation of A-FINE against CKDN, DRGQA, and Q-Insight (Sec. 3.4, Appendix C.1, Table B), human content-fidelity checks on 500 sampled groups (Appendix D.2, Table I).
  • Added sensitivity analyses for τssim\tau_{\mathrm{ssim}} and τpost\tau_{\mathrm{post}} (Sec. 5.1, Appendix C.2–C.3, Tables C–D).
  • Clarified test-set independence, and official training settings (Sec. 5.1, Appendix B.3).

Dear Area Chairs,

For ease of reviewing and evaluation, we provide a summary of the main strengths of our work, followed by an overview of the reviewers' key concerns and how we have addressed them through our rebuttal and manuscript revisions.

R1 denotes Reviewer WXPG, R2 denotes Reviewer TsZS, and R3 denotes Reviewer Tdcm.

Key Strengths

  1. Addressing a challenging real-world problem. Our work targets complex, large-scale moiré patterns that are insufficiently represented in existing datasets and for which clean ground truth is often difficult to obtain.

  2. A multi-prior generative supervision framework. We leverage complementary generative models and systematic quality control to construct WildMoiré, containing 6,832 training pairs, together with an independent 247-pair captured test set.

  3. Broad applicability across model architectures. Incorporating WildMoiré improves demoiréing performance across five CNN, Transformer, and diffusion-based models, without modifying their deployed architectures.

  4. Comprehensive experimental validation. Our experiments cover challenging real-world captures, public benchmarks, specialized demoiréing methods, and extensive ablations on supervision quality and training data composition.

Key Concerns and Our Responses

We have carefully addressed all the reviewers' key concerns through additional ablations, quantitative analyses, and comparisons. The corresponding updates are also incorporated into the revised manuscript, as summarized in our "Update Summary" comment.

  1. Reliability of A-FINE selection and generated supervision (R1 Q1, R1 W1, R2 W3, R3 Q3, R3 W1). We compared A-FINE with CKDN, DRGQA, and Q-Insight on 1,000 randomly sampled candidate groups. A-FINE achieved highest agreement with human selections, outperforming the alternatives. We further conducted an independent human evaluation on 500 groups, examining residual moiré, text alterations, changes in human details, and structural changes. These experiments provide additional evidence supporting the reliability of our supervision selection and filtering pipeline. Details are provided in R1 Q1, R1 W1, R2 W3, R3 Q3, R3 W1.

  2. Insufficient generalization evidence and public benchmark coverage (R1 Q2, R1 W2, R2 W1, R2 W4, R3 Q4). We added evaluations on the external FHDMi benchmark, involving unseen capturing devices, screens, and scenes. We also extended evaluations to the complete UHDM and DCID test sets rather than relying only on selected challenging cases. The results demonstrate performance improvements across different model architectures and test distributions. Details are provided in R1 Q2, R1 W2, R2 W1, R2 W4, R3 Q4.

  3. Lack of comparisons with recent specialized demoiréing methods (R1 Q3, R1 W3). We incorporated two recent specialized methods, SADT and FaDeNet, into our quantitative evaluations. Beyond retraining them with and without WildMoiré, we also compared their performance against their publicly released checkpoints. Both methods achieve significant improvements on our captured test set and the challenging cases from UHDM and DCID when WildMoiré is incorporated. Details are provided in R1 Q3, R1 W3.

  4. Effect of training data composition and data quantity–quality trade-off (R2 Q1, R2 W2). We added an ESDNet baseline trained exclusively on WildMoiré and compared it with models trained on UHDM+DCID and their combination with WildMoiré. The results show that WildMoiré alone provides effective supervision for complex moiré removal, while joint training achieves better overall performance across different test distributions. We also clarified the trade-off between supervision quality and retained dataset size. Details are provided in R2 Q1, R2 W2.

  5. Threshold selection, experimental settings, and dataset independence (R2 W1, R2 W3, R3 Q1, R3 Q2). We added additional ablations and analyses for the SSIM pre-filtering threshold and post-filtering threshold. We clarified that threshold settings were initially determined through human assessment during early dataset construction rather than optimization on the final test set. We also provided detailed training configurations, and explanations of the independence between training and test data. Details are provided in R2 W1, R2 W3, R3 Q1, R3 Q2.

We also expanded the discussion of the novelty and practical value of generative supervision for complex demoiréing (R2 W5) and clarified why improvements on challenging cases from UHDM and DCID are smaller than those on our more complex test set (R2 W4).

We sincerely thank all reviewers for their constructive feedback and valuable suggestions. We believe that the additional experiments and revisions have substantially strengthened the empirical evidence and clarified the contributions of our work.

We also thank the Area Chairs for their time and effort in reviewing and evaluating our submission.

Authors of Submission 40222

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.