- 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
- How were the post-filtering thresholds and training configurations determined before inspecting the final test results, and which independent validation data were used?
- 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.
- 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?
- Do evaluations on complete existing benchmarks, or evaluations with devices and screen content held out, still support the reported restoration improvements?
4 comments
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.
Table R3. Results on the complete DCID test set.
Table R4. Results on FHDMi test set.
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}}
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.
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.
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 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 10−4 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}}