acceptodds
Under review as a conference paper at ICLR 2027

Dual-Scale Condition-Shift Editing for Robust Real-World Image Demoir\'eing

13 views

Abstract

Robust demoiréing of screen-recaptured images remains challenging: moiré varies in frequency, color, and spatial support across display-camera pairs and capture conditions. Dedicated restoration networks trained on finite paired data struggle with this diversity, especially when low-frequency color banding and high-frequency stripes co-occur. We present a content-preserving editing framework that adapts the natural-image prior of a pretrained flow-matching editor to demoiréing. First, an endpoint-supervised training objective supplements velocity matching with pixel, perceptual, and frequency-domain restoration losses. Second, a dual-scale condition-shift strategy performs global correction at low resolution, then refines the image at full resolution by shifting the denoising condition from the first-stage output to the original input for detail recovery. We also introduce DailyMoiré130, a benchmark of challenging moiré in everyday screen photography, comprising 130 smartphone captures across diverse displays with low-frequency color banding, fine high-frequency stripes, and mixed-frequency interference. On the challenging DailyMoiré130 benchmark, our method achieves the lowest residual moiré area and receives 71.3% overall-quality preference in a blind Best-of-7 study with 31 participants. We also evaluate on paired FHDMi/UHDM, reducing residual moiré confidence by 47.1%/9.3% and increasing MUSIQ by 4.3%/2.0% over the best baseline for each metric.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Demoir\'e, Image Editing, Image Restoration, Low-level

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_NgStOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper addresses real‑world image demoiréing by leveraging a pre‑trained large generative model. Novel strategies are proposed for both training and inference, yielding improved subjective scores over several general‑purpose and task‑specific demoiréing methods, on both the self‑constructed no‑reference test set and two public benchmarks. Nevertheless, the method achieves slightly lower full‑reference fidelity (PSNR / SSIM) compared with pixel‑wise regression‑based approaches.

Critical strengths
  1. The newly introduced dataset consists of real‑world acquisitions from various mainstream smartphones. Further scaling and better alignment with real‑user data distribution would make it a valuable asset for the community.
  2. The integration of condition‑shift in Flow‑Matching image editing and dual‑scale inference is well‑tailored to moiré‑specific degradation. Driven by clear motivations, the method offers a targeted solution to real-world demoiréing challenges.
Critical weaknesses

Despite the authors’ emphasis on subjective evaluation as the key evidence for the superiority of their approach over competing methods, I still have concerns regarding the quality of the reported results:

  1. The proposed method yields generally inferior pixel‑wise metrics compared with competing baselines. While the authors argue that higher PSNR of other methods stems from "a trade‑off between pixel fidelity and artifact removal", it does not fully rule out potential detail degradation and undesired content distortion in model outputs.
  2. The baseline results reported on FHDMi and UHDM deviate from the numbers originally published in their respective papers. It appears that retraining these baselines under the same setup may degrade baseline performance, which could lead to unfair comparisons. Moreover, the paper lacks comparisons against Freqformer — another dual‑scale demoiréing approach — because its code and weights were unavailable at submission time. According to the original Freqformer paper, it achieves competitive metrics with substantially lower computational overhead.
  3. Visually, noticeable color shifts can be observed in most examples in Figure 6, and observable loss of fine‑grained details appears in Figure 5.
Questions for authors
  1. I'm wondering the motivation of establishing a newly‑built DailyMoiré130 test set: although the paper argues existing benchmarks suffer from limited conditions and scenes, I remain unconvinced that merely 130 images are sufficient to represent the data distribution encountered under realistic scenarios. I would appreciate further explanations on the sampling strategy and statistical properties of this dataset, to justify why such a small‑scale collection can reliably evaluate real‑world generalization.
  2. Why are the objective metrics (PSNR, etc.) for retrained baselines degrading? This is the most confusing problem for me. Ideally, comparisons should be performed using officially released baseline weights. If the authors choose to retrain all baselines from scratch, thorough analysis must be provided to pinpoint where this performance drop originates and rule out unfair comparison.
  3. The paper only ablates hard binary condition‑shifting with different r values (0,0.3,0.5,0.70, 0.3, 0.5, 0.7), but does not compare against the more classical strategy of continuously interpolating between the two conditioning latents over denoising timesteps, which is a well‑established scheduling paradigm within diffusion and Flow‑Matching image editing. It raises the question: can comparable performance be achieved with continuous weighted scheduling? If interpolation yields similar results, the necessity of the hard‑shift design would be weakened. I suggest the authors to add this ablation and corresponding analysis.
7 comments
AuthorsOpenReview ↗

Response to Reviewer NgSt (W1)

We thank Reviewer NgSt for the valuable feedback and detailed suggestions. Below we address each of the concerns and questions in detail.

W1 Pixel fidelity and content preservation

To examine whether lower PSNR reflects content degradation, we evaluate detail and structure preservation using dedicated human judgments and GT-referenced perceptual metrics, which support competitive content fidelity overall (Tables 1 and 2).

In the DailyMoiré130 Best-of-7 study, participants assess text, edges, and detail preservation separately from moiré suppression. Our method receives 68.5% of the 558 detail preservation selections from 31 participants (Table 1, Appendix E). On UHDM, TOPIQ-FR and AHIQ measure perceptual quality, while PieAPP measures perceptual error.

Our method achieves the highest TOPIQ-FR and the second best PieAPP and AHIQ among the compared methods (Table 2, Appendix G, manuscript Table 13). These complementary evaluations support content preservation beyond pixelwise agreement.

Table 1. Detail preservation on DailyMoiré130

Method or responseSelectionsPercentage
Ours38268.5%
ESDNet-L488.6%
Uni-control386.8%
MZNet162.9%
OSEDiff193.4%
MoCE-IR132.3%
AdaIR142.5%
No clear best285.0%
Total558100.0%

Table 2. GT referenced perceptual evaluation on UHDM

MethodTOPIQ-FR ↑PieAPP ↓AHIQ ↑
ESDNet-L0.6391.1370.433
Uni-control0.6411.0450.439
MZNet0.6640.9910.431
MoCE-IR0.6580.9640.433
OSEDiff0.5881.3100.284
Ours0.6700.9840.435
AuthorsOpenReview ↗

Response to Reviewer NgSt (W2, Q2)

W2 and Q2 Retraining and official checkpoints

We have supplemented the FHDMi and UHDM comparisons with official ESDNet-L checkpoints to address the concern about relying on retrained baselines (Table 4). We also conducted additional evaluations on DailyMoiré130 to assess cross-domain generalization under unconstrained real-world capture conditions (Tables 5 and 6).

The PSNR drop is consistent with a trade-off between single-domain specialization and multi-domain generalization. Official ESDNet-L checkpoints are trained on individual benchmarks, whereas our unified training shares model capacity across 12 heterogeneous streams, including eight public or synthetic moiré streams, two document streams, and two self-reconstruction streams (Appendix C.1).

Learning across these domains may reduce specialization to benchmark-specific pixel statistics. On unseen DailyMoiré130 captures, retrained ESDNet-L nevertheless achieves lower RMC and RMA than the official UHDM-trained checkpoint (Table 5), while CLIP-IQA+ is slightly lower.

In our human evaluation across all 130 images (Table 6), retrained ESDNet-L is preferred nearly twice as often as the official checkpoint across moiré removal (96 vs. 51), content preservation (91 vs. 46), and overall quality (94 vs. 51), achieving GSB scores of 1.41 to 1.43. This confirms that mixed training strengthens rather than degrades baseline performance on unconstrained real-world captures.

Crucially, our single unified model reduces RMC by 55.2% on FHDMi and 63.5% on UHDM against the official ESDNet-L checkpoints, with higher MUSIQ and CLIP-IQA+ on both datasets. TOPIQ is higher on FHDMi and tied on UHDM (Table 4). On DailyMoiré130, RMC and RMA are reduced by 67.3% and 66.6%, with higher CLIP-IQA+ (Table 5). Separately, the Best-of-7 study gives our method 71.3% overall quality preference versus 8.8% for retrained ESDNet-L.

These results support the generalization of our framework to challenging in-the-wild moiré. Aligned visual comparisons against official ESDNet-L are included in the supplementary material (Appendix L, manuscript Figure 12).

Table 4. Official and retrained comparison on paired benchmarks

DatasetMethodPSNR ↑MUSIQ ↑CLIP-IQA+ ↑TOPIQ ↑RMC ↓
FHDMiOfficial ESDNet-L24.8858.310.5420.4770.0143
FHDMiRetrained ESDNet-L20.6755.560.5230.4080.0175
FHDMiOurs21.0961.550.5850.5080.0064
UHDMOfficial ESDNet-L22.4265.580.6430.5850.0107
UHDMRetrained ESDNet-L21.1665.440.6250.5560.0139
UHDMOurs19.9967.430.6560.5850.0039

TOPIQ denotes the no-reference metric, and official ESDNet-L checkpoints are evaluated under the same test protocol as our main comparisons.

Table 5. DailyMoiré130 objective comparison

MethodRMC ↓RMA ↓CLIP-IQA+ ↑
Official ESDNet-L0.03520.16520.5517
Retrained ESDNet-L0.03400.15990.5478
Ours0.01150.05510.5928

Table 6. Retrained against official ESDNet-L

CriterionGood retrainedSameBad retrainedGSB ↑
Moiré removal96 (37.65%)108 (42.35%)51 (20.00%)1.43
Content preservation91 (35.69%)118 (46.27%)46 (18.04%)1.43
Overall quality94 (36.86%)110 (43.14%)51 (20.00%)1.41

This evaluation contains 255 image pair judgments from 15 participant IDs covering all 130 images. GSB = (G + 0.5S)/(B + 0.5S). Good and Bad refer to the retrained checkpoint. Values above 1 favor retraining.

Released ESDNet-L checkpoints are marked with a superscript in manuscript Table 2. For completeness, Table 7 lists the published results of MZNet-Large [1] and Freqformer [2] (whose code and weights remain unavailable) as reported in their original papers.

Table 7. Original paper results for literature reference

MethodDatasetPSNR ↑SSIM ↑LPIPS ↓Source
MZNet-LargeFHDMi26.120.86240.1042[1] Table 6
MZNet-LargeUHDM23.630.80960.2237[1] Table 6
FreqformerFHDMi25.260.85180.1253[2] Tables 1 and 2
FreqformerUHDM22.240.80210.2424[2] Tables 1 and 2

Results are cited directly from the original papers [1] and [2] under their respective benchmark-specific training and evaluation protocols.

References

[1] Lee et al. Moiré Zero. An Efficient and High Performance Neural Architecture for Moiré Removal. arXiv 2507.22407.

[2] Liu et al. Freqformer. Image Demoiréing Transformer via Effective Frequency Decomposition. arXiv 2505.19120.

AuthorsOpenReview ↗

Response to Reviewer NgSt (W3, Q1)

W3 Color shifts and fine details

We evaluate color accuracy and detail fidelity using GT-referenced metrics and human judgments to examine how local color shifts and fine-detail loss affect restoration quality (Tables 1 to 3). The results show that our method achieves GT color errors comparable to ESDNet-L and competitive detail preservation overall.

Regarding color shifts, we agree that some moiré-affected regions can appear lighter or whiter after restoration. This behavior can be understood in terms of the pretrained natural image prior. When severe corruption obscures the underlying colors, the pretrained prior may favor a clean, uniform appearance while removing the corruption.

Importantly, a color change relative to the corrupted input does not necessarily imply a larger error relative to the clean GT. Our UHDM chroma and hue errors are comparable to ESDNet-L (Table 3, Appendix G, manuscript Table 13).

Table 3. Color fidelity on UHDM

MethodMean absolute chroma error ↓Mean hue error ↓
ESDNet-L4.4542.776
Ours4.4222.643

Regarding fine details, we also acknowledge that fine details can be lost in individual cases. However, the apparent detail loss in Figure 5 should be considered together with the removal of severe moiré. We therefore provide dedicated detail preservation judgments and GT referenced perceptual evaluation to evaluate content fidelity alongside artifact removal.

Our method receives 68.5% of the detail preservation selections and performs competitively on TOPIQ-FR, PieAPP, and AHIQ (Tables 1 and 2). We discuss the restoration and fidelity tradeoff in Sec. 5 and have made the possibility of fine detail loss explicit.

Q1 Dataset motivation, sampling, and statistical properties

Our motivation is to evaluate challenging moiré on unconstrained in-the-wild displays, particularly those in malls and stations, where fine stripes and broad color bands often coexist. These conditions are underrepresented in the designated monitor settings of TIP2018, FHDMi, and UHDM. Yue et al. [1] similarly note that large public screens exhibit more visible moiré while lacking accessible clean ground truth.

We photographed displays operating in unconstrained real-world environments using five smartphone models at ordinary viewing positions and selected captures with visibly apparent moiré. The collection covers four scene categories (Tables 8 and 9). This curated sampling specifically targets hard, mixed-frequency real-world cases, complementing large-scale controlled monitor benchmarks (FHDMi and UHDM) while remaining compact enough for dense human evaluation across all 130 images (558 total judgments).

Table 8. Benchmark capture conditions

DatasetScreen typesCapture settingsCapture devicesImage resolution
TIP2018Designated monitors and laptop screensControlled screen recaptureiPhone 6, Samsung Galaxy S7 Edge, Sony Xperia Z5 Premium Dual415p to 968p
FHDMiDesignated monitors and laptop screensControlled screen recaptureiPhone 7, Samsung Galaxy S10, Huawei Mate 201080p
UHDMDesignated desktop monitorsControlled screen recaptureiPhone XR, iPhone 13, Redmi K30 Pro2160p
DailyMoiré130Commercial signage, public display walls, monitors, and laptopsUnconstrained in-the-wild settingsvivo X200s, iPhone 17, iPhone 16 Pro Max, HONOR 90 Pro, Huawei Enjoy 90 Pro Max864p to 4284p

Table 9. Overall quality preferences by scene

SceneImagesJudgmentsOursESDNet-LUni-controlMZNetOSEDiffMoCE-IRAdaIRNo clear best
Public or commercial indoor8635573.8%7.6%6.2%2.5%2.0%2.5%2.3%3.1%
Outdoor public188558.8%10.6%5.9%2.4%7.1%3.5%3.5%8.2%
Transportation135979.7%3.4%6.8%1.7%3.4%0.0%0.0%5.1%
Office or home135966.1%18.6%10.2%3.4%0.0%0.0%0.0%1.7%

Selection rates are percentages of all judgments within each scene, including no clear best. Scene proportions are 66.15%, 13.85%, 10.00%, and 10.00%.

We additionally stratified the Best-of-7 study by scene. The 31 participants provide 558 overall quality judgments covering all 130 images. Our method receives the highest preference rate in every scene category (Table 9). Moiré suppression and detail preservation show the same leading method across all four categories.

These results support generalization across the included in-the-wild capture conditions and complement paired evaluation on FHDMi and UHDM. We have included the complete scene statistics and all three dimensions of scene stratified human evaluation in the supplementary material.

References

[1] Yue et al. Unsupervised moiré pattern removal for recaptured screen images. Neurocomputing, 2021.

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

The paper adapts Qwen-Image-Edit with LoRA for image demoiréing, adding clean-image endpoint losses to a five-step unrolled flow-matching objective. Inference first corrects a downsampled image, then runs a full-resolution process that shifts its condition from the coarse restoration to the original input. Validation covers paired FHDMi/UHDM benchmarks, a new unpaired DailyMoiré130 collection, component ablations, and a randomized blind study. The method receives 71.3% of overall Best-of-7 votes on the real-image collection and reduces residual moiré on paired benchmarks, while sacrificing some conventional fidelity metrics. The full 4K recipe takes approximately 260 seconds on one H100 in FP32.

Critical strengths
  1. The evaluation triangulates real-world quality using complementary evidence. DailyMoiré130 includes challenging unpaired captures, and the randomized blind study covers all images with participant-cluster uncertainty and attention-check sensitivity. A separate detail-preservation question and paired content/color metrics reduce reliance on a learned moiré detector. This materially supports the practical restoration benefit.
  2. The method offers a concrete adaptation and inference recipe with useful controls. Endpoint-loss removal, condition-shift ratios, black-reference removal, and a native Qwen-Edit negative control are reported. The perceptual/artifact improvements over both feed-forward restorers and OSEDiff, together with acknowledged fidelity trade-offs, make the contribution more substantive than an unconstrained image-editing demonstration.
Critical weaknesses
  1. The main training comparison does not isolate why the adaptation works. Vanilla FM versus the proposed objective changes both single-timestep versus unrolled training and the endpoint losses, as Appendix B acknowledges. Removing individual endpoint losses is useful but does not supply unrolled velocity-only or matched single-step endpoint controls. Consequently, the claimed role of endpoint supervision relative to multi-step optimization remains uncertain.
  2. The incremental benefit and cost of the complete two-stage recipe are not adequately separated. In Table 4, the stage-2-only setting already reaches 20.02 PSNR, 0.135 LPIPS, and 67.42 MUSIQ on UHDM, versus 19.99, 0.134, and 67.43 for the full method; the clearer improvement is in detector-based residual moiré. Appendix C reports 260 seconds per 4K image, but no matched runtime/memory breakdown or direct human preference between one and two stages. These omissions matter to the practical justification for the central dual-scale component.
Questions for authors
  1. Under matched updates and data, how do unrolled velocity-only training and single-step endpoint supervision compare with the full objective? Which part explains the gain over Vanilla FM?
  2. What are the per-stage runtime and peak memory, and does a blinded comparison favor the full two-stage recipe over the stage-2-only setting? This would establish whether its extra cost is justified beyond the residual detector scores.
4 comments
AuthorsOpenReview ↗

Response to Reviewer XHCX (W1, Q1)

We thank Reviewer XHCX for the insightful comments and thorough evaluation. We now address the concerns and questions raised.

W1 and Q1 Training contributions

We have added controlled ablation experiments to distinguish the individual and combined contributions of endpoint supervision and five-step unrolling.

Under matched training data and 3k optimizer updates, we isolate single-step endpoint supervision, five-step unrolled velocity supervision, and their joint training configuration (Table 1, manuscript Table 3). Endpoint supervision is the more broadly effective isolated component, improving PSNR, SSIM, LPIPS, RMC, and RMA over Vanilla FM.

However, supervising decoded losses after a single step from high noise suffers from coarse one-step approximation errors, which limits its suppression ceiling. Conversely, unrolling velocity supervision alone improves PSNR and SSIM by aligning training with multi-step sampling, but degrades LPIPS, MUSIQ, and residual moiré because small velocity errors accumulate across unrolled steps without an image-space anchor. Joint training resolves both bottlenecks. Five-step unrolling produces a well-refined latent prediction, while the endpoint pixel, perceptual, and frequency losses anchor the end of this trajectory in image space to prevent multi-step drift.

Consequently, their combination reduces RMC to 0.0047 and raises PSNR to 20.02 dB, substantially outperforming either isolated control. The subsequent inference configurations show the additional contribution of Condition Shift and Black Image. We have revised the ablation discussion to clarify this mechanism.

Table 1. Component ablation on UHDM

VariantEndpointUnrollingCondition ShiftBlack ImagePSNR ↑SSIM ↑LPIPS ↓RMC ↓RMA ↓MUSIQ ↑
Vanilla FM××××17.510.6690.1950.01620.095365.29
+ Endpoint✓×××18.630.7040.1610.01020.057465.03
+ Unrolling×✓××18.930.7140.2080.01780.106860.26
Joint training✓✓××20.020.7930.1350.00470.025667.42
+ Condition Shift✓✓✓×20.000.7370.1890.00360.019566.81
Full✓✓✓✓19.990.7890.1340.00390.021267.43
AuthorsOpenReview ↗

Response to Reviewer XHCX (W2, Q2)

W2 and Q2 Incremental benefit and computational cost

We have added matched H200 runtime profiling and a blinded human study, showing that the two-stage recipe provides a clear perceptual gain in moiré removal for a moderate runtime increase (Tables 2 and 3). After one warm up run, timings are averaged over three runs. Full takes 181.34 seconds, adding 21.8% relative to Stage-2 only. Table 2 reports the stage breakdown and the peak allocated memory of each complete configuration. In Full, the Stage 2 time (165.90s vs. 148.88s) includes the overhead of upsampling and VAE-encoding the coarse Stage 1 output prior to full-resolution denoising.

Table 2. Runtime on H200 in BF16

ConfigurationStage 1 secondsStage 2 secondsTotal secondsPeak allocated GiB
Stage-2 onlyNot used148.88148.8864.81
Full15.45165.90181.3464.82

Timings use the same fixed 4K input and batch size 1. Peak memory is measured over each complete configuration.

Full reduces RMC and RMA by 17.0% and 17.2% on UHDM relative to Stage-2 only. On DailyMoiré130, RMC decreases from 0.0195 to 0.0115 and RMA from 0.1025 to 0.0551, corresponding to reductions of 41.0% and 46.2%. To further verify the perceptible benefit, we conducted an additional anonymous comparison covering all 130 images.

For overall quality, Full is rated better in 29.3% of judgments, comparable in 66.9%, and worse in 3.7%. Among the 375 image pair judgments, Full receives 122 favorable versus 10 unfavorable judgments for moiré removal and 110 versus 14 for overall quality. The GSB scores are 1.85 and 1.69, retaining comparable judgments. Content preservation is predominantly rated as comparable (Table 3). Our qualitative analysis identifies difficult color band and mixed frequency cases as particularly benefiting from the additional stage.

These results support a perceptible restoration benefit for the extra runtime. We have added the complete statistics in Appendix I, manuscript Table 18, and aligned visual comparisons in Appendix L, manuscript Figure 11.

Table 3. Full against Stage-2 only

CriterionGood FullSameBad FullGSB ↑
Moiré removal122 (32.53%)243 (64.80%)10 (2.67%)1.85
Content preservation41 (10.93%)309 (82.40%)25 (6.67%)1.09
Overall quality110 (29.33%)251 (66.93%)14 (3.73%)1.69

GSB = (G + 0.5S)/(B + 0.5S). G, S, and B denote Good, Same, and Bad judgments for Full. Values above 1 favor Full.

Reviewer_XHCXOpenReview ↗

Response

Thank you for the comprehensive rebuttal and additional experiments. The newly added additive ablations provide a much clearer understanding of the individual and joint contributions of endpoint supervision and multi-step unrolling, addressing my primary concern regarding the training objective. Overall, the authors have adequately addressed my main concerns, and I am inclined to raise my rating.

Dear Reviewer XHCX,

Thank you again for reviewing our rebuttal and additional experiments, and for raising your rating. Your insightful feedback has greatly helped us improve the clarity and completeness of our work.

Best regards,

Authors of Submission 41038

Reviewer_p7wvOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a image demoireing method by utilizing content-preserving editing framework that leverages a pretrained flow-matching image editor (i.e., Qwen-Edit) to remove complex moiré artifacts (e.g. low freq color banding and high freq stripes) from real-world screen captures. The authors introduce an endpoint-supervised training objective incorporating pixel, perceptual, and frequency-domain losses, alongside a dual-scale condition-shift inference strategy that performs global moiré correction at low resolution and refines details at high resolution. Further, the paper introduces DailyMoiré130, a new benchmark dataset comprising 130 real-world smartphone captures with mixed-frequency interference, to evaluate the model's performance.

Critical strengths
  • By combining the proposed dual-scale condition-shift strategy with the prior of a pretrained generative model, the method achieves visually dramatic moiré removal improvements compared to existing dedicated demoiréing networks.

  • The introduction of the DailyMoiré130 dataset, paired with a blind user study, provides a robust and practical evaluation setting for real-world moiré patterns where low-frequency color banding and high-frequency stripes co-occur.

Critical weaknesses
  • A visible color shift can be observed in the generated images compared to the original inputs, as seen in Figures 5 and 6. For instance, in the last example of Figure 6 (Laptop Display), the result from the proposed method appears noticeably whitened. I am curious about whether this is due to the inherent prior of the underlying Flow-Matching model.

  • First processing low-frequency components at a low resolution and then refining them is a widely used approach in existing demoiréing methods. It needs a deeper discussion on why this specific framework yields such dramatic effects to convince readers. Also, an analysis is needed on how well the Demoiréing task could be solved by utilizing Qwen-Edit alone, without the proposed dual-scale architecture.

  • The paper omits an analysis of the computational complexity (e.g., inference time, computing power required), making it difficult to assess the practicality of the method. Also, the paper needs justification regarding the specific new benefits the DailyMoiré130 dataset provides over existing moiré benchmarks.

Questions for authors
  • The performance improvement seems almost too dramatic compared to existing methods. I am concerned that the evaluation protocol and baselines were strictly fair.

  • In Tables 1 and 2, it would be highly beneficial for the analysis to include the performance metrics of the baseline 'Vanilla FM' for reference.

  • The ablation study in Table 3 is presented by subtracting ("w/o") one component at a time. Because the performance still seems vastly superior to Vanilla FM even when removing a single component, it is difficult to tell if other hidden factors are contributing to the performance gain. Adding elements one by one to the baseline, rather than subtracting, would make the individual contributions much easier to understand.

4 comments
AuthorsOpenReview ↗

Response to Reviewer p7wv (W1, W2)

We thank Reviewer p7wv for the constructive feedback and positive assessment of our work. We address each of the concerns and questions below.

W1 Color shifts and the pretrained prior

We agree that local color shifts relative to the corrupted input stem from the pretrained flow-matching prior, while our color errors relative to the clean GT remain comparable to feed-forward baselines (Table 1).

White interface backgrounds commonly appear clean and uniform, whereas moiré introduces abnormal bands and tones. When this corruption obscures the underlying colors, restoration may depend more strongly on the prior and produce a lighter or whiter appearance while removing the bands.

However, a color change relative to the corrupted input does not necessarily imply a larger error relative to the clean GT. On UHDM, our mean chroma and hue errors are comparable to those of ESDNet-L, as shown in Table 1 (Appendix G, manuscript Table 13).

Table 1. Color fidelity on UHDM

MethodMean absolute chroma error ↓Mean hue error ↓
ESDNet-L4.4542.776
Ours4.4222.643

W2 Motivation and value of dual scale inference

While coarse-to-fine architectures are common in existing demoiréing networks, they differ fundamentally from our framework in both prior knowledge and cross-scale interaction. First, conventional multi-scale networks learn feature extractors from scratch on finite paired data, so severe out-of-distribution color bands often survive the coarse branch. In contrast, our Stage 1 leverages a pretrained editor's broad natural-image prior on downsampled inputs (where high-frequency stripes fold into the lower band) paired with a black reference, reconstructing clean global structure and uniform colors even under heavy corruption. Second, rather than statically fusing coarse features with full-resolution inputs throughout the network, our Stage 2 exploits the temporal progression of flow-matching denoising by decoupling conditions across timesteps. Conditioning on the Stage 1 output during early high-noise steps locks in artifact-free global structure and color, whereas shifting to the original input during later low-noise steps recovers sharp text and edges without reintroducing broad color bands.

Regarding using Qwen-Edit without dual-scale inference, native zero-shot Qwen-Edit produces severe block artifacts (Appendix F, manuscript Figure 10), and standard single-step adaptation (Vanilla FM) leaves noticeable residual moiré (manuscript Tables 1 and 2). Our task-adapted single-scale model (Stage-2 only, trained with our full objective) still struggles when broad color bands and fine stripes coexist.

On all 130 DailyMoiré130 images, Full reduces RMC from 0.0195 to 0.0115 and RMA from 0.1025 to 0.0551 relative to Stage-2 only, corresponding to reductions of 41.0% and 46.2%. To further verify the perceptible benefit, we conducted an additional anonymous comparison with 375 image pair judgments from 16 participant IDs. For overall quality, Full is rated better in 29.3% of judgments, comparable in 66.9%, and worse in 3.7%. Among judgments indicating a difference, Full is preferred by 122 to 10 for moiré removal and 110 to 14 for overall quality. The complete GSB statistics retain all comparable judgments (Table 2). Our qualitative analysis indicates that the additional benefit is particularly relevant to difficult cases with broad color bands and mixed frequency patterns, while easier fine stripe cases are often comparable. We have added aligned visual comparisons to the supplementary material (Appendix L, manuscript Figure 11).

Table 2. Full against Stage-2 only

CriterionGood FullSameBad FullGSB ↑
Moiré removal122 (32.53%)243 (64.80%)10 (2.67%)1.85
Content preservation41 (10.93%)309 (82.40%)25 (6.67%)1.09
Overall quality110 (29.33%)251 (66.93%)14 (3.73%)1.69

GSB = (G + 0.5S)/(B + 0.5S). G, S, and B denote Good, Same, and Bad judgments for Full. Values above 1 favor Full.

AuthorsOpenReview ↗

Response to Reviewer p7wv (W3)

W3 Computational cost and the value of DailyMoiré130

Runtime was reported in Appendix C.3. We have updated this section with matched measurements on a single NVIDIA H200 GPU in BF16. For the same 4K input, Full takes 181.34 seconds compared with 148.88 seconds for Stage-2 only, adding 21.8% runtime with nearly identical peak allocated memory (Table 3). Quantization and distillation to fewer sampling steps are ongoing to improve practicality.

Table 3. Runtime on H200 in BF16

ConfigurationStage 1 secondsStage 2 secondsTotal secondsPeak allocated GiB
Stage-2 onlyNot used148.88148.8864.81
Full15.45165.90181.3464.82

Timings use the same fixed 4K input and batch size 1. Peak memory is measured over each complete configuration.

For the value of DailyMoiré130, compared with existing benchmarks such as TIP2018, FHDMi, and UHDM, which recapture designated monitors and laptops under controlled setups, DailyMoiré130 offers three main benefits for real-world evaluation (Tables 4 and 5).

First, it targets operating displays across four unconstrained real-world scene categories, particularly malls and stations, covering 96 public large screens or display walls and 80 challenging mixed-frequency cases where broad color bands and fine stripes coexist. Yue et al. [1] similarly note that large public screens exhibit more visible moiré while lacking accessible clean ground truth.

Second, its captures use recent smartphones such as vivo X200s, iPhone 17, and iPhone 16 Pro Max up to 4284p, reflecting artifacts that persist with modern mobile optics.

Third, stratifying our 558 Best-of-7 human judgments by scene reveals clear generalization differences across environments (Table 5). While ESDNet-L achieves 18.6% preference in office or home scenes that resemble conventional monitor benchmarks, its preference falls to between 3.4% and 10.6% in public indoor, outdoor, and transportation scenes, whereas our method consistently leads in all four categories.

Table 4. Benchmark capture conditions

DatasetScreen typesCapture settingsCapture devicesImage resolution
TIP2018Designated monitors and laptop screensControlled screen recaptureiPhone 6, Samsung Galaxy S7 Edge, Sony Xperia Z5 Premium Dual415p to 968p
FHDMiDesignated monitors and laptop screensControlled screen recaptureiPhone 7, Samsung Galaxy S10, Huawei Mate 201080p
UHDMDesignated desktop monitorsControlled screen recaptureiPhone XR, iPhone 13, Redmi K30 Pro2160p
DailyMoiré130Commercial signage, public display walls, monitors, and laptopsUnconstrained in-the-wild settingsvivo X200s, iPhone 17, iPhone 16 Pro Max, HONOR 90 Pro, Huawei Enjoy 90 Pro Max864p to 4284p

Table 5. Overall quality preferences by scene

SceneImagesJudgmentsOursESDNet-LUni-controlMZNetOSEDiffMoCE-IRAdaIRNo clear best
Public or commercial indoor8635573.8%7.6%6.2%2.5%2.0%2.5%2.3%3.1%
Outdoor public188558.8%10.6%5.9%2.4%7.1%3.5%3.5%8.2%
Transportation135979.7%3.4%6.8%1.7%3.4%0.0%0.0%5.1%
Office or home135966.1%18.6%10.2%3.4%0.0%0.0%0.0%1.7%

Selection rates are percentages of all judgments within each scene, including no clear best. Scene proportions are 66.15%, 13.85%, 10.00%, and 10.00%.

References

[1] Yue et al. Unsupervised moiré pattern removal for recaptured screen images. Neurocomputing, 2021.

AuthorsOpenReview ↗

Response to Reviewer p7wv (Q1)

Q1 Comparison fairness

We strictly control the baseline comparison and evaluation protocols to ensure fairness. As stated in Sec. 4.1, the retrained baselines use the same mixed training data while retaining their architectures and training objectives. They are trained for 50k optimizer updates with batch size 64. Our pretrained editor is fine tuned for 3k updates with batch size 96, as reported in Sec. 4.1 and Appendix C. Our suppression advantage also holds against the official ESDNet-L checkpoints, with lower RMC on FHDMi and UHDM (Table 6) and lower RMC and RMA on DailyMoiré130 (Table 7). All methods use the same test images and metric implementations.

Appendix E describes the anonymous candidate presentation, randomized and counterbalanced positions, and common viewing and crop settings used in the Best-of-7 study. We also provide aligned visual comparisons of our method with official and retrained ESDNet-L in Appendix L, manuscript Figure 12.

Table 6. Official and retrained comparison on paired benchmarks

DatasetMethodPSNR ↑MUSIQ ↑CLIP-IQA+ ↑TOPIQ ↑RMC ↓
FHDMiOfficial ESDNet-L24.8858.310.5420.4770.0143
FHDMiRetrained ESDNet-L20.6755.560.5230.4080.0175
FHDMiOurs21.0961.550.5850.5080.0064
UHDMOfficial ESDNet-L22.4265.580.6430.5850.0107
UHDMRetrained ESDNet-L21.1665.440.6250.5560.0139
UHDMOurs19.9967.430.6560.5850.0039

TOPIQ denotes the no-reference metric, and official ESDNet-L checkpoints are evaluated under the same test protocol as our main comparisons.

Table 7. DailyMoiré130 objective comparison

MethodRMC ↓RMA ↓CLIP-IQA+ ↑
Official ESDNet-L0.03520.16520.5517
Retrained ESDNet-L0.03400.15990.5478
Ours0.01150.05510.5928
AuthorsOpenReview ↗

Response to Reviewer p7wv (Q2, Q3)

Q2 Vanilla FM in the main comparisons

We have added Vanilla FM to manuscript Tables 1 and 2 to directly compare standard FM adaptation with our complete method. Its paired benchmark and DailyMoiré130 results are reported in Table 8.

Table 8. Vanilla FM supplementary reevaluation

DatasetPSNR ↑MUSIQ ↑CLIP-IQA+ ↑TOPIQ ↑RMC ↓
FHDMi18.5857.530.5420.4750.0188
UHDM17.5165.290.6130.5320.0162
DatasetCLIP-IQA+ ↑RMC ↓RMA ↓
DailyMoiré1300.54930.02070.1055

UHDM PSNR, MUSIQ, and RMC retain the submitted manuscript values, consistent with Table 9. CLIP-IQA+ and TOPIQ are additional supplementary measurements.

Q3 Additive component ablations

We have restructured manuscript Table 3 and added corresponding ablations to explicitly show the configurations bridging Vanilla FM and Full (Table 9). Starting from Vanilla FM with PSNR 17.51 and RMC 0.0162, single-step endpoint supervision alone raises PSNR to 18.63 and lowers RMC to 0.0102, though its gain is bounded by coarse one-step decoding approximations at high noise levels. Five-step unrolling alone raises PSNR to 18.93, yet degrades LPIPS, MUSIQ, and RMC because unrolled latent velocities accumulate drift without an image-space anchor. Combining both in Joint training (used as the Stage-2 only reference) resolves these limitations by applying pixel, perceptual, and frequency supervision to a refined multi-step endpoint, reaching PSNR 20.02 and lowering RMC to 0.0047. Adding Condition Shift reduces RMC to 0.0036. Full then improves SSIM, LPIPS, and MUSIQ relative to the configuration without Black Image, with a small suppression tradeoff. This presentation distinguishes the isolated training controls, their joint configuration, and subsequent inference components. We have revised manuscript Table 3 and the ablation discussion accordingly.

Table 9. Component ablation on UHDM

VariantEndpointUnrollingCondition ShiftBlack ImagePSNR ↑SSIM ↑LPIPS ↓RMC ↓RMA ↓MUSIQ ↑
Vanilla FM××××17.510.6690.1950.01620.095365.29
+ Endpoint✓×××18.630.7040.1610.01020.057465.03
+ Unrolling×✓××18.930.7140.2080.01780.106860.26
Joint training✓✓××20.020.7930.1350.00470.025667.42
+ Condition Shift✓✓✓×20.000.7370.1890.00360.019566.81
Full✓✓✓✓19.990.7890.1340.00390.021267.43

Comments on the paper

2 comments
AuthorsOpenReview ↗

Revision Summary

We thank the reviewers for their valuable comments and suggestions. We have incorporated the following changes in the revised paper.

  1. Added Vanilla FM to the main comparisons (Reviewer p7wv). See manuscript Tables 1 and 2.
  2. Expanded component ablations to six configurations (Reviewers p7wv and XHCX). See manuscript Table 3.
  3. Clarified shared training data and training budgets (Reviewers p7wv and NgSt). See Sec. 4.1 and Appendix Sec. C.1.
  4. Added matched H200 runtime and memory measurements (Reviewers p7wv and XHCX). See manuscript Table 7 and Appendix Sec. C.3.
  5. Added Full versus Stage-2 only human evaluation (Reviewers p7wv and XHCX). See Appendix Sec. I.
  6. Added released ESDNet-L comparisons on FHDMi and UHDM (Reviewers p7wv and NgSt). See manuscript Table 2 and Appendix Sec. J.
  7. Added released and retrained ESDNet-L comparisons on DailyMoiré130 (Reviewers p7wv and NgSt). See manuscript Table 20.
  8. Added human evaluation of retrained versus released ESDNet-L (Reviewer NgSt). See manuscript Table 21.
  9. Added benchmark capture comparisons and dataset statistics (Reviewers p7wv and NgSt). See Appendix Secs. D and H.
  10. Added human preference analysis across four scene categories (Reviewer NgSt). See Appendix Sec. H.
  11. Added Linear and Cosine scheduling ablations and sampling weights (Reviewer NgSt). See Appendix Sec. K.
  12. Added aligned comparisons against Stage-2 only and ESDNet-L checkpoints (Reviewers p7wv, XHCX, and NgSt). See manuscript Figures 11 and 12.
  13. Added MZNet-Large and Freqformer literature results separately from checkpoint evaluations (Reviewer NgSt). See manuscript Table 22.
  14. Clarified possible color shifts and fine detail loss (Reviewers p7wv and NgSt). See Sec. 5.
  15. Updated the paper title and refined the framing of dual-scale detail-reanchored inference as a task-specific design tailored to demoiréing (Reviewer NgSt). See Title, Abstract, Sec. 1, Sec. 3.3, and Appendix Sec. K.
  16. Clarified the targeted hard-case benchmark positioning of DailyMoiré130 alongside paired FHDMi and UHDM datasets and noted ongoing multi-device benchmark expansion (Reviewer NgSt). See Sec. 5 and Appendix Sec. H.

We have addressed the reviewers' concerns below and would be happy to clarify any further questions during the discussion.

Authors of Submission 41038

AuthorsOpenReview ↗

Summary for Area Chairs (Updated at End of Discussion)

Dear Area Chairs,

For ease of reviewing and evaluation, we provide a concise summary of the key strengths highlighted by the reviewers, followed by an overview of the concerns raised and where they are addressed in our rebuttal.

R1 -> Reviewer p7wv

R2 -> Reviewer XHCX

R3 -> Reviewer NgSt

[Update at the End of Discussion]

Thanks to the constructive engagement of the reviewers and Area Chairs, our manuscript has been substantially strengthened and most concerns have been addressed. Specifically, Reviewer XHCX (R2) confirmed that our step-by-step component ablations and two-stage evaluations addressed their main concerns and raised their rating to join Reviewer p7wv (R1) in recommending acceptance, while Reviewer NgSt (R3) confirmed that our official-checkpoint comparisons, scheduling ablations, color/detail metrics, and stratified dataset analyses improved the paper and resolved most of their original concerns. Regarding Reviewer NgSt's two remaining follow-up notes on contribution phrasing and benchmark scope, we have provided final clarifications and details of our ongoing 500+ multi-device open-source expansion in their thread, alongside synchronized manuscript updates (items 15 and 16 in "Revision Summary").

Key strengths

  1. Well-tailored integration of condition-shift editing and dual-scale inference for real-world mixed-frequency moiré (R1, R2, R3).
  2. Visually dramatic moiré removal improvements compared with existing dedicated demoiréing networks (R1, R2).
  3. Perceptual and artifact improvements over both feed-forward restorers and diffusion restoration baselines (R2, R3).
  4. Concrete adaptation and inference recipe supported by useful controls (R2).
  5. Introduction of the DailyMoiré130 dataset capturing challenging real-world moiré across mainstream smartphones (R1, R2, R3).
  6. Robust evaluation pairing DailyMoiré130 with a randomized blind Best-of-7 study and uncertainty analysis (R1, R2).
  7. Dedicated detail-preservation judgments and paired content and color metrics that reduce reliance on a learned detector (R2).

Key Concerns We list below the primary concerns shared across reviewers. We have taken care to address all these concerns in our rebuttal and in the revised version of the manuscript. The revision details can be found in our global comment titled "Revision Summary". In the points below, W refers to "Weakness" and Q refers to "Question".

  1. Unclear how endpoint supervision and five-step unrolling individually contribute over Vanilla FM (R1 Q2, R1 Q3, R2 W1, R2 Q1). We added Vanilla FM to the main benchmark comparisons and expanded the component ablation to six additive configurations isolating single-step endpoint supervision and five-step velocity unrolling under matched data and updates. The results show that five-step unrolling provides a refined multi-step trajectory while endpoint supervision anchors the decoded image, and their combination achieves substantially stronger moiré suppression and fidelity than either isolated control. Detailed explanations can be found in R1 Q2, R1 Q3, R2 W1, R2 Q1.

  2. Incremental benefit and computational cost of the complete two-stage recipe over single-stage Qwen-Edit (R1 W2, R1 W3, R2 W2, R2 Q2). We provide matched H200 runtime and memory breakdowns along with a new blinded human comparison between Full and Stage-2 only across all 130 DailyMoiré130 images. These results show that the low-resolution stage adds moderate runtime with nearly identical peak memory, while delivering substantial reductions in residual moiré and strong human preference in both moiré removal and overall quality. More details can be found in R1 W2, R1 W3, R2 W2, R2 Q2.

  3. Comparison fairness of retrained baselines versus officially released checkpoints (R1 Q1, R3 W2, R3 Q2). We provide additional evaluations using officially released ESDNet-L checkpoints on FHDMi, UHDM, and DailyMoiré130, alongside a new human comparison between retrained and official ESDNet-L. The results confirm that our method maintains a clear moiré suppression and perceptual advantage over official benchmark-specialized checkpoints, while unified mixed training improves baseline generalization on unseen real-world captures. Details can be found in R1 Q1, R3 W2, R3 Q2.

In response to the remaining questions by the reviewers, we have provided further comparisons, quantitative analyses, ablations, and clarifications, which we summarize below.

  1. Additional ablations (R1 Q3, R2 W1, R2 Q1, R3 Q3).
  2. Additional comparisons (R1 W2, R1 Q1, R1 Q2, R2 W2, R2 Q2, R3 W2, R3 Q2).
  3. Additional quantitative results (R1 W2, R1 W3, R1 Q1, R1 Q2, R2 W2, R2 Q2, R3 W2, R3 Q1, R3 Q2, R3 Q3).
  4. Additional clarifications (R1 W1, R1 W2, R1 W3, R1 Q1, R3 W1, R3 W2, R3 W3, R3 Q1, R3 Q2, R3 Q3).

We thank the reviewers for their valuable comments and feedback. We also thank the Area Chairs for their time and effort invested in reviewing and evaluating our work.

Authors of Submission 41038

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.