acceptodds
Under review as a conference paper at ICLR 2027

Combined Flicker-banding and Moire Removal for Screen-Captured Images

11 views

Abstract

Capturing display screens with mobile devices has become increasingly common, yet the resulting images often suffer from severe degradations caused by the coexistence of moiré patterns and flicker-banding, leading to significant visual quality degradation. Due to the strong coupling of these two artifacts in real imaging processes, existing methods designed for single degradations fail to generalize to such compound scenarios. In this paper, we present the first systematic study on joint removal of moiré patterns and flicker-banding in screen-captured images, and propose a unified restoration framework, named CLEAR. To support this task, we construct a large-scale dataset containing both moiré patterns and flicker-banding, and introduce an ISP-based flicker simulation pipeline to stabilize model training and expand the degradation distribution. Furthermore, we design a frequency-domain decomposition and re-composition module together with a trajectory alignment loss to enhance the modeling of compound artifacts. Extensive experiments demonstrate that the proposed method consistently outperforms existing image restoration approaches across multiple evaluation metrics, validating its effectiveness in complex real-world scenarios.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Moiré, Flicker banding, Diffusion

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_PMM8OpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the coexistence of flicker-banding and moiré artifacts in screen-captured images, examines their joint removal, and proposes a unified restoration framework, CLEAR. The authors construct MIRAGE, a dataset containing 3,000 pairs of real captured images covering three degradation categories: flicker-banding only, moiré only, and compound degradations. Built on the PiSA-SR backbone, the method combines ISP-based RAW-domain banding simulation, a frequency-domain segmentation and weighted recomposition module (FS), and a multi-level wavelet loss (MWL). The experimental results show that CLEAR outperforms the compared methods on all six metrics. Comparisons with cascaded single-task models and component ablations further support the need for joint processing and the effectiveness of the individual components.

Critical strengths
  1. A practically relevant problem formulation and dataset contribution. The coexistence of moiré and flicker-banding is an important restoration scenario in screen photography. The paper compares cascaded single-task models in both processing orders, providing direct evidence for the need for joint processing. MIRAGE covers three degradation categories, six content categories, and five smartphones, giving the work value as a dataset contribution in addition to the proposed model.
  2. Targeted modeling of compound degradations and innovative component designs are well supported by ablation results. I find CLEAR’s technical contributions worthy of recognition. ISP-based RAW-domain banding simulation incorporates considerations of the imaging process into the method design. FS uses cumulative spectral-energy quantiles to determine segment boundaries, constructs frequency bands with smooth Butterworth masks, and combines task priors with spectral statistics to initialize recomposition weights. MWL constrains residual artifacts in the output through multi-scale, orientation-sensitive subband supervision, complementing frequency-domain processing of the input. The design and complementary roles of these components constitute a meaningful methodological contribution to joint banding and moiré removal.
Critical weaknesses
  1. The parameter definitions and configuration of the core method are incomplete. FS is one of the paper’s main technical contributions, yet the paper does not explain the choice of six frequency segments, specify the data over which the spectral statistics are computed, or clarify whether the segment boundaries and recomposition weights are fixed, computed per image, or trainable. A weight initialization formula alone does not specify how the module actually operates. The description of MWL also omits the decomposition depth used in training, the subband weights, and the weighting relative to the other loss terms. A four-level decomposition visualization is not a substitute for the training configuration.
  2. The evaluation of practical applicability is incomplete. The paper does not report inference time, peak GPU memory usage, or the strategy for processing high-resolution images, nor does it analyze representative failure cases. Consequently, the reported results are insufficient to assess the method’s deployment costs and reliability on challenging inputs.
Questions for authors
  1. FS parameter determination: Why are six frequency segments used? Are the spectral statistics used to determine the boundaries computed over the entire training set or for each input? After initialization, are the segment boundaries and recomposition weights fixed, recomputed per image, or optimized during training? Please specify the actual procedure, key parameter values, and the rationale for their selection; the current description is insufficient to reproduce the module.
  2. MWL training configuration: How many wavelet decomposition levels are used during training? What are the weights of the LL, LH, HL, and HH subbands, and the relative weights of MWL, pixel loss, and perceptual loss? How were these values determined? The paper provides only the loss formulation and decomposition visualization, without the complete configuration needed to specify the actual training objective.
  3. Deployment costs: What are CLEAR’s per-image inference time and peak GPU memory usage under specified hardware, input resolution, and batch size? Are original high-resolution photographs processed as complete images or in tiles? The paper emphasizes practical applicability but does not report the corresponding resource requirements. Please provide these measurements and the inference strategy actually used.
  4. Failure cases and operating limits: What are CLEAR’s main failure modes? Please provide representative failure cases, analyze whether residual banding, moiré, or detail loss occurs, and identify the method’s operating limits.
6 comments
AuthorsOpenReview ↗

Response to Reviewer PMM8 (Part 1)

Q3-1: FS configuration, spectral statistics, parameter updates, and the choice of six frequency segments.

A3-1:

Thank you for asking us to clarify the FS settings. (1) Statistics and boundaries. We average smoothed grayscale log-power profiles across the training dataset. After reducing the DC peak, we apply radial compensation. We then divide the cumulative energy into six equal parts to set the band boundaries. (2) Weights and updates. We first compute the energy of each frequency band using the same training images. We apply log(1 + E) to each energy value E, then use min-max normalization to obtain normalized energies. We multiply these by the negative energy scale, apply the exponential function, and divide by the sum across bands to obtain statistical weights. We combine these weights with preset prior weights based on the frequency characteristics of flicker-banding and moiré. The resulting weights control each band's contribution when the frequency bands are recombined. The boundaries and weights remain fixed during training and inference. They are neither recomputed for each input nor learned by the optimizer. (3) Key settings. We use six bands, Butterworth order 2, energy scale 2.2, and prior balance 0.45. Appendix A.7 and Table 12 provide the remaining values and implementation details. (4) Choice of six bands. We chose six bands empirically based on our observations of the spectra of real images with flicker-banding and moiré. This number was not derived from an exact calculation. Our design mainly needs to distinguish low-, mid-, and high-frequency regions. We observed flicker-banding mainly in the low and middle frequencies. Within the high frequencies, we further distinguish the range where moiré appears from the frequencies above it. Six bands provide enough separation for this purpose. Too few bands can group degradation components with normal image structures, making them harder to distinguish. Too many bands can make the statistically estimated weights more sensitive to random variation. We therefore use six bands as an empirical balance between sufficient frequency separation and stable weight estimation. Table 4 shows the benefit of FS with this setting.

Table 4. Effect of FS and MWL.

Modelms-SSIM↑LPIPS↓DISTS↓
Baseline0.79090.31010.2211
MWL0.79790.29010.2087
FS0.80820.29580.2158
MWL + FS0.81580.28360.2011

Q3-2: MWL decomposition depth, subband weights, loss weights, and their selection.

A3-2:

Thank you for pointing out these missing details. (1) Decomposition depth. We use four levels of Haar decomposition during training, consistent with the figure in the paper. At each level, we decompose the LL output from the previous level. (2) Subband weights. We assign weight 1 to each LL, LH, HL, and HH subband at all four levels. LL supervises coarse luminance, while LH, HL, and HH supervise directional details. We average squared errors within each subband over the batch, channels, and spatial positions, then sum the sixteen subband losses. (3) Overall loss. The pixel MSE, perceptual LPIPS, and MWL weights are 1, 2, and 0.1, respectively. Pixel and MWL losses are used from the start; perceptual loss starts after 5,000 steps. (4) How we chose the settings. We chose four decomposition levels and equal subband weights to supervise flicker-banding across multiple scales and directions. We observed that four levels clearly reveal the directional features of flicker-banding. This already meets our design goal, so we consider further decomposition unnecessary for this purpose. We regard the features at each level and in each subband as equally important and therefore assign them equal weights. These choices follow our empirical observations and design goals, rather than exact calculations or a systematic parameter search. The overall MWL weight of 0.1 was also set empirically. Table 4 shows that MWL improves all three reported metrics with these settings. We will explore more refined designs and systematic parameter selection in future work. Appendix A.7 provides the implementation details, loss weights, and training schedule.

Table 4. Effect of FS and MWL.

Modelms-SSIM↑LPIPS↓DISTS↓
Baseline0.79090.31010.2211
MWL0.79790.29010.2087
FS0.80820.29580.2158
MWL + FS0.81580.28360.2011
AuthorsOpenReview ↗

Response to Reviewer PMM8 (Part 2)

Q3-3: Inference time, peak GPU memory, and the high-resolution inference strategy.

A3-3:

Thank you for asking us to clarify the deployment costs and high-resolution inference strategy. (1) Test setup and resource requirements. We evaluate CLEAR on an RTX 4090 48GB using FP16 and batch size 1. For a complete image at 4096 × 3072, inference takes 173.18 s and uses 46.83 GiB of GPU memory. Table 9 also reports the parameter count and FLOPs. (2) High-resolution inference. To process images at their native resolution, we use tiled inference in latent space. This follows the same strategy used in diffusion-based image restoration methods such as OSEDiff [2]. We process overlapping 64 × 64 latent tiles with an overlap of 32 latent pixels. We then combine the predictions using Gaussian weights in the overlapping regions to reduce tile-boundary artifacts and reconstruct the full-resolution image. The reported runtime and memory usage correspond to this tiled inference setup. Appendix A.4 provides the settings and resource measurements.

Table 9. CLEAR complexity and inference efficiency at 4096 × 3072 on RTX 4090 48GB.

ModelParams (M)FLOPs (T)Time (s)GPU (GiB)
CLEAR1292.00478.98173.1846.83

[2] Rongyuan Wu, Lingchen Sun, Zhiyuan Ma, and Lei Zhang. One-Step Effective Diffusion Network for Real-World Image Super-Resolution. In NeurIPS, 2024.


Q3-4: Representative failure cases, residual artifacts, detail loss, and operating limits.

A3-4:

Thank you for asking about failure cases and the limits of our method. (1) Observed errors. Appendix A.9 and Figure 13 show two examples under severe banding. The first retains band boundaries on the road, while the second has blurred details. (2) Causes and limitations. These examples illustrate two limitations. First, extremely dark bands can erase visible scene structures, leaving too little information for accurate correction and causing residual banding. Second, under poor capture conditions, severe banding and moiré may coexist with blur and other degradations. Even after the target artifacts are removed, these other degradations can leave the output quality low. Recovering blurred details is beyond the scope of this paper, which focuses on banding and moiré removal. We nevertheless recognize these outputs as practical failure cases and will explore how to address such additional degradations in future work.


AuthorsOpenReview ↗

Call for further discussion with Reviewer PMM8

Dear Reviewer PMM8,

Thank you for your time and thoughtful feedback. In our rebuttal, we have provided the following details and evidence in response to your concerns:

  1. FS configuration and design rationale, including spectral statistics, frequency boundaries, recomposition weights, and the choice of six bands.
  2. The complete MWL training configuration, including decomposition depth, subband weights, loss weights, and the rationale for these choices.
  3. Inference costs and high-resolution processing details, including runtime, GPU memory usage, and the tiled inference strategy.
  4. Representative failure cases and limitations, with an analysis of residual artifacts and detail loss.

As we approach the midpoint of the rebuttal discussion period, we would appreciate your feedback on whether these responses address your concerns. If you have any remaining questions or suggestions, we would welcome the opportunity to discuss them.

Thank you again for your time and consideration.

Best regards,

Authors

Reviewer_cGgAOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

A method to remove the Moiré patterns

Critical strengths

I was in this field for many years. I do not find very obvious strengths.

Critical weaknesses

Actually, I also did quite a lot of research on image restoration. However, I do not think this is still a good research topic. For example, no matter what designs you use, the LLM with image generation capability can produce much better results based on the specific prompts. I tried some samples from this paper in GPT. Then, I obtained very good results that are able to outperform the results shown in this paper.

Maybe in the future, it is better to call image restoration "Conditional image generation".

I suggest the authors consider some other research topics.

Questions for authors

No questions. I have to keep the comments as simple as possible to show that this is written by me. I never use any AI for the review.

13 comments
AuthorsOpenReview ↗

Response to Reviewer cGgA (Part 1)

Q2-1: Comparison with GPT-based image generation and editing.

A2-1:

Thank you for testing GPT on our examples. (1) GPT's performance does not invalidate this work. Your observation establishes GPT as a promising solution and a useful comparison method. It does not justify dismissing our dataset, simulation pipeline, and restoration method. Strong performance from a large, closed-source general-purpose system is understandable given its large-scale pretraining and extensive engineering. Comparing it directly with an SD 2.1-based method under unmatched training and deployment resources is inherently unequal. Requiring CLEAR to surpass GPT in every aspect is not a valid criterion for judging these contributions. (2) Systematic comparison. GPT can improve sharpness beyond our target degradations, but both quantitative results and visual comparisons show that CLEAR achieves clearly better content fidelity. Fidelity to the original content is a key criterion in image reconstruction. CLEAR also retains native resolution and supports local inference with lower observed latency under the tested conditions. GPT's limited output resolution and cloud dependence restrict native-resolution processing and on-device deployment, both key requirements for image reconstruction applications. Appendix A.10, Tables 14–15, and Figure 14 provide the full comparisons and settings. (3) Our view of GPT's results. We recognize the high visual quality of GPT's outputs. Its success offers a promising direction for future model design, but does not diminish the value of CLEAR. Recent work supports this view: GGT-100K [3] uses image editing models to create training targets for restoration models, while Ren et al. [4] show that high visual quality can still come with identity errors in restored images. Reliable restoration with image editing models therefore remains an open research problem. As the first systematic study of joint moiré and flicker-banding removal in screen-captured images, our choice of an established restoration approach as the baseline is reasonable. Better-looking GPT outputs alone do not invalidate this choice or our contributions. We recognize GPT's strong visual results, but they do not mean that our work lacks value. We hope the additional evidence helps you reconsider the value of our work. Thank you for your time and consideration.

Table 14. Image editing comparisons on original 4K inputs.

MethodSSIM↑ms-SSIM↑LPIPS↓DISTS↓FSIM↑GMSD↓
GPT0.60260.68360.44500.16170.84250.2133
Step1X0.66300.68200.45230.20880.79030.1841
CLEAR0.79420.84160.28290.11060.90960.1371

Table 15. Model complexity and efficiency at 1536 × 1024.

MethodParams (B)FLOPs (T)Latency (s)Execution
GPT––22.08Hosted service
Step1X20.80415979.30157.65Local RTX 4090
CLEAR1.292353.9814.64Local RTX 4090

[3] Xiangtao Kong, Jixin Zhao, Lingchen Sun, Rongyuan Wu, and Lei Zhang. GGT-100K: Generative Ground Truth for Generalizable Real-World Image Restoration. arXiv preprint arXiv:2605.31039, 2026.

[4] Mengwei Ren, Xuaner Zhang, and Zhihao Xia. Persistent Identity Preservation in Generative Image Models: A Benchmark and Evaluation System. arXiv preprint arXiv:2609.04151, 2026.

AuthorsOpenReview ↗

Response to Reviewer cGgA (Part 2)

Q2-2: Research contributions, the value of image restoration, and conditional image generation.

A2-2:

Thank you for your perspective. (1) Clear and meaningful contributions. Our contributions are clear and meaningful. GPT's strong results on selected examples do not invalidate them. We provide a paired dataset, an ISP-based simulation pipeline, and a restoration method for joint flicker-banding and moiré removal (Sections 3–5). Reviewer Vrtp recognizes the practical problem and the dataset's potential value. Reviewer PMM8 also recognizes the dataset contribution. (2) Image restoration remains valuable. Image editing models can produce high-quality images, but they do not always preserve the original content. Ren et al. [4] report that restored images can look good while changing the subject's identity. GGT-100K [3] uses generative models to provide training data that improve existing restoration models. This supports the continued value of restoration architectures and research into their design. This research is still developing. Producing high-quality images while faithfully recovering the original content remains a challenge. Progress in image editing therefore does not mean that image restoration is solved or no longer worth studying. (3) Practical value and efficiency. Image editing models like GPT rely on cloud infrastructure and are unavailable for direct deployment on mobile devices. CLEAR supports local inference on consumer hardware and can also be deployed on edge devices such as mobile phones. Efficient local processing and faithful native-resolution output are important needs in screen-image restoration. CLEAR addresses these needs, and our quantitative results, visual comparisons, and ablations demonstrate its effectiveness. Our work makes substantial contributions, and recent progress in image editing further supports the value of this research. GPT's strong results on a few examples do not invalidate these contributions. We would appreciate it if you could reconsider the contributions of our work in light of these clarifications and additional results. Thank you for your time and careful consideration.

  1. I do not agree that comparing with GPT is unfair. When working on a problem with clear practical value, one cannot ignore the competing solutions already available on the market. Treating closed-source models as an unfair comparison is, in itself, a way of avoiding the issue. You may argue that your method has advantages in speed, but that does not justify not comparing against it.

  2. Are you truly confident that your own model does not alter the original scene content? Unless you can demonstrate experimentally that your results preserve the original content, this remains my concern.

  3. The comparison with GPT and Step1X is limited to your own test set, and the setup is unclear: the GPT version is not specified, their low-resolution outputs are upsampled before evaluation, and no quantitative content-preservation metric is reported for them. The comparison is therefore not convincing.

  4. Restoration results can fluctuate considerably depending on which data are added to or removed from the training and test sets. It is therefore difficult to tell whether the reported improvements come from your method design or from your choice of data.

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

The paper studies joint flicker-banding and moiré removal in screen-captured images. It introduces MIRAGE, a dataset of 3,000 paired captures (Section 3.2), and an ISP-based pipeline for synthesizing banding on moiré images. CLEAR adapts PiSA-SR with LoRA, frequency segmentation, and a multi-level wavelet loss. Experiments report improvements over seven retrained baselines and two debanding–demoiréing cascades on cropped MIRAGE test images (Tables 2 and 5).

Critical strengths

The paper addresses a practical restoration problem in which banding and moiré coexist. MIRAGE is a potentially useful contribution, covering six content categories, five capture smartphones, and both individual and combined degradations (Section 3). The simulation pipeline provides controllable banding augmentation, with mathematical descriptions of five mask families in Appendix A.2. CLEAR achieves the best reported results on all six metrics in Table 5, and the visual examples show substantial artifact suppression. Comparisons with cascades in both orders and ablations of simulated data, frequency segmentation, and wavelet supervision provide useful initial evidence.

Critical weaknesses
  1. The training-data comparison is unclear. Baselines are described as retrained on MIRAGE, while CLEAR additionally uses simulated data derived from UHDM. Table 3 shows substantial gains from these data, so Table 5 does not clearly isolate the proposed components’ contribution.
  2. The ablations need reconciliation. Table 3’s Both row and Table 4’s Baseline row have identical LPIPS/DISTS but different ms-SSIM values (0.8219 versus 0.7909). If they represent the same configuration, the full model’s 0.8158 is below the former value. The unexplained discrepancy weakens the component-level conclusions.
  3. Evaluation and reproducibility remain limited. The selection of 512 × 512 test crops (Section 5.1) and the GT alignment procedure are unspecified. Single-artifact results are qualitative, with no held-out-device or full-resolution evaluation. FS optimization, wavelet weights, training schedule, and synthetic-data mixing are insufficiently documented. Matched-data controls and a RAW-versus-sRGB synthesis comparison would help establish which design choices drive the improvements.
Questions for authors
  1. Did all Table 5 baselines receive the same real and simulated data as CLEAR? Can you provide matched-data results for ESDNet and RIFLE?
  2. How do Table 3’s Both and Table 4’s Baseline configurations differ? If they are identical, which ms-SSIM value is correct, and what does IISP denote?
  3. How are test crops selected and GT images aligned? Please specify FS optimization, wavelet weights, and training settings, including the simulated-data mixing ratio. Can you report full-resolution runtime and text-fidelity results? Clarifying these points and the comparison fairness could materially change my assessment.
5 comments
AuthorsOpenReview ↗

Response to Reviewer Vrtp (Part 1)

Q1-1: Training data and matched-data comparisons.

A1-1:

Thank you for pointing out the difference in training data. We agree that using the same training data is important for a fair comparison of model design. (1) Training data in the main comparison. We view ISPS as a data augmentation method for joint flicker-banding and moiré removal. It is part of the CLEAR framework and one of our main contributions. We therefore used simulated data only for CLEAR in Table 5. The baselines were trained on real-world MIRAGE data, while CLEAR also used simulated data generated from UHDM. (2) Comparison with the same training data. We added experiments for ESDNet, RIFLE, and CLEAR under two settings: real data only, and the same real and simulated data. We add appendix A.1 to describe these experiments. (3) Results. Adding simulated data improves several metrics for all three methods. With the same real and simulated data, CLEAR achieves the best results on all six metrics. This supports the effectiveness of our model design under the same training conditions.

Table 6. Comparisons under matched training data. Bold marks the best result in each metric across all rows.

MethodTraining dataSSIM↑ms-SSIM↑LPIPS↓DISTS↓FSIM↑GMSD↓
ESDNetReal0.73070.79120.44920.27210.81610.1709
ESDNetReal + Simulated0.73500.81040.34980.22680.82680.1682
RIFLEReal0.68160.78720.37450.23710.82510.1679
RIFLEReal + Simulated0.73520.80180.33040.21750.83280.1793
CLEARReal0.74400.79850.30420.21610.81200.1720
CLEARReal + Simulated0.76000.81580.28360.20110.83520.1616

Q1-2: Clarification on the ablations.

A1-2:

Thank you for carefully checking our ablation results. (1) Errors and corrections. We sincerely apologize for the incorrect ms-SSIM value and the naming error. The incorrect value made the baseline appear better than the full model and it is reasonable to question our conclusions. The Both row in Table 3 and the Baseline row in Table 4 use the same configuration: real and simulated data, without FS or MWL. We have corrected the ms-SSIM in Table 3 from 0.8219 to 0.7909. We have also corrected IISP to ISPS, our ISP-based simulation pipeline. The ISPS row refers to training on simulated data only. (2) Analysis of the corrected results. All settings in Table 3 use the baseline without FS or MWL. Mixed training gives slightly lower ms-SSIM than training on simulated data alone, but better LPIPS and DISTS (Table 3). It combines diverse simulated degradations with the realism and complexity of real data. We therefore regard mixed training as the better overall choice for its data diversity and perceptual quality. Section 5.4 now explains this choice, with the changes highlighted. (3) Apology and conclusions. We sincerely apologize again for these errors and the extra effort they caused during your review. We should have caught them before submission and take full responsibility. The corrected results support our use of mixed training for diverse data and better perceptual quality. They also support the effectiveness of our proposed components, FS and MWL

Table 3. Effect of simulated training data.

Trainingms-SSIM↑LPIPS↓DISTS↓
MIRAGE0.72400.35560.2420
ISPS0.80280.31580.2233
Both0.79090.31010.2211

Q1-3: Test crop selection and GT alignment.

A1-3:

Thank you for asking us to clarify these procedures. (1) Random selection. We randomly sample four 512 × 512 patches from each of the 50 images in the MIRAGE test set, yielding 200 test crops. (2) LQ and GT alignment. We construct the GT by aligning the original displayed image with the captured LQ image. For feature matching, we convert both images to grayscale, apply 5 × 5 Gaussian smoothing, and equalize their histograms. We match features using the Scale-Invariant Feature Transform (SIFT), with a nearest-neighbor ratio threshold of 0.75. Random Sample Consensus (RANSAC) then estimates a homography from LQ coordinates to source-image coordinates. We use this mapping to resample the original displayed image onto the LQ pixel grid with bilinear interpolation. The resulting GT is spatially aligned with the LQ image, whose pixels remain unchanged. We then crop each aligned pair at the same coordinates. Appendix A.11 provides the equations and full settings.

AuthorsOpenReview ↗

Response to Reviewer Vrtp (Part 2)

Q1-4: Quantitative results on single-artifact removal.

A1-4:

Thank you for suggesting numerical comparisons for the two individual tasks. (1) Flicker-banding removal. We evaluate CLEAR and RIFLE, the SOTA debanding method, on the cropped MIRAGE flicker-banding test set. CLEAR performs better on all six metrics (Table 10). The visual examples show fewer residual bands and more uniform brightness (Appendix A.5, Figure 11). (2) Moiré removal. We also evaluate CLEAR and ESDNet, the SOTA demoiré method, on the cropped MIRAGE moiré test set. CLEAR performs better on all six metrics (Table 11). The visual examples show fewer moiré patterns and less color interference (Appendix A.6, Figure 12). (3) These results show that CLEAR is effective on both individual tasks as well as the joint task.

Table 10. Flicker-banding removal on cropped MIRAGE test images.

MethodSSIM↑ms-SSIM↑LPIPS↓DISTS↓FSIM↑GMSD↓
RIFLE0.71940.81370.35540.22170.83900.1525
CLEAR0.81470.86330.19740.15570.87140.1379

Table 11. Moiré removal on cropped MIRAGE test images.

MethodSSIM↑ms-SSIM↑LPIPS↓DISTS↓FSIM↑GMSD↓
ESDNet0.74120.83030.34200.21430.84470.1381
CLEAR0.82000.89440.16960.13110.89360.1141

Q1-5: Evaluation on held-out devices.

A1-5:

Thank you for suggesting a test on an unseen device. (1) Test setup. We collected a new flicker-banding and moiré test set of 10 images using another smartphone. This phone was not used to collect any training images. We evaluated ESDNet, RIFLE, and CLEAR on this set. (2) Results. CLEAR performs best on all six metrics in Table 7. The visual examples also show better removal of the artifacts (Appendix A.2, Figure 9). (3) Limitations and future work. Although we acknowledge that this new test set is small, its results still demonstrate CLEAR's strengths in cross-device generalization to some extent. As a direction for future work, we will consider expanding the test set to include more images and a wider range of unseen devices.

Table 7. Cross-device evaluation on 10 images from an unseen smartphone.

MethodSSIM↑ms-SSIM↑LPIPS↓DISTS↓FSIM↑GMSD↓
ESDNet0.78590.80750.27580.24210.83970.1351
RIFLE0.61150.74790.64040.34310.78410.1566
CLEAR0.80210.83300.24540.19950.84750.1258

AuthorsOpenReview ↗

Response to Reviewer Vrtp (Part 3)

Q1-6: Full-resolution quality, efficiency, and text fidelity.

A1-6:

Thank you for requesting full-resolution and text-fidelity results. (1) Why we use test crops. The main paper reports numerical results and visual comparisons on cropped test images (Table 5 and Figure 8). Our test images have a high native resolution. Scaling full images to fit the page makes residual artifacts and differences in fine details harder to see. Cropped regions show these differences more clearly. Some competing methods also lose restoration quality when processing images at full resolution. We therefore evaluate all methods on the same test crops to provide a fairer comparison of artifact removal. (2) Full-resolution restoration. We also provide results on complete 4K test images to assess performance at the original resolution. CLEAR achieves the best results on all six image-quality metrics and the OCR metric in Table 8. The visual comparisons also show better artifact removal while preserving scene details (Appendix A.3, Figure 10). These results support CLEAR's advantages in both numerical performance and visual quality under full-resolution evaluation. (3) Text fidelity. We use PP-OCR [1] to recognize text in the same regions of each output and its GT. OCR exact-match accuracy (EM) is the percentage of text regions whose recognized text exactly matches that from the corresponding GT region after Unicode normalization and whitespace removal. CLEAR reaches 14.24%, compared with 8.90% for ESDNet and 10.68% for RIFLE. This shows better agreement with the GT text. Appendix A.3 describes the details. (4) Runtime. CLEAR takes 173.18 s per image at 4096 × 3072 on an RTX 4090 48GB, using FP16 and batch size 1. Appendix A.4 and Table 9 give the inference settings and resource measurements.

Table 8. Full-sized image-quality and OCR results on the original 4K test set. EM is OCR exact-match accuracy between the restored and GT regions.

MethodSSIM↑ms-SSIM↑LPIPS↓DISTS↓FSIM↑GMSD↓EM (%)↑
ESDNet0.76640.79190.38750.19810.83880.16108.90
RIFLE0.54090.69690.72310.21600.84410.186710.68
CLEAR0.79420.84160.28290.11060.90960.137114.24

Table 9. CLEAR complexity and inference efficiency at 4096 × 3072.

ModelParams (M)FLOPs (T)Time (s)GPU (GiB)
CLEAR1292.00478.98173.1846.83

[1] Yuning Du, Chenxia Li, Ruoyu Guo, Xiaoting Yin, Weiwei Liu, Jun Zhou, Yifan Bai, Zilin Yu, Yehua Yang, Qingqing Dang, and Haoshuang Wang. PP-OCR: A practical ultra lightweight OCR system. arXiv preprint arXiv:2009.09941, 2020.


Q1-7: Implementation, training schedule, and data mixing.

A1-7:

Thank you for identifying the missing training details. (1) FS and MWL. We acknowledge that the original submission did not explain the implementation details and hyperparameters of FS and MWL sufficiently, making our method harder to reproduce. We add Appendix A.7 to explain the FS construction procedure and MWL implementation, and Table 12 lists the FS hyperparameters. (2) Optimization. We use LoRA rank 32 and train for 100,000 steps with AdamW. The learning rate stays at 5 × 10⁻⁵, and the total batch size is 12. AdamW uses beta values of 0.9 and 0.999, weight decay 0.01, and epsilon 10⁻⁸. Pixel and MWL losses are active from the start. Perceptual loss starts after 5,000 steps. (3) Data mixing. Training uses 2,790 real MIRAGE pairs and 2,790 simulated pairs, giving a 1:1 ratio. We shuffle the pairs and apply matched random 512 × 512 crops and horizontal flips. Appendix A.7 also provides the full training settings.


Q1-8: RAW-versus-sRGB synthesis comparison.

A1-8:

Thank you for suggesting a comparison of the simulation domains. (1) Experiment. We compared banding simulation in RAW and sRGB using the baseline model. Both settings use the same source images, banding gain maps, and training configuration. This keeps the comparison focused on the synthesis domain. (2) Results. RAW-domain simulation performs better on all six metrics in Table 13.These results support simulating banding in RAW before ISP conversion. Appendix A.8 gives the experiment and results.

Table 13. RAW- versus sRGB-domain simulation under full-sized 4K evaluation.

DomainSSIM↑ms-SSIM↑LPIPS↓DISTS↓FSIM↑GMSD↓
sRGB0.71460.80150.37000.17090.86140.1687
RAW0.71780.80610.33410.16370.86630.1659

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.