Improving Complex Moiré Removal with Generative Supervision

Xinyang Gu; Zhilu Zhang; Honglei Xu; Yanting Mei; Yukang Ding; Wangmeng Zuo

cs.CVAug 18, 2026

Improving Complex Moiré Removal with Generative Supervision

Authors: Xinyang Gu, Zhilu Zhang, Honglei Xu, Yanting Mei, Yukang Ding, Wangmeng Zuo

Organizations: Harbin Institute of Technology, Harbin, China · Alibaba Group - Taobao & Tmall Group

Abstract

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

Figures & tables

Figure 1: Results on moiré images. The (a) column presents three moiré inputs with large-scale chromatic bands and overlapping interference patterns. The (b) and (c) columns show Qwen-Image-Edit trained without and with our constructed WildMoiré dataset, respectively. Training only on the previous dataset leaves artifacts and suppresses details, while incorporating WildMoiré removes the broad interference and preserves original contents better.

Figure 2: Overview of WildMoiré dataset construction. We first collect real screen-captured images containing complex moiré and localize the display regions. Five image-conditioned generative models then produce candidate GT images. The candidates are processed at matched spatial locations through spatial and color alignment, synchronized patch extraction, pre-filtering, and optimal GT selection. This offline procedure yields 6.8K moiré-GT pairs at 1024 $\times$ 1024 resolution. The plots on the right summarize the improvements obtained by training ESDNet, SDXL, and Qwen-Image-Edit with constructed WildMoiré dataset.

Model	GPT-Image-2	Nano-Banana-2	FLUX.2	Qwen-Image-Edit	SDXL
Ratio (%)	62.73	17.97	9.10	5.96	4.23

Table 1: Distribution of models for the final selected supervisions in WildMoiré.

Dataset	Environment	Supervision	Resolution	Size
UHDM	Controlled	Digital source	$\sim$ 4328 $\times$ 3248	4,271
DCID	Controlled	Digital source	$\sim$ 4096 $\times$ 3072	7,176
WildMoiré	In the wild	Generation	1024 $\times$ 1024	6,832

Table 2: Datasets used in the experiment. Size of UHDM is counted after our alignment-quality filtering.

	Models	Training Datasets	Full-reference Fidelity Metrics			No-reference Perceptual Metrics
	Models	Training Datasets	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	MUSIQ $\uparrow$	TOPIQ $\uparrow$	Q-Align $\uparrow$
Closed Source	GPT-Image-2	–	15.1965	0.5265	0.4636	65.4375	0.5831	4.3665
Closed Source	Nano-Banana-2	–	16.8793	0.5679	0.4696	60.9621	0.5002	4.4743
Open Source	ESDNet (CNN-based)	UHDM+DCID	21.5002	0.7341	0.3471	38.1971	0.2770	3.6343
Open Source	ESDNet (CNN-based)	UHDM+DCID+WildMoiré	23.3714	0.7821	0.2614	44.2109	0.3413	3.9607
Open Source	SDXL (Diffusion-UNet-based)	UHDM+DCID	21.9034	0.7522	0.2951	41.4789	0.3203	3.8128

Table 3: Comparison of quantitative results. The optimal results in every part is highlighted in bold.

Figure 3: Comparison of qualitative results. The first example emphasizes logo and text fidelity under broad color interference, while the second contains severe mixed chromatic patterns over fine scene textures. Nano-Banana-2 is relatively content-faithful but may retain moiré, whereas GPT-Image-2 removes artifacts aggressively while replacing scene content and producing an overly digital-image appearance. Using WildMoiré suppresses residual bands more effectively across all 3 trainable models.

GT Source	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	MUSIQ $\uparrow$	TOPIQ $\uparrow$	Q-A. $\uparrow$
Nano-Banana-2	22.5640	0.7647	0.2921	40.9345	0.3087	3.8122
GPT-Image-2	23.0013	0.7721	0.2742	42.0071	0.3236	3.8834
Optimal of 5 Models	23.3714	0.7821	0.2614	44.2109	0.3413	3.9607

Table 4: Effect of candidate GT sources.

Training Datasets	Scale Aug.	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	MUSIQ $\uparrow$	TOPIQ $\uparrow$	Q-ALIGN $\uparrow$
UHDM+DCID	✗	21.5002	0.7341	0.3471	38.1971	0.2770	3.6343
UHDM+DCID	✓	21.6414	0.7468	0.3230	39.3647	0.2941	3.7478
UHDM+DCID+WildMoiré	✗	23.3164	0.7710	0.2611	43.1129	0.3224	3.9091
UHDM+DCID+WildMoiré	✓	23.3714	0.7821	0.2614	44.2109	0.3413	3.9607

Table 5: Effect of scale transformation augmentation.

$\tau$	#Pairs	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	MUSIQ $\uparrow$	TOPIQ $\uparrow$	Q-A. $\uparrow$
5	8,991	23.2863	0.7750	0.2720	43.0117	0.3309	3.8603
10	8,143	23.3411	0.7807	0.2643	43.7824	0.3365	3.9431
15	6,832	23.3714	0.7821	0.2614	44.2109	0.3413	3.9607
20	5,284	23.3647	0.7786	0.2684	43.7306	0.3374	3.9513
25	2,818	23.3021	0.7707	0.2925	42.9205	0.3315	3.8341

Table 6: Sensitivity to the threshold for GT Selection.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Generative Models	Prompt
GPT-Image-2	Remove the moiré patterns from the image. Strictly keep all other content unchanged. Preserve the original structure, colors, and fine details exactly. Ensure that all text on the screen remains clear, accurate, and completely unmodified. Preserve the original aspect ratio.
Nano-Banana-2	Remove the moiré patterns, together with any abnormal stripes, color bands, or color blocks caused by the moiré effect. Strictly keep all other content unchanged, including the original structure, colors, and fine details. Ensure that all text on the screen remains clear, accurate, and completely unmodified. Preserve the original aspect ratio.
SDXL, FLUX.2, and Qwen-Image-Edit	Remove only the moiré patterns from the image. Keep all other content unchanged, including the colors, text, and fine details.

Appendix

Table A: Prompts used for candidate GT generation.

Stage	Groups	Retained
Initial synchronized groups	13,615	100.00%
$\sigma_{img}$ and $\sigma_{lap}$ filtering	11,778	86.51%
ESDNet-response filtering	10,889	79.98%
A-FINE filtering ( $\tau=15$ )	6,832	50.18%

Appendix

Table B: Statistics of data filtering on constructing WildMoiré.

$\tau$	Total	GPT- Image-2	Nano- Banana-2	FLUX.2	Qwen- Image-Edit	SDXL
	(Number)	(%)	(%)	(%)	(%)	(%)
5	8,991	53.05	19.04	10.92	9.16	7.82
10	8,143	57.40	19.26	10.86	6.95	5.54
15	6,832	62.73	17.97	9.10	5.96	4.23
20	5,284	65.93	19.00	7.51	4.86	2.69
25	2,818	60.15	24.59	8.59	4.15	2.52

Appendix

Table C: Statistics of the optimal GT distribution under different A-FINE filtering thresholds. Total denotes the number of selected priors, while the model columns report row-wise percentages.

	Models	Training Datasets	Full-reference Fidelity Metrics			No-reference Perceptual Metrics
	Models	Training Datasets	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	MUSIQ $\uparrow$	TOPIQ $\uparrow$	Q-Align $\uparrow$
Closed Source	GPT-Image-2	–	16.4847	0.6288	0.3597	47.6108	0.4396	4.4022
Closed Source	Nano-Banana-2	–	19.5363	0.7320	0.2859	45.6903	0.4112	4.2289
Open Source	ESDNet (CNN-based)	UHDM+DCID	26.5372	0.8722	0.2483	34.1202	0.3006	3.9790
Open Source	ESDNet (CNN-based)	UHDM+DCID+WildMoiré	26.9181	0.8784	0.2441	34.9631	0.3109	4.0698
Open Source	SDXL (Diffusion-UNet-based)	UHDM+DCID	26.5779	0.8706	0.2467	36.5014	0.3082	4.0967

Appendix

Table D: Quantitative results on challenging cases from the UHDM and DCID test sets. The better result within each model pair is highlighted in bold.

Figure A: Additional qualitative comparisons on our test set. Incorporating WildMoiré consistently improves complex moiré suppression across ESDNet, SDXL, and Qwen-Image-Edit while preserving the captured content.

Figure B: Additional qualitative comparisons on challenging cases from the UHDM and DCID test sets. Models trained only on UHDM and DCID may retain broad color interference, local waves, or mixed high- and low-frequency artifacts. Incorporating WildMoiré improves complex moiré suppression across all three trainable architectures while maintaining the captured content.

Image Restoration

Image restoration uses computational methods to recover useful visual content from noisy, blurred, low-resolution, or otherwise degraded images.

See all 372 papers

Dual-Scale Condition-Shift Editing for Robust Real-World Image Demoiréing
Oct 19, 2026Shiyang Zhou, Haijin Zeng, Congpei Qiu +6Image Restoration Image Corruption Robustness
Harbin Institute of Technology, Shenzhen · Shenzhen Loop Area Institute · Xi’an Jiaotong University
Fabric Image Demoiréing Benchmark from Synthesis to Restoration
Jun 23, 2026Pengchao Wei, Xiaojie GuoSynthetic Benchmark Generation Image Restoration
Tianjin University, Tianjin, China
Freqformer: Image-Demoiréing Transformer via Effective Frequency Decomposition
May 25, 2025Xiaoyang Liu, Bolin Qiu, Zheng Chen +5Frequency-Domain Image Processing Image Restoration
Shanghai Jiao Tong University

Oct 19, 2026J/K move · Enter open · S save

Dual-Scale Condition-Shift Editing for Robust Real-World Image Demoiréing

Shiyang Zhou, Haijin Zeng, Congpei Qiu, Zhengcen Li+5

Harbin Institute of Technology, Shenzhen · Shenzhen Loop Area Institute · Xi’an Jiaotong University

Image Restoration Image Corruption Robustness

Open paper AskPDF arXiv

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 DemoiréEdit, 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.

Abstract

Image Restoration

Explore similar work

Dual-Scale Condition-Shift Editing for Robust Real-World Image Demoiréing

Fabric Image Demoiréing Benchmark from Synthesis to Restoration

Freqformer: Image-Demoiréing Transformer via Effective Frequency Decomposition