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

Shiyang Zhou; Haijin Zeng; Congpei Qiu; Zhengcen Li; Ping Chen; Xiangming Wang; Zheng Li; Yongyong Chen

cs.CVOct 19, 2026

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

Authors: Shiyang Zhou, Haijin Zeng, Congpei Qiu, Zhengcen Li, Ping Chen, Xiangming Wang, Zheng Li, Yongyong Chen,

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

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 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.

Figures & tables

Figure 2: Moiré at different frequency scales.

Figure 3: Overview of our dual-scale inference. Stage 1, Low-frequency Demoiré: the low-quality (LQ) moiré image is downsampled and paired with a black auxiliary reference as the condition. Stage 2, Condition Shift Refined Demoiré: the low-frequency output of Stage 1 is used as a condition for the first $n$ steps for structured guidance. The condition is shifted back to the original LQ image for the remaining $N{-}n$ steps for high-frequency content and detail recovery.

Figure 4: DailyMoiré130 captures challenging moiré on everyday displays (signage, monitors, laptops) using multiple smartphones. The dataset is publicly available; evaluation details are provided in Appendix D .

	User preference $\uparrow$			Automatic metrics
Method	Moiré	Detail	Overall	CLIP-IQA+ $\uparrow$	RMC $\downarrow$	RMA $\downarrow$
Input	—	—	—	.5690	.0523	.2507
ESDNet-L	8.6	8.6	8.8	.5478	.0340	.1599
MoCE-IR	1.8	2.3	2.2	.5277	.0341	.1602
AdaIR	1.4	2.5	2.0	.5464	.0447	.2163
OSEDiff	2.3	3.4	2.7	.5671	.0163	.0794

Table 1: Evaluation on DailyMoiré130. Blind Best-of-7 preference (%) from 31 participants, with 18 cases each. Uncertain responses remain in the denominator. The input is shown only as a reference.

	FHDMi					UHDM
Method	PSNR $\uparrow$	MUSIQ $\uparrow$	CLIP- IQA+ $\uparrow$	TOPIQ $\uparrow$	RMC $\downarrow$	PSNR $\uparrow$	MUSIQ $\uparrow$	CLIP- IQA+ $\uparrow$	TOPIQ $\uparrow$	RMC $\downarrow$
Input	18.00	56.76	.528	.442	.0395	17.12	64.04	.617	.537	.0652
ESDNet-L	20.67	55.56	.523	.408	.0175	21.16	65.44	.625	.556	.0139
MoCE-IR	22.40	56.33	.521	.424	.0150	20.92	64.23	.600	.521	.0189
AdaIR	21.80	56.00	.508	.425	.0205	20.61	63.27	.587	.509	.0342
OSEDiff	18.04	59.00	.545	.476	.0121	17.57	65.16	.612	.555	.0043

Table 2: Paired-benchmark comparison. Best in bold , second-best underlined . TOPIQ is no-reference. Uni-control denotes UniDemoiré-style augmentation of ESDNet-L. Remaining metrics are in Appendix G .

Panel of Figure 5 — Figure 5: Qualitative comparison on UHDM paired benchmark. Our method suppresses moiré while retaining visible texture. The lower-row crops show wood grain and fur around the toy’s face.

Panel of Figure 6 — Figure 6: Qualitative results on DailyMoiré130. Samples are cropped from the original 4K images for better visualization. These everyday screen captures illustrate challenging mixed-frequency moiré.

Setting	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	RMC $\downarrow$	RMA $\downarrow$	MUSIQ $\uparrow$
Vanilla FM (1 step)	17.51	0.669	0.195	0.0162	0.0953	65.29
w/o condition shift	20.24	0.751	0.189	0.0047	0.0255	66.84
w/o black image	20.00	0.737	0.189	0.0036	0.0195	66.81
Ours	19.99	0.789	0.134	0.0039	0.0212	67.43

Table 3: Training and inference ablations on UHDM.

$r$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	RMC $\downarrow$	RMA $\downarrow$	MUSIQ $\uparrow$
0.7	19.81	0.727	0.193	0.0032	0.0170	66.89
0.5	19.88	0.732	0.191	0.0036	0.0192	66.87
0.3 (ours)	19.99	0.789	0.134	0.0039	0.0212	67.43
0 (Stage-2 only)	20.02	0.793	0.135	0.0047	0.0256	67.42

Table 4: Condition-shift ablation on UHDM, including Stage-2 only (

r=0

Panel of Figure 7 — Figure 7: Visual ablation on DailyMoiré130. The full model reduces color banding in this example.

Method	0.02	0.03	0.05	0.07	0.10
ESDNet-L	0.1627	0.1193	0.0754	0.0534	0.0361
MZNet	0.1425	0.1142	0.0808	0.0621	0.0454
MoCE-IR	0.2203	0.1693	0.1110	0.0797	0.0531
OSEDiff	0.0469	0.0355	0.0239	0.0174	0.0118
Ours	0.0440	0.0327	0.0212	0.0155	0.0106

Table 6: RMA at different detector thresholds on UHDM. Lower is better.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Method / response	Moiré suppression	Detail preservation	Overall quality
Ours	417	382	398
ESDNet-L	48	48	49
Uni-control	30	38	37
MZNet	15	16	14
OSEDiff	13	19	15
MoCE-IR	10	13	12

Appendix

Table 8: Best-of-7 selection counts on DailyMoiré130. Each column contains 558 responses, including uncertain responses. Percentages are reported in Tab. 1 .

Criterion	Method	Rate	Participant CI	Image CI
Moiré suppression	Ours	74.7	[70.4, 78.9]	[69.8, 79.3]
	ESDNet-L	8.6	[6.1, 11.3]	[6.0, 11.4]
Detail preservation	Ours	68.5	[62.9, 73.7]	[63.4, 73.4]
	ESDNet-L	8.6	[5.9, 11.5]	[5.9, 11.6]
Overall quality	Ours	71.3	[65.1, 76.9]	[66.3, 76.2]
	ESDNet-L	8.8	[5.7, 12.2]	[6.1, 11.8]

Appendix

Table 9: Best-of-7 selection rates and separate cluster-bootstrap 95% confidence intervals (%). Participant clustering is the primary analysis, and image clustering is a sensitivity analysis.

	FHDMi			UHDM
Method	SSIM $\uparrow$	LPIPS $\downarrow$	MANIQA $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$	MANIQA $\uparrow$
Input	.707	.265	.311	.560	.335	.380
ESDNet-L	.789	.183	.315	.837	.118	.411
MoCE-IR	.831	.151	.307	.838	.145	.370
AdaIR	.808	.180	.299	.815	.173	.358
OSEDiff	.739	.201	.340	.784	.205	.400

Appendix

Table 10: Additional metrics on FHDMi and UHDM. Best in bold , second-best underlined . Uni-control denotes UniDemoiré-style augmentation of ESDNet-L.

Method	TOPIQ-FR $\uparrow$	PieAPP $\downarrow$	AHIQ $\uparrow$	$\|\Delta C^{*}\|\downarrow$	$\Delta H^{*}\downarrow$
ESDNet-L	0.639	1.137	0.433	4.454	2.776
Uni-control	0.641	1.045	0.439	4.273	2.865
MZNet	0.664	0.991	0.431	3.995	2.407
MoCE-IR	0.658	0.964	0.433	4.322	2.727
OSEDiff	0.588	1.310	0.284	6.352	4.034
Ours	0.670	0.984	0.435	4.422	2.643

Appendix

Table 11: Content and color fidelity on UHDM. TOPIQ-FR is full-reference, whereas TOPIQ in Tab. 2 is no-reference. Best values are bold. Uni-control denotes UniDemoiré-style augmentation of ESDNet-L.

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

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
Improving Complex Moiré Removal with Generative Supervision
Aug 18, 2026Xinyang Gu, Zhilu Zhang, Honglei Xu +3Image Restoration Synthetic Data Augmentation
Harbin Institute of Technology, Harbin, China · Alibaba Group - Taobao & Tmall Group
Fabric Image Demoiréing Benchmark from Synthesis to Restoration
Jun 23, 2026Pengchao Wei, Xiaojie GuoSynthetic Benchmark Generation Image Restoration
Tianjin University, Tianjin, China

May 25, 2025J/K move · Enter open · S save

Freqformer: Image-Demoiréing Transformer via Effective Frequency Decomposition

Xiaoyang Liu, Bolin Qiu, Zheng Chen, Libo Zhu+4

Shanghai Jiao Tong University

Frequency-Domain Image Processing Image Restoration

Open paper AskPDF arXiv

Image demoiréing remains a challenging task due to the complex interplay between texture corruption and color distortions caused by moiré patterns. Existing methods, especially those relying on direct image-to-image restoration, often fail to disentangle these intertwined artifacts effectively. While frequency-aware approaches offer a promising direction, their potential is hindered by the discrete transform (e.g., Haar wavelet or block-based DCT), which may suffer from spatial discontinuity, channel redundancy, and further cause error accumulation during their fixed inverse processes. In this paper, we present Freqformer, a Transformer-based framework specifically designed for image demoiréing through targeted frequency separation. Our method performs an effective frequency decomposition that splits moiré patterns into high-frequency spatially-localized textures and low-frequency scale-robust color distortions, which are then handled by a dual-branch architecture and an asymmetric training scheme tailored to their distinct characteristics. We further propose a learnable Frequency Composition Transform (FCT) module to adaptively fuse the frequency-specific outputs, enabling consistent and high-fidelity reconstruction. To better aggregate the spatial dependencies and the inter-channel complementary information, we introduce a Spatial-Aware Channel Attention (SA-CA) module that refines moiré-sensitive regions without incurring high computational cost. Extensive experiments on various demoiréing benchmarks demonstrate that Freqformer achieves state-of-the-art performance with a compact model size. The code will be made publicly available at https://github.com/xyLiu339/Freqformer.

Abstract

Image Restoration

Explore similar work

Freqformer: Image-Demoiréing Transformer via Effective Frequency Decomposition

Improving Complex Moiré Removal with Generative Supervision

Fabric Image Demoiréing Benchmark from Synthesis to Restoration