The source is different
Screen moiré comes from rigid display lattices. Fabric moiré comes from non-rigid garments with dense, anisotropic, semi-periodic weave structures, folds, and pose changes.
Motivation
Why fabric moiré needs its own benchmark
Existing screen-camera demoiréing work assumes a more regular sampling source and a cleaner separation between artifact and scene content. The paper argues that fabrics break both assumptions.
The artifact is coupled
In fabrics, aliasing overlaps the genuine textile spectrum. Removing interference without deleting authentic stripes, grids, and herringbone details becomes substantially more ill-posed.
The data pipeline breaks
Supervised demoiréing needs pixel-aligned pairs, but real garments deform. Global homography-style registration is not enough for large-scale fabric data collection.
PRISM Benchmark
Physics-based residual injection for aligned supervision
PRISM avoids the impossible part of real fabric capture: obtaining a perfectly aligned clean garment target after non-rigid deformation. It simulates the imaging chain, extracts a moiré residual through a round-trip construction, and injects that residual back into the original clean grid.
1. Viewpoint and sampling
Warp, downsample, apply CFA mosaicking, noise, and demosaicking to induce realistic aliasing.
2. Round-trip residual
Compare aliased and anti-aliased branches under matched geometry to isolate moiré components.
3. Residual injection
Add the residual to the clean image so texture remains pixel-aligned and supervision stays stable.
Dataset construction
The clean pool starts from web-collected garment images with high-frequency textile patterns, followed by duplicate removal, quality filtering, garment localization, text removal, semantic filtering, and expert inspection. The final release contains 14,587 training pairs and 1,463 testing pairs.
- Candidate images
- 42,939
- Clean GT pool
- 1,963
- Retained PSNR range
- 10-20 dB
FaDeNet
A conservative restoration baseline for textile detail
The method is designed around the risk identified by the motivation analysis: aggressive restoration can erase real fabric texture. FaDeNet therefore separates base and detail components and gates corrections spatially.
Base/detail decomposition
A content-adaptive split lets the model correct low-frequency color and illumination shifts while keeping high-frequency textile structures explicit.
Mask-gated updates
Corrections are applied where the confidence mask predicts moiré-dominant regions, limiting unnecessary edits on clean fabric areas.
Spectral-anisotropic blocks
SAGB combines directional spatial filters with coarse-scale FFT gating to model stripe-like and narrow-band periodic interference.
Results
Fabric-aware modeling improves both fidelity and user preference
On PRISM, FaDeNet outperforms representative screen-camera demoiréing baselines across PSNR, SSIM, and LPIPS. On real unpaired fabric captures, it also receives the highest MOS in the user study.
| Method | PSNR ↑ | SSIM ↑ | LPIPS ↓ | Params | FPS |
|---|---|---|---|---|---|
| Input | 16.998 | 0.7268 | 0.3495 | - | - |
| P-BiC | 28.590 | 0.9689 | 0.0334 | 4.922M | 38.17 |
| ESDNet-L | 28.809 | 0.9711 | 0.0268 | 10.623M | 19.22 |
| FaDeNet | 32.159 | 0.9859 | 0.0169 | 7.051M | 38.10 |
Open Challenge
Severe spectral entanglement remains difficult
The failure cases reinforce the paper's central motivation: when colored moiré overlaps authentic garment patterns, restoration can still over-suppress valid textile details. Future work needs stronger texture-aware priors and real-world adaptation.
Citation
Cite this work
Proceedings metadata can be updated after the final camera-ready release.
@inproceedings{wei2026fabric,
title = {Fabric Image Demoir{\'e}ing Benchmark from Synthesis to Restoration},
author = {Wei, Pengchao and Guo, Xiaojie},
booktitle = {European Conference on Computer Vision},
year = {2026}
}