umu/DailyMoire130

DailyMoiré130

Challenging moiré in everyday screen photography.

Project page · Code · Model adapter

DailyMoiré130 is a contribution of Dual-Scale Condition-Shift Editing for Robust Real-World Image Demoiréing. It contains 130 original-resolution smartphone photographs of everyday displays, including signage, monitors and laptops. It targets low-frequency color banding, fine high-frequency stripes and mixed-frequency interference across display–camera combinations.

Contents

• 130 input photographs in a single test split, using neutral IDs 0000–0129 (112 JPEGs and 18 PNGs).
• test/metadata.jsonl: neutral image IDs, display-oriented dimensions, file sizes and SHA-256 digests.
• manifest.json: complete file inventory and integrity information.
• EVALUATION.md: reference-free evaluation protocol and interpretation.

There are no aligned clean ground-truth images, no training split, and no per-image artifact-type labels. This public collection is derived from the selected input set used for the paper's DailyMoiré130 evaluation, with local image edits described below. It is distinct from the paired FHDMi and UHDM benchmarks. Model outputs and participant-level questionnaire records are not included.

Load with Hugging Face Datasets

from datasets import load_dataset

images = load_dataset("umu/DailyMoire130", split="test")
print(len(images))  # 130
sample = images[0]
image = sample["image"]
print(sample["image_id"], image.size)

The repository uses the standard ImageFolder layout. The metadata file_name column links each photograph to its row. Image IDs are stored as four-character strings.

For an offline copy, download a pinned snapshot:

from huggingface_hub import snapshot_download

folder = snapshot_download(
    repo_id="umu/DailyMoire130",
    repo_type="dataset",
    local_dir="datasets/DailyMoire130",
)

Use the repository commit hash as revision for a fixed benchmark copy. Verify the SHA-256 digests in manifest.json when preparing a benchmark copy.

Image processing and fidelity

This release preserves the original display-oriented resolutions. 18 images have local blur applied: screen portraits use small regions such as one eye, and selected incidental people and reflections use face regions. The remaining 112 images retain their decoded scene pixels, including 0012, 0058, and 0094. The edited images are saved losslessly as PNG; the other images retain their JPEG image payloads.

Pixels outside the annotated regions were verified unchanged against the decoded source images. JPEG EXIF/XMP/IPTC and comments are removed; only numeric display orientation is retained where needed. Source ICC color profiles are retained, including Display P3, so the unedited areas keep their original color interpretation. Files and manifests use neutral names and contain no source-machine paths.

The local edits are partial occlusion, not comprehensive anonymization. People, reflections, display text, and third-party advertising may remain visible. This release does not assert that all identifiable faces have been blurred. These are screen photographs, not camera RAW sensor data.

Local blur can affect restoration outputs and no-reference metrics. Historical paper scores were obtained on the historical input set; evaluations on this public derivative must identify the dataset revision and disclose these edits. They should not be assumed numerically identical to the historical run.

Evaluation

Use this set for unpaired real-world evaluation. PSNR, SSIM and paired LPIPS against the noisy input are not clean-target restoration metrics. The paper combines no-reference perceptual metrics, detector-based residual moiré measurements and a blind human preference study; see EVALUATION.md.

Terms and citation

Photographed display content retains its applicable rights. No additional dataset-specific license has been declared. Cite the accompanying manuscript when using the collection:

@misc{zhou2026dualscale,
  title  = {Dual-Scale Condition-Shift Editing for Robust Real-World Image Demoiréing},
  author = {Shiyang Zhou and Haijin Zeng and Congpei Qiu and Zhengcen Li and Ping Chen and Xiangming Wang and Zheng Li and Yongyong Chen and Jingyong Su},
  year   = {2026},
  note   = {Manuscript},
  url    = {https://clausy9.github.io/DemoireEdit/}
}