LAION-Mobile

A metadata-only, publicly-derivable benchmark of real smartphone photographs, extracted from re-LAION-5B by EXIF camera make/model. It is intended for measuring the false-positive rate of AI-image / deepfake detectors on genuine, ISP-processed phone photos.

Anonymous release. This repository accompanies a double-blind conference submission. It contains no author or institution information. A non-anonymised version with full citation will replace it after review.

What this release contains (and does not)

Per LAION redistribution policy we share manifests, metadata and scores only — no imagery. Downstream users reconstruct the image set by fetching the original URLs in the metadata.

File	Rows	Description
metadata/laion_mobile_full_metadata.parquet	935,399	The full usable dataset. One row per cleaned image with the source url, make/model, curated EXIF capture fields (ISO, f-number, focal length, exposure, lens model, dimensions, white balance, etc.), and coarse country (ISO-2) + continent for the 41.5 % of rows that had EXIF coordinates — enough to filter by region without exposing anyone's location. URL/make/model 100 % populated; capture params ~99 %. No precise GPS and no PII (exact coordinates, camera/lens serials, owner, artist, copyright, user comment and maker notes are excluded by design).
metadata/laion_mobile_1M_keep_ids.csv	935,399	Bare cleaned image-ID list (re-LAION-5B indices); a convenience subset of the parquet's image_id column.
metadata/laion_mobile_1M_nonphoto_flags.csv	69,407	The entries removed by the non-photo filter, with reason (logo/screenshot/aspect-ratio/EXIF-incomplete) and matched caption.
metadata/eval_sample_metadata.csv	9,115	EXIF-enriched metadata for the evaluation sample: image_id, make/model, capture parameters (focal length, aperture, ISO, exposure), image dimensions, CLIP similarity, and the source url. Run the build script with --hash to add a per-image content_sha256.
metadata/ai_pipeline_era.csv	44	Per-phone-model mapping to a computational-photography era (none/light/heavy) with sourced manufacturer notes.
scores/laion_mobile_scores.csv	100,265	Long-format per-image detector scores: one row per (image, detector), for the 11 detectors that scored the full evaluation set. Columns: dataset, detector, label, sensor, image_type, condition, iso, image_id, score, extra_json. label is always 0 (all images are real photos).

image_id is the join key across all files (it is the re-LAION-5B / manifest index).

The evaluation sample is the complete-case set of 9,115 images that carry a valid score for every one of the 11 detectors; eval_sample_metadata.csv and laion_mobile_scores.csv cover exactly this set.

How the subset was built (funnel)

1. re-LAION-5B (~5.5 B image–text pairs, safety-reviewed re-release of LAION-5B).
2. High-resolution subset (> 1024×1024) → ~170 M.
3. Entries carrying EXIF metadata → ~18 M.
4. EXIF Make/Model matching one of 36 smartphone manufacturers, deduplicated (perceptual hash + URL canonicalisation), dead URLs dropped, per-manufacturer balanced → ~1 M (LAION-Mobile-1M, 1,035,208 IDs).
5. Non-photo filter (caption regex + aspect ratio + EXIF completeness) → 935,399 cleaned real-photo IDs.

Reconstructing the images

import pandas as pd, requests
# full 935k set:
meta = pd.read_parquet("metadata/laion_mobile_full_metadata.parquet")
# (or the 9,115-image evaluation sample: pd.read_csv("metadata/eval_sample_metadata.csv"))
for _, r in meta.iterrows():
    img = requests.get(r["url"], timeout=15).content   # fetch from original URL
    # eval sample only: verify hashlib.sha256(img).hexdigest() == r["content_sha256"]

Expect link rot (~5–15 % of URLs dead at any time). If a content_sha256 column is present, use it to confirm you fetched the original bytes.

Filter by region using the coarse location columns, e.g. every African photo:

africa = meta[meta["continent"] == "Africa"]      # ~6,800 images
japan  = meta[meta["country"] == "JP"]

Intended use & limitations

• Use: as a real-only false-positive-rate benchmark for synthetic-image detectors, and for per-manufacturer / per-capture-condition drill-down.
• Bias: the pool skews toward older phones; only ~10.9 % come from heavy-ISP devices (iPhone XS and later), and the heavy tier tops out at 2020-era hardware. It therefore does not represent the most aggressive 2023+ computational-photography pipelines.
• Era labels are manufacturer-disclosed and may misclassify phones whose firmware later added ML processing.
• Images are in-the-wild internet uploads (already resaved / recompressed).

Licence & ethics

LAION metadata is released under CC-BY 4.0; individual images retain their original web licences and are not redistributed here. We rely on the re-LAION-5B maintainers' content filtering and recommend the safety-reviewed re-release for reproduction.

Privacy. We deliberately exclude precise GPS coordinates and every other potentially identifying EXIF field (camera/lens serial numbers, owner/artist names, copyright strings, user comments, maker notes); only camera capture parameters are kept. For images whose EXIF carried coordinates we publish only a coarsened location — the country and continent derived offline from the coordinates — so the data supports regional filtering (e.g. "all photos in Africa") while no individual's precise location is ever released. No imagery is hosted — users fetch images from the photographers' original public URLs. If you are the rights holder of an image and want its row removed, open an issue on the repository's discussion tab.

Citation

@misc{laionmobile_anon,
  title        = {{LAION-Mobile}: A re-LAION-5B-Derived Smartphone Real-Photo Benchmark},
  author       = {Anonymous},
  year         = {2026},
  note         = {Anonymous double-blind submission; metadata-only release.}
}