devmandan/broken-qr-codes-104

Broken QR Codes (104)

104 real QR code images that ordinary decoders fail on, with byte-level ground truth where it could be established, geometric annotations, and per-image results for eight reference decoders.

This is a test set, not a training set. It exists to make QR decoder claims falsifiable: every image here was collected because something failed on it, so the distribution is deliberately adversarial rather than representative.

Why this is hard

59 of the 104 images are read by none of the eight reference decoders — not Apple Vision, not Apple CIDetector, not WeChat's CNN or classical detector, not any of the three ML Kit configurations, not marian_mark_decoder. The best single decoder in the set reads 38.

The failure modes are the interesting part: artistic QR codes where the module grid is deliberately obscured by generated imagery, physical damage, extreme perspective, low contrast, motion blur, partial occlusion, screen glare, and resolutions from 0.05 to 16 megapixels.

What's in each row

field	notes
image	the picture itself
file_name, image_id, width, height
has_payload, payload, payload_bytes_hex	56 of 104 have an established payload
payload_source	how the payload was established, where recorded
qr_version	QR version 1–9, on 30 images
quad	the symbol's four corners in pixels, on 33 images
finder_centers	the three finder-pattern centres, on 30 images
u_px_per_module	pixels per module, on 30 images
quad_source	which decoder or process established the geometry
precision_note	stated tolerance where the geometry is approximate
solved_by_*	eight booleans, one per reference decoder
n_decoders_solving	0–8

The 56 / 104 split matters for scoring

Only 56 images have an established payload, so 56 is the real ceiling — a decoder cannot be credited for reading an image nobody can check. Reporting "X/104" without saying that 48 images are uncheckable overstates every score, including ours. Score against 56, or state the denominator you used.

payload and payload_bytes_hex are not equally populated: 56 rows carry a payload string and 32 carry the raw bytes. Where both exist they agree; where only the string exists, binary payloads may not round-trip through UTF-8, so compare bytes when you have them.

Reference decoder results

Measured on one machine (Apple M2, MacBook Air 13" 2022, 8 GB, macOS 14.3.1), decode call only, image already resident in memory.

decoder	solves / 104	false	p50 µs
marian_mark_decoder (default, 1T)	38	0	8,982
Apple Vision	21	0	11,195
Apple CIDetector	18	0	7,168
WeChat (CNN)	14	0	—
WeChat (classical)	13	0	—
ML Kit iOS	6	0	—
ML Kit Android (bundled)	4	0	—
ML Kit Android (GMS)	2	0	—

Taller is not a superset. Apple Vision alone reads 4 images that the 38-solve decoder misses, and 17 of that decoder's 38 are unique to it. Everything in the table together reads 45 of the 56 checkable images. Rank by the column you actually care about, and check the intersection before concluding one decoder dominates another.

Scoring is not uniform across these rows, and the difference matters. marian_mark_decoder returns bytes and is compared byte-for-byte. Apple returns a string plus a raw codeword stream and is scored on both surfaces separately. WeChat is string-only. ML Kit coverage was imported from physical-device runs and its payloads were never recorded, so those three rows cannot be byte-checked at all and are coverage claims only.

Usage

from datasets import load_dataset

ds = load_dataset("devmandan/broken-qr-codes-104", split="test")

# the checkable subset — the only fair denominator
checkable = ds.filter(lambda r: r["has_payload"])          # 56

# the images every reference decoder failed
unsolved  = ds.filter(lambda r: r["n_decoders_solving"] == 0)   # 59

# geometry-supervised subset
geo = ds.filter(lambda r: r["quad"] is not None)           # 33

Provenance, licensing, and takedown

Please read this before redistributing.

The annotations in this dataset — payloads, quads, finder centres, versions, module scales, decoder results — are original work and are offered under CC BY 4.0.

The images are a different matter and are not uniformly owned by the publisher of this dataset. They were collected from public web sources as examples of QR codes that fail to decode, including artistic/generated QR codes whose individual authorship is not documented. They are published here for research and benchmarking, which is the purpose for which they were gathered. No claim of ownership is made over them.

If you own an image in this set and want it removed, open a discussion on this repository and it will be taken down.

Payload contents are real. Several decoded payloads are live third-party URLs that were encoded in the original codes, some carrying opaque identifiers. They are reproduced verbatim because a QR benchmark whose payloads have been rewritten cannot be used to check a decoder. Do not treat any URL here as endorsed, safe, or maintained, and be careful about resolving them automatically.

What is deliberately not here

A separate 19-image validation set exists and is withheld on purpose. It has never been trained on, tuned against, or scored more than once, and publishing it would destroy the only clean holdout this line of work has. A benchmark you can iterate against stops measuring generalization; that set exists so there is one number that was not fit.

Citation

@misc{broken_qr_codes_104,
  title  = {Broken QR Codes (104): a hard real-world QR decoding benchmark},
  author = {Legut, Daniel},
  year   = {2026},
  url    = {https://huggingface.co/datasets/devmandan/broken-qr-codes-104}
}

The decoder this benchmark was built alongside, marian_mark_decoder, is named for David Marian Legut.