QR by Jean Pierre Kolb

1. Encoding Modes

Before the first bit of your payload is stored, the code picks an encoding mode. Different modes pack characters at different densities, so the right choice can shrink the code dramatically.

Mixed content: a single QR code may switch modes mid-stream to stay compact — e.g. numeric for a product code followed by byte for a URL slug.

2. The Bitstream

The payload is built as a single stream of bits, then chopped into 8-bit codewords. The layout of that stream is always the same:

• Mode indicator (4 bits) — which encoding mode was chosen.
• Character count (8-16 bits, depending on version and mode) — how many characters follow.
• Encoded data — the actual payload, bit-packed according to the mode.
• Terminator — up to four 0 bits marking the end of data.
• Padding — fills the remaining capacity with the repeating bytes 0xEC and 0x11. The pattern is arbitrary; what matters is that it’s predictable.

3. Error Correction Levels

QR codes are built to survive damage. A chunk of the capacity is always spent on error-correction codewords, and you get to pick how much:

Higher ECL costs capacity: the same text in H needs a bigger code than in L. See the live comparison further down.

Where the Level Is Recorded

Note: before storage the 5 header bits (ECL + mask) are expanded to 15 bits via BCH(15, 5), then the whole string is XOR’d with the fixed pattern 101010000010010 so the format info can never accidentally resemble a finder. A decoder reverses the XOR, runs BCH correction, and reads the top two bits.

4. Reed-Solomon Error Correction

The redundancy behind error correction comes from Reed-Solomon codes operating over a finite field called GF(256). You don’t need the maths to use QR codes, but the one-sentence intuition is useful:

Treat the data codewords as coefficients of a polynomial. Reed-Solomon appends more coefficients so that any n missing or wrong values can be reconstructed from the rest — as long as the total errors stay under a known limit.

Practical consequence: you can punch a hole through a QR code, cover part of it with a logo, or print it on crumpled paper, and as long as the damage fits within the ECL’s budget, scanners recover the full payload.

5. Interleaving

Before being placed into the grid, data and error-correction codewords are interleaved in blocks. Why? A large scratch tends to destroy adjacent modules. By spreading codewords from different blocks across the code, no single burst of damage can wipe out a whole block’s-worth of error correction. It’s the same trick used on scratched CDs.

6. Zig-Zag Data Placement

7. Masking

If the raw bitstream happens to produce a pattern that looks like a finder pattern, or a big block of one colour, scanners will struggle. To avoid this, the encoder XORs the data modules with one of 8 standard mask patterns, then scores the result against a penalty function. The mask with the lowest penalty wins — and its index is stored in the format info so the scanner knows how to undo it.

The four penalty rules (simplified):

1. Long runs of same-colour modules in rows / columns.
2. Blocks of 2×2 same-colour modules.
3. Patterns that resemble a finder (1 : 1 : 3 : 1 : 1 ratio).
4. Overall imbalance between dark and light (should be near 50 / 50).

Versions 1-40 — Grid Sizes

Every version adds 4 modules per side. Alignment pattern count grows roughly quadratically with the grid size.

Maximum Capacity (Version 40)

Worst-case payload the largest QR code can hold, per encoding mode and error-correction level.

Alphanumeric Character Set

The alphanumeric mode is stricter than it sounds: lowercase letters are not allowed. Stick to this set and you save ~30% of bits per character vs. byte mode.

Common Content Formats

QR codes carry plain strings. To trigger actions on a phone, apps agreed on a handful of text conventions. The generator above implements the most common ones.

Practical Tips

• Always keep the quiet zone. Four modules of background on every side. Violating this is the #1 reason real-world codes fail.
• Scan-distance rule of thumb: 10 : 1. A 2 cm code is reliable from ~20 cm, a 20 cm poster code from ~2 m, and so on.
• Contrast matters more than colour. Dark-on-light is safest; inverted (light-on-dark) is legal but not every scanner app handles it. Pure black / white is the most compatible.
• Use ECL H only when you need it. Adding a logo? Use H. Printing on glossy paper that will scratch? Use Q or H. Otherwise M is a sensible default.
• Shorten URLs before encoding. The difference between a 30- and 80-character URL is the difference between a tiny V3 code and a much larger one — and smaller codes scan faster and from further away.
• Test on real devices. iOS Camera, Android Camera, and the popular scanner apps each have quirks. A code that looks fine on your screen can fail when printed.

History & Standards

• 1994 — Invented by Masahiro Hara at Denso Wave (Japan) to track car parts on the Toyota assembly line. The name stands for “Quick Response”.
• 1999 — Published as the Japanese standard JIS X 0510.
• 2000 — Adopted internationally as ISO/IEC 18004; still the authoritative spec today (current revision: ISO/IEC 18004:2024).
• Micro QR — A smaller variant (11×11 to 17×17) for minimal payloads, one finder pattern only.
• rMQR (Rectangular Micro QR) — ISO/IEC 23941, rectangular shape for labels with constrained height.
• iQR, SQRC, Frame QR — Denso Wave’s own extensions: larger capacity, encrypted data, or an image-carrying frame.
• Denso Wave does not enforce the QR Code trademark on generators and readers — that’s a big part of why QR took off while competing 2-D barcodes didn’t.

Further Reading

• qrcode.com — Denso Wave’s official site, overview of QR variants.
• Wikipedia: QR code — comprehensive encyclopedic entry.
• qr.blinry.org — interactive walk-through of decoding a QR code by hand, highly recommended.
• Thonky’s QR Code Tutorial — long-form, step-by-step tutorial through the whole pipeline.