BlinkCode AI

Point your camera, get the file.

A file is compressed into a compact File DNA container, painted as SpectraCode colour matrices, and recovered by pointing a camera at the sender's screen — no server, no network after the first launch.

Live demo  ·  Landing page  ·  All projects

Download: Windows (.exe)  ·  Android (.apk)

---

BlinkCode AI is the high-capacity member of the optical-transfer family. Where a QR code carries black-and-white modules, a SpectraCode carries colour cells — five bits per cell instead of one — so a single frame holds kilobytes rather than hundreds of bytes.

file ─▶ arithmetic coder ─▶ File DNA ─▶ frames ─▶ RS(255,223) ─▶ colour grid
                                                                      │
        file ◀─ arithmetic decoder ◀─ CRC-32 ◀─ collector ◀─ RS ◀──────┘
                                              ▲
        camera ─▶ bullseye detect ─▶ homography ─▶ colour correction ─▶ classify

What it does

• SpectraCode transport — a 36-colour palette, 5 bits/cell, bullseye markers and a calibration strip, rendered at two densities (120×80 and 60×40).
• Real compression — an order-1 arithmetic coder with a match (LZP) model, plus a perceptual image mode. See the benchmark below.
• Reed-Solomon RS(255,223) — repairs up to 16 damaged bytes per block, which is what lets a frame survive glare, shadow and mis-classified cells.
• Full optical decoder — marker detection → homography → 3×4 colour correction → cell classification → RS, all in a Web Worker.
• Rateless (fountain) frame stream — an LT code with a systematic prefix, so the receiver rebuilds the container from any k + a little overhead symbols. A missed frame costs one more symbol, not a whole loop. CRC-32 verifies the result.
• Offline-first PWA — the model bundle is downloaded once and cached; after that the app never touches the network.

Compression benchmark

npm run bench — the arithmetic coder with and without the match model, against gzip. Measured on the project's own corpora:

corpus	orig	order-1	+ match	gzip
english prose (repeated)	18 000	6.6×	53.9×	136×
english prose (varied)	14 379	2.6×	12.5×	13.3×
JSON records	71 556	4.0×	8.6×	15.5×
source code	102 286	2.1×	2.9×	3.3×
already-compressed (gzip)	156	1.3×	2.0×	1.8×
random bytes	16 000	0.93×	0.92×	1.00×
total	222 377	2.4×	3.5×	4.2×

The match model added ~31% overall over the original order-1 coder. It is still beaten by gzip on long repeats, and cannot beat entropy — see the honest limits below.

Robustness

npm run robustness pushes a simulated camera (perspective tilt, colour cast, vignetting, blur, sensor noise) until the decoder breaks. Current margins on a standard-density code:

condition	still decodes up to
sensor noise	±32 (8-bit levels)
blur	a 3-pixel box radius
low light (gain)	0.3× – 1.0×

The decoder samples each cell with a median (robust to edge bleed and glare), stretches the capture's contrast before binarising, and tries several thresholds when locating the bullseye markers.

Perceptual mode

The lossy image mode quantises a photo into a 64-value latent. npm run perceptual measures reconstruction error:

image	RMSE	PSNR
smooth gradient	3.9	36.3 dB
photo-like	7.7	30.4 dB
2-pixel stripes	57.5	12.9 dB

The last row is the honest limit: a 64-value latent cannot represent near-Nyquist detail, and no reconstruction trick brings it back. Reconstruction now uses a bicubic upscale — bilinear left visible faceting on gradients and cost about 9 dB there.

Honest limits

These matter more than the pitch, so they are stated plainly here and in the app's "How it works" tab:

• One SpectraCode frame holds 5,111 payload bytes (1,097 in robust density). Throughput is roughly 5 KB × frame rate — about 15 KB/s at 3 fps. Frames are rateless, so a missed frame costs an extra symbol rather than a full loop.
• 1000:1 lossless compression of arbitrary files is impossible. Lossless coding cannot go below the entropy of the data. Real measured ratios are in the table above: strong on structured text, ~1× on already-compressed data.
• Extreme ratios exist only in perceptual mode, which is genuinely lossy — an image becomes a 64-value latent and returns as a soft reconstruction.
• The "neural" model is a mock with a real interface. The bundle format, one-time download, offline caching and the coder's model API are production shaped; the weights are trained order-1 statistics plus DCT quantiser scales (~129 KB), not a large neural net. Order1Model in src/lib/arith.js, Order1MatchModel in src/lib/predict.js and the transforms in src/lib/image-codec.js are the swap points.
• Camera decoding is validated in simulation, not against physical phone optics: the tests render a code, apply perspective tilt, a colour cast and vignetting, and decode it end to end. Real-world glare, motion blur and low-end sensors may still need per-cell voting and multi-frame averaging.

Run it

npm install
npm run build:model   # (re)train and package the downloadable model bundle
npm run serve         # http://localhost:8080
npm test              # codec + full optical loopback tests
npm run bench         # compression benchmark
npm run lint

Open the app, pick a file (or paste text), hit Compress, then Show SpectraCode on the sender and Start scanning on the receiver. Without a second device, the loopback self-test pushes every rendered frame back through the real optical decoder.

Project layout

src/
  index.html  app.js  styles.css  sw.js  manifest.webmanifest
  lib/
    arith.js          arithmetic coder + order-1 model
    predict.js        order-1 + match (LZP) model — the compression upgrade
    codec.js          File DNA encode/decode orchestration
    format.js         container header
    fountain.js       LT fountain codes (rateless frame stream)
    image-codec.js    perceptual latent encoder/decoder
    geometry.js       linear solver, homography, colour-correction fit
    palette.js        the 36 reference colours
    rs.js             Reed-Solomon GF(256)
    spectracode.js    layout, symbol packing, renderer
    spectra-decode.js marker detection → payload bytes
    storage.js        one-time model download + offline persistence
    stream.js         frame header, CRC-32, out-of-order frame collector
    surface.js        canvas / buffer drawing targets
  workers/            model loader, compression, decompression, optical decode
  models/ucg-v1.bin   downloadable model bundle
tools/                build.js, build-model.js, bench.js, tune.js, robustness.js, perceptual.js, update-scales.js, serve.js
tests/                codec, stream, predict + end-to-end optical loopback
docs/                 landing page + built demo (GitHub Pages)

Licence

BlinkCode AI is free for personal, academic, hobby and non-profit use under the PolyForm Noncommercial License 1.0.0.

Commercial use needs a paid licence — see COMMERCIAL.md.

Credits

Built by Dr Sohil Momin (@drsamonline).