@pixagram/paph-js

PAPH 5.1 — an integer-only perceptual hash for detecting plagiarised pixel art, with a calibrated three-axis comparator on top of it.

One wire (version 5). One comparator (51). One shipped calibration (CAL-051-PROVISIONAL). JavaScript reference only in this release — the golden vectors in docs/golden/ are the contract the Rust/WASM engine will be held to.

github.com/pixagram-blockchain/paph-js · spec: docs/SPEC-005.1.md

What it answers

Given two pixel-art images, compare returns a verdict — Identical, Copy, Related, Suspected, Unrelated, or Indeterminate — with the evidence that produced it: seven structural channels, a staged D4-similarity geometry stage with an inversion family, and a diversity axis that demotes repetition-driven matches. Everything is integer arithmetic; the same bytes in give the same bytes out, on every platform, every time.

The comparator survives what pixel-art plagiarists actually do: any of the eight D4 transforms (rotations, mirrors, transposes), color inversion, tone shifts, nearest-neighbour upscales, crops pasted into new compositions — and combinations, including D4-plus-inversion.

Install and use

npm install @pixagram/paph-js

// CommonJS
const { hash, compare } = require('@pixagram/paph-js');
// ESM
import { hash, compare } from '@pixagram/paph-js';

const a = hash(rgbaPixels, width, height);   // { t1, t2, detail }
const b = hash(otherPixels, w2, h2);
const report = compare(a, b);
console.log(report.verdict, report.evidenceClass, report.assetRisk.level);

t1 is 2560 fixed bytes built for index search (screen(a, b) runs on Tier 1 alone); t2 is 32 + 32n bytes of evidence keypoints, n ≤ 512. Store both; compare accepts { t1, t2 } pairs or bare Tier-1 buffers.

Subpath imports: @pixagram/paph-js/wire (extraction only), @pixagram/paph-js/comparator (comparison only). Both are UMD — use default imports from ESM.

Browser

npm run bundle writes dist/paph53.browser.js (no dependencies, attaches window.PAPH51, byte-conformant with the Node build). Then open demo/playground.html: drop an image on side A, derive side B with one of the transform buttons (or drop a second image), and read the full verdict — geometry models by name, channel evidence, the report JSON. Everything runs in the tab; nothing is uploaded.

Calibration

The shipped profile is CAL-051-PROVISIONAL: thresholds and LUTs set from the conformance fixtures, honest about not having seen a production corpus. tools/calibrate.cjs fits a real profile from a labelled corpus:

node tools/calibrate.cjs corpus.json outdir            # draft
node tools/calibrate.cjs corpus.json outdir --bless    # release, if gates pass

corpus.json lists families of images (same work within a family, unrelated works across). The SPEC-005.1 §24.4 gates — at least 24 families, 60 images, 60 positive pairs, 10 measurable positives per channel — must all pass before the tool will emit a non-provisional profile. Fitted profiles round-trip through profileEncode/profileDecode and are passed to compare(a, b, { profile }).

Conformance

npm test

runs five suites: wire (parse/serialize, D4 algebra, determinism), assignment (the 004.2 exact objective against the dense oracle), comparator (19 verdict scenarios including D4×inversion composites, crops, repetition), golden (13 frozen vectors, tier hashes plus entire reports field-for-field), and the PNG tool codec. npm run golden regenerates the vectors — only lawful on a spec bump.

License

MIT © Pixagram SA.