Rapier JXL

JPEG XL encoder for integer and floating-point RGBA samples and JPEG coefficients. JavaScript, optional inlined WebAssembly, no runtime dependencies. Native 8-, 10-, 12- and 16-bit integers, IEEE binary16 and binary32, PQ and HLG transfer functions, and Rec. 2020 primaries. MIT.

npm install rapier-jxl

import {encode} from 'rapier-jxl';

const {data, width, height} = canvas.getContext('2d').getImageData(0, 0, canvas.width, canvas.height);
const bytes = encode(data, width, height, {quality: 100});
const jxl = new Blob([bytes], {type: 'image/jxl'}); // save as image.jxl

quality: 100 (default) is lossless; 1–99 is lossy. Alpha stays exact. For smaller lossless files, import encode from rapier-jxl/effort and set effort: 1–9 (default 1). rapier-jxl/wasm runs that search with optional integer kernels and identical output bytes.

Native precision and HDR

import {readSource} from 'rapier-jxl/source';
import {encode} from 'rapier-jxl';

const {data, width, height, ...source} = await readSource(fileBytes);
const bytes = encode(data, width, height, {...source, quality: 100});

Colour and alpha declarations describe the supplied samples. Encoding does not convert colour, tone-map, clamp HDR values or unpremultiply RGB. The optional source reader supports PNG16, including Adam7, and uniform HALF/FLOAT OpenEXR scanlines with NONE, RLE, ZIPS or ZIP compression. Source formats, error bounds and decoder qualifications.

Encoding under a CPU budget

The fixed-work benchmark uses Node 22.23.3 on an AMD EPYC 9V74 x86-64 host, with affinity to four virtual CPUs and an aggregate limit of 0.5 process CPU seconds per wall second. Four original-size inputs total 1,157,120 pixels: 640 × 448 text, 640 × 432 screen, 640 × 448 drawing and a 512 × 600 photograph. The text, screen and drawing inputs are authored corpus rasters. Every call uses quality 100.

Each table entry is the sum of four per-image medians, with one warmup and three measured repetitions per image. The worker count is additional threads; the caller also codes groups. Worker startup belongs to each job. These results describe this constrained x86 host; phone latency, browser scheduling, thermal performance and energy require target-device measurements.

Caller-thread CPU, milliseconds

Aggregate process CPU, milliseconds

Elapsed time under the CPU budget, milliseconds

The benchmark records complete stream hashes, source module hashes, input hashes, individual CPU and wall samples, caller profiles and actual governor accounting. node bench/phone-budget.mjs --inputs=inputs.json --out=results --max-pixels=1157120 --efforts=1,2,3,4,5,6,7,8,9 --workers=0,1,2,4 --rounds=3 --cpu-share=0.5 --cpus=0-3 --profile runs the same declared work. The manifest is a JSON array; each entry names an id, a kind, width, height, the rgbaHash (SHA-256 of the raw RGBA bytes) and pixels, the path of those bytes relative to the manifest. Split the efforts into smaller invocations when limiting individual command duration.

Encoder payload

Executable payloads, gzip 9; external packages measured 2026-09-30. These are encoder download sizes. All comparisons, capabilities and measurement method.

API, options and errors · Imports, sizes and inlining · Machine-readable facts · Browser example · Worker example · Contributing