wlensfun

lensfun compiled to WebAssembly — lens distortion, transverse chromatic aberration and vignetting correction in the browser, driven by lensfun's database of measured lens profiles.

Try it: lensfun.wasmedia.org — load a photo, pick the camera and lens, and correct it entirely in the browser.

Part of the wasmedia family of single-threaded + SIMD WASM ports that work on any site without COOP/COEP headers.

Install

npm install @wasmedia/wlensfun
npx wlensfun-fetch-db --out public   # download the lens profile database

The profile database is not bundled in the package. It is licensed CC BY-SA 3.0 rather than LGPL, upstream releases it independently of the library, and it is larger than the rest of the package combined. The CLI fetches it and repackages it for browsers as lensfun-db-v2.json.gz, written to --out <dir> (the current directory if omitted). Serve it alongside your app.

Usage

import { initWLensfun } from '@wasmedia/wlensfun';

const lf   = await initWLensfun();
const db   = await lf.loadDatabase('/lensfun-db-v2.json.gz');

const cam  = db.findCamera('Canon', 'Canon EOS 5D Mark III');
const lens = db.findLens(cam, null, 'Canon EF 24-70mm f/2.8L II USM');

const mod  = db.modifier(cam, lens, {
  focal: 35, aperture: 5.6, distance: 3,
  width: image.width, height: image.height,
});

const corrected = mod.apply(pixels);   // reusable across frames

mod.dispose();
db.dispose();

Both handles wrap raw pointers into WASM memory, so dispose() is not optional — nothing else frees them.

The database can also be loaded straight from lensfun's XML files — an OPFS directory in a worker, or a host directory under Node — with lf.loadDatabaseFiles(). It uses wopfs to read the files in place rather than copying them into memory.

Cameras match on their EXIF spelling first, then fuzzily, so both ('Canon', 'Canon EOS 5D Mark III') and wlibraw's ('Canon', 'EOS 5D Mark III') work. Lenses always match fuzzily.

From a raw file

examples/node/correct-raw.mjs decodes a camera raw with wlibraw, takes the camera, lens, focal length and aperture from its metadata, and corrects it in linear light:

npm run fetch-db
node examples/node/correct-raw.mjs photo.cr2      # add --lens "…" if the file records none

It writes photo.original.png and photo.corrected.png for comparison. Raw data is linear, so it corrects with colorSpace: 'linear'. wlibraw's decode() can't produce linear output yet (wlibraw#1), so the example gets it from LibRaw's dcraw_emu -4 through execute().

Frames shot in a crop mode (Nikon DX, 1:1, video crops) need --crop: the camera's crop factor describes its full sensor, and wlibraw doesn't yet report the 35 mm-equivalent focal length that would reveal the smaller frame (wlibraw#2). Pass the equivalent focal length over the real one, e.g. --crop 1.5 for DX on a full-frame body.

What it does

apply() runs every correction the lens has calibration data for, in lensfun's documented order: vignetting first, then distortion and TCA fused into a single resampling pass so the image is interpolated once rather than twice. Perspective correction and projection conversion are available through the corrections argument.

If you would rather resample yourself — on the GPU, say — coordinatesForRow() hands back the raw coordinate map instead.

Correction is CPU-bound and synchronous; @wasmedia/wlensfun/worker runs it off the main thread.

Documentation

• API reference
• Building from source

Scope

wlensfun corrects; it does not decode. Raw handling belongs to wlibraw, format conversion to wvips and wimagemagick. This library takes and returns raw pixel buffers — 8-bit, 16-bit or 32-bit float, 3 or 4 channels.

License

LGPL-3.0-or-later, matching upstream lensfun. See LICENSE.

The lens profile database is separately licensed CC BY-SA 3.0 and carries its own attribution and share-alike obligations. It is fetched at build time rather than bundled, which keeps the two licenses distinct.