pdfmini

Compress and combine PDFs in your browser. Files are processed on your device and never uploaded.

Use it: https://den-run-ai.github.io/pdfmini/

Features

• Compress – keeps text, vector graphics, links and form fields intact and only re-encodes the images, at a resolution that matches how large each image is actually shown on the page.
• Never makes a file bigger – if no meaningfully smaller version is possible, pdfmini says so instead of handing you a worse file.
• Three simple levels – Light, Balanced (default) and Strong. Everything else is under Advanced settings: image resolution, image quality, grayscale, removing document info, and Convert pages to images for scans.
• Combine – merge PDFs in any order (drag to reorder on desktop, arrow buttons on touch screens), then compress the result in one tap.
• Same results everywhere – browsers use different built-in JPEG encoders (Safari uses Apple's, Chrome and Firefox use libjpeg-turbo), so the same quality setting gives different files. pdfmini ships its own encoder, which produces byte-identical output in every browser (checked in CI), so sizes and quality match on iPhone, Android, Mac and Windows.

Compression levels

Level	Images resampled to	JPEG quality	Good for
Light	200 dpi	82	Printing, zooming in, archiving
Balanced	144 dpi	72	Email, sharing, reading on any screen
Strong	110 dpi	55	Upload limits, the smallest file that stays readable

Images are only resampled when they are more than 15% sharper than the target for the size they are drawn at, and JPEGs that are already at or below the target quality are left alone, so repeated compression does not keep degrading a file. Lossless images such as charts and screenshots are re-encoded losslessly unless JPEG is dramatically smaller.

See Calibration for how these values were chosen.

How it works

1. pdf-lib parses the PDF.
2. Page, form and annotation content streams are interpreted to find the size at which every image is drawn, which gives its effective DPI.
3. Images above the target DPI are decoded, resampled and re-encoded with pdfmini's JPEG encoder (in Web Workers). Charts and screenshots also get a lossless (PNG-predicted Flate) candidate, and soft masks are resampled to match their images.
4. Images browsers cannot decode (CMYK, JPEG 2000, Lab, ...) are decoded by PDF.js, which is only downloaded when such an image or Convert pages to images is needed.
5. Lossless cleanups:
   • Form XObjects that a page draws only once are drawn directly in the page's content. Flattened form fields (tax returns and other filled government forms) leave one small XObject per field, and each pays for its own stream header and compresses poorly on its own.
   • The Adobe XFA copy of a static form is removed. The form's fields and page content stay (they are what every viewer except Acrobat shows); dynamic XFA forms are left alone. Adobe Reader usage rights, which any rewrite invalidates, are removed so Reader doesn't warn.
   • Unused objects, duplicate streams (common after merging), private application data and embedded thumbnails are removed.
   • Uncompressed, ASCII-encoded and LZW streams are re-compressed, and objects are packed into object streams (except for PDF/A-1 files, which do not allow them).
6. If the result is not at least 1% smaller, the original file is kept. The result's Details show how the original divides into images, fonts, text and drawings, and other data.

Privacy

Everything runs locally in the page. The only network requests are for the open-source libraries (pdf-lib, and PDF.js when needed) from the jsDelivr CDN, pinned to exact versions and checked with Subresource Integrity. A Content-Security-Policy enforces this: the page may only run its own code and those two pinned packages, and may only fetch from the pinned PDF.js package, so a document cannot be sent anywhere. Your settings are remembered in your browser's local storage.

Browser support

Current versions of Chrome, Edge, Firefox and Safari on desktop, iOS and Android. Every change is tested in Chromium, Firefox and WebKit, plus emulated Android and iPhone, in CI.

• Compressing and combining need Safari 14.1+, Chrome/Edge 85+ or Firefox 90+.
• CMYK/JPEG 2000 images and Convert pages to images use PDF.js, which needs Safari/iOS 16.4+, Chrome/Edge 110+ or Firefox 115+. On older browsers those images are left as they are.

Limitations

• Password-protected or encrypted PDFs cannot be optimized directly. If the file opens without a password, Convert pages to images works.
• Fonts are not subset or merged. A document combined from many forms embeds a copy of the same fonts for each form, and text-only PDFs usually shrink only a little.
• Removing XFA drops Acrobat-only form scripts (such as calculations); the fields and their values stay.
• Digital signatures do not survive any modification; pdfmini tells you when a signed PDF was compressed.
• Combine keeps pages and their content, but not bookmarks or interactive form fields of the merged files.
• Very large files (hundreds of MB) can exceed a phone's memory.

Development

It is a static site with no build step: index.html, js/app.js (UI), js/engine.js (PDF processing) and js/jpeg.js (JPEG encoder).

npm install
npm run serve            # http://localhost:8080
npx playwright install   # once, to get Chromium, Firefox and WebKit
npm test                 # all browsers + mobile emulation
npm run test:chromium    # unit tests + Chromium (desktop and mobile)

Tests serve the CDN libraries from node_modules, so they run offline against the exact versions the app uses. UI tests fail on any Content-Security-Policy violation.

Deployment

The site is published to GitHub Pages by GitHub Actions (.github/workflows/ci.yml): every pull request runs the full test suite, and each push to main that passes it deploys an artifact containing only the files the site serves (index.html, js/, LICENSE). The repository's Pages source must be set to GitHub Actions (Settings → Pages → Build and deployment). Dependabot keeps the workflow's actions and Playwright up to date.

Updating pdf-lib or PDF.js

The libraries are pinned in package.json (for tests), in the CDN URLs and SRI hashes in index.html and js/engine.js, and in the Content-Security-Policy in index.html. After changing the version in package.json and running npm install, node tools/sri.mjs prints the new URLs and hashes; a unit test fails until all of them match.

Calibration

The levels were calibrated on a corpus of real and synthetic documents (phone scans, photo albums, reports with photos and charts, slides, screenshots, CMYK print files, text-only papers, forms, a JPEG 2000/JBIG2 book scan) by measuring output size and visual similarity (SSIM of pages rendered by poppler, an independent renderer, at 110 and 200 dpi), next to Ghostscript's presets, and by inspecting zoomed-in text and photo crops.

Results on the eight image-heavy files (100 MB in total):

	Size of original	SSIM @110 dpi	SSIM @200 dpi
pdfmini Strong (110 dpi, q55)	3.0%	0.932	0.878
pdfmini Balanced (144 dpi, q72)	6.0%	0.941	0.913
pdfmini Light (200 dpi, q82)	13.8%	0.965	0.948
Ghostscript /screen	5.7%	0.869	0.855
Ghostscript /ebook	19.8%	0.883	0.918
Ghostscript /printer	68.5%	0.936	0.959

Higher SSIM is closer to the original (1.0 is identical). Each level is about half the size of the one above it; going from 144 dpi/q72 to 150 dpi/q75 added 16% in size for no visible difference, and below 110 dpi small scanned text starts to blur. Text-only documents gain 0-37% from the lossless steps alone (fonts are left untouched); files that are already tightly packed are returned unchanged.

Forms gain the most from the lossless steps, with pages that render pixel-identically in poppler and PDF.js: fillable IRS forms (W-2, W-9, 1099, 1040 and schedules) shrink 41-75% once the static XFA copy is removed, and a 32-page return made of filled and flattened IRS forms shrinks 57%, mostly by drawing its 1,716 field graphics inline.

To repeat the measurements:

node tools/run-engine.mjs --level balanced --out out/ my.pdf   # run the engine
python tools/quality.py my.pdf out/my.pdf                       # SSIM, text and qpdf checks
python tools/calibrate.py --out cal/ --dpi-eval 110,200 \
  --settings "110:55,144:72,200:82" --gs screen,ebook,printer corpus/*.pdf

run-engine.mjs --cpu-throttle 4 approximates a mid-range phone, tools/make-form-return.mjs builds the flattened IRS-form return, tools/profile-engine.mjs prints a CPU profile, and tools/screenshots.mjs captures the UI on mobile and desktop in light and dark mode.

The Python tools need poppler-utils, numpy, pillow, scikit-image and pikepdf (and Ghostscript for the reference presets).

License

MIT