Repository navigation
Releases
LumenPDF 0.3.1
LumenPDF is now under the Apache License 2.0. Free for anything, including commercial use, with no obligation to ask.
There are no code changes in this release. 0.3.0 was built before the relicense, so the LICENSE.txt inside its installer and portable zip still said Business Source License 1.1 — telling anyone who downloaded it that commercial use required a paid licence. The repository said one thing and the artefact said another, and the artefact is what people actually get. This release fixes that.
Why the licence changed
The Business Source License reserved commercial rights until 2030. It was protecting revenue that was not going to arrive: those terms only bite when a company comes to buy a licence, and enterprise procurement wants a support contract, an SLA, and someone to hold liable — all of which need a company behind them. In practice the clause did not make anyone pay; it made them not use it.
Meanwhile it was charging three real costs:
- Linux distributions cannot package a non-OSI licence. Debian, Fedora and Arch all require one, and Linux is on the roadmap.
- Free code signing for open source requires an OSI licence.
- It was a barrier to the help this project most needs — more interface translations, and the macOS and Linux ports.
The argument for publishing the source survives the change and is better served by it. A tool that claims to destroy text on a page is asking to be trusted with something irreversible, and "read the code yourself" is the only honest answer to that. That argues for the most permissive licence, not a restricted one.
It also happens now rather than later because relicensing needs every copyright holder's agreement, and today that is one person.
The name
Section 6 of Apache 2.0 grants no trademark rights. Fork it, change it, sell it — but an unofficial build should not call itself LumenPDF. See NOTICE.
If you want to help
The two things this project cannot do alone are translations — one JSON file each, see translations/README.md — and the macOS and Linux ports, for which every platform-specific call is already isolated in src/platform/.
The installer is still not code-signed, so Windows SmartScreen warns on first run — More info → Run anyway. Verify your download against SHA256SUMS.txt:
(Get-FileHash .\LumenPDF-0.3.1-win64-setup.exe -Algorithm SHA256).Hash.ToLower()LumenPDF 0.3.0
Six things that made LumenPDF feel broken rather than unfinished.
Printing
There was none. A PDF application without printing is a strange thing to ship.
Page ranges, copies, greyscale, fit-to-page or actual size — and printing to a
PDF file keeps the document's own page size instead of reflowing it onto A4.
Pages are rasterised through PDFium at the printer's own resolution rather than
handing the printer a PDF, so what is on screen is what comes out: annotations,
filled form fields and OCR text layers included, instead of depending on how
complete the printer's own PDF interpreter is.
Links
They did nothing when clicked.
Internal jumps now follow. External links always show you the real target
first, because what a PDF prints on the page and where its link actually
points are two unrelated strings. Only http, https and mailto are ever
handed to the system, and a Launch action — one that names a program to run —
is not offered as a link at all.
Encrypted documents
A password-protected PDF was refused with a message. It now asks for the
password. It is used once and written nowhere: this application has no keystore
of its own, and storing someone else's document password without one is not a
trade worth making.
It remembers things now
Theme, window size, the last dozen files you opened, and where you stopped
reading in each of them. None of that survived a restart before — switching to
the light theme and finding it dark again next launch was the daily version of
this.
繁體中文・简体中文・日本語
The interface was English only. It now ships fully translated in three more
languages and follows your system locale by default.
Only complete translations are offered. A partial one falls back to English
string by string and produces a window in two languages with no error anywhere.
Adding a language is a single JSON file — see
translations/README.md.
Tests run on their own
Every push now builds and runs the full suite on GitHub Actions — 107
assertions, up from 65. Releases are built the same way and are not published
if they fail.
Two real bugs this found
CMakeLists.txt still said 0.1.0 after v0.1.0 and v0.2.0 had shipped, so
the v0.2.0 binary reported itself as 0.1.0. The new update checker would have
told those users an update was available forever. The release workflow now
refuses to publish a tag that does not match the declared version.
annotationAt() could close a page held by the form-fill environment.
PDFium caches page objects per document, so that would have destroyed the
widget state of whoever was editing that page — the same cause as an earlier
run of form bugs.
Please read before trusting it with something
The signature feature is an ink stamp, not a cryptographic digital
signature. It does not identify who signed and does not break if the document
is altered afterwards. If you need a legally recognised signature, this is not
it. The wording in the README and on the site has been corrected to say so.
The installer is not code-signed, so Windows SmartScreen will warn on first
run — More info → Run anyway. A certificate is issued to an identified legal
entity and costs a few hundred dollars a year; the support is implemented and
waiting for one. See
docs/CODE-SIGNING.md.
Verify your download against SHA256SUMS.txt:
(Get-FileHash .\LumenPDF-0.3.0-win64-setup.exe -Algorithm SHA256).Hash.ToLower()Still not built
Annotation review sidebar, sticky notes and shapes · tabs · cryptographic
signatures · page composition (insert blank, crop, watermark, split) ·
PDF↔Office conversion · more UI languages · macOS and Linux.
LumenPDF 0.2.0
Adds OCR, Google Drive, a recorded demo, and coverage for writing systems
beyond Latin.
The installer is 1 MB smaller than 0.1.0 despite the new features, because
OCR uses the recogniser Windows already ships rather than bundling one.
OCR — scanned pages become searchable
An invisible text layer is written underneath the image. The scan looks
identical, and becomes searchable, selectable and copyable.
Verified by flattening a page to a bare scan and recognising it: the page went
from 0 characters to 254, and a search that found 6 matches found 8.
Backed by Windows' own OCR engine — offline, nothing extra to download, and it
uses whichever languages your system already has. The app lists the languages
your machine actually has rather than pretending every language works. The
trade is stated plainly: available languages are Windows' to decide, and this
is the one part with no cross-platform story yet.
Every script, not just Latin
"CJK support" was the wrong framing for what is a correctness problem in every
writing system. Ten writing systems are now regression-tested on every change,
each with a phrase chosen for how it can fail: right-to-left order in Arabic
and Hebrew, stacked tone marks in Vietnamese, clusters in Devanagari and Thai,
final sigma in Greek, and scripts written without spaces.
Google Drive
Open from Drive, edit, and save back to the same file, keeping its ID and
sharing. Scoped to drive.file, so it can only touch files you explicitly open
or that LumenPDF created — it cannot see the rest of your Drive, and that
boundary is enforced by Google rather than promised by us.
Requires your own OAuth client, which takes a few minutes to create. A
client ID identifies an application and cannot be obtained on your behalf;
see docs/GOOGLE-DRIVE-SETUP.md.
Until credentials are present, the Drive actions stay hidden rather than
appearing and failing.
Sign-in uses PKCE in your real browser, not an embedded one, and the refresh
token is encrypted with the Windows Data Protection API.
Demo
The video on https://scotteliu.github.io/LumenPDF/ is the application
recording its own window frame by frame while a scripted timeline drives the
real controllers. Nothing is staged.
Verification
65 assertions in scripts/run-tests.ps1, none of them screenshot comparisons.
LumenPDF 0.1.0
First public build of LumenPDF.
Measured on this build (scripts/benchmark.ps1, median of 5 runs, RTX 4090 @ 143 Hz):
| 3-page document | 1000-page document | |
|---|---|---|
| First page on screen | 287 ms | 451 ms |
| Whole process lifetime | 536 ms | 527 ms |
| Memory | 132 MB | 145 MB |
| Scrolling 1000 pages | — | 144 fps, no dropped frames |
Opening a 1000-page document costs the same as opening a 3-page one.
What it does
Reading, search, outline, thumbnails, text selection and copy, highlight /
underline / strike-through, page rotate / reorder / delete, merge, single-page
export, signatures as vector paths, AcroForm filling, image and text export,
image downsampling, and redaction that destroys the text rather than covering
it. Everything that modifies the document is undoable.
CJK text is a first-class case: extraction normalises the Kangxi-radical
duplicates that make copied Chinese match nothing when pasted, single-ideograph
search works, and double-click selects one character rather than the whole line.
What it does not do
- Text editing corrects rather than typesets. Real PDFs embed font subsets,
so typing a glyph the font lacks would produce garbage. Every edit is read
back and rolled back if it does not survive re-encoding. - Windows only for now. macOS and Linux are planned.
- No OCR yet, so scanned documents are not searchable.
Verification
59 assertions in scripts/run-tests.ps1, none of them screenshot comparisons.
Ten deliberately broken PDFs — empty, truncated, corrupted, self-referential
page tree, a page declaring itself 200000 inches square — are refused or
recovered without crashing, and none can push memory past 250 MB.
Requirements
Windows 10 or 11, 64-bit. The installer installs per-user (no UAC prompt) and
does not seize the .pdf default association.
Licensed under BSL 1.1 — free for personal, academic and non-profit use;
commercial use requires a licence; converts to Apache 2.0 on 2030-07-30.