FPX Converter

CI Latest release Licence: Apache-2.0

Converts Kodak .fpx (FlashPix) photographs — the format the DC200 and DC210 wrote around 2000 — into archival TIFFs and shareable JPEGs, carrying the metadata across instead of throwing it away.

The FPX Converter window

• Decodes .fpx files that nothing modern opens. Its own decoder, not a wrapper round a broken one.
• Writes two useful things: a lossless TIFF with every pixel the camera captured, and a JPEG that opens anywhere.
• Keeps the metadata as standard EXIF, XMP and IPTC — and is careful about dates, because a wrong date is worse than none.
• Never writes to your originals. The folder you point it at is only ever read from. Nothing under it is written, moved, renamed or deleted, and a destination inside it is refused with an explanation.
• Runs unattended over thousands of files, never stopping the batch for one bad file, and picking up where it left off if it is interrupted.

Get fpx-converter-<version>.exe from the latest release . One file. Nothing to install, no Python needed.

The app is not code-signed, so Windows treats it as an unknown program. Signing is not currently in place — a paid certificate is part of the reason, but free signing for open-source projects (SignPath) and cheaper commercial options both exist and have not been ruled out. You will see up to three warnings, in this order:

1. Your browser says the file "isn't commonly downloaded" or similar. Choose Keep — in Chrome and Edge this is behind the ... menu next to the download, then Keep anyway.
2. A blue full-screen box: "Windows protected your PC". The only button you can see is Don't run. Click the small More info link above it, then the Run anyway button that appears.
3. Possibly your antivirus. One-file bundles built with PyInstaller are a well-known false-positive trigger for heuristic scanners. If yours objects, the honest answer is to check the source and build it yourself — see docs/BUILD.md.

None of that is a judgement about the file. It is what Windows shows for every program that has not paid for a certificate.

You do not have to take any of that on trust. Each release publishes a .sha256 file next to the .exe. On Windows, in PowerShell:

Get-FileHash fpx-converter-<version>.exe

and compare the result against the contents of the published .sha256 file. Get-FileHash prints its answer in capitals and the file holds the same characters in lower case; that difference is the display, not the hash.

Each release is also signed with GitHub build provenance, so if you have the GitHub CLI this also works:

gh attestation verify fpx-converter-<version>.exe --repo sremich/fpx-converter

It prints nothing when it passes. Silence there is the good outcome — a file that has been altered, or one that did not come from this repository, fails with an error instead.

Both prove the same thing: that this is the exact file this repository's release workflow built from this source, not one assembled by somebody else and given the same name. Neither stops the SmartScreen warning above — that is about the certificate, not the contents.

ExifTool writes the metadata. It has its own licence, and this project keeps it separate rather than bundling it, so it has to be installed once.

The app checks for ExifTool when it opens. If it is not there, it says so and offers to install it — one button, and you are done. The steps below are for anyone who would rather run it themselves, or whose machine has no winget. If ExifTool is already installed somewhere the app cannot find on its own, its Locate exiftool.exe... option lets you point at it directly, and remembers the choice for next time.

Open a terminal — press Win, type powershell, press Enter — and run:

winget install --id OliverBetz.ExifTool

Check it worked by closing that window, opening a new one, and running:

exiftool -ver

A version number means you are done. If winget is not recognised (older Windows 10), download the Windows package from exiftool.org instead, unzip it, and point the app at exiftool.exe with the FPX_EXIFTOOL setting or the --exiftool flag.

Without ExifTool, convert refuses to start and tells you so. It does not write a run's worth of images and then report every one of them failed.

Choose the folder holding your .fpx photos, choose where the converted photos should go, pick one of the three output choices, and press Convert.

Full walkthrough, in plain language: docs/GUIDE.md.

<destination>/
 archive/2002/Summer 2002/2002-07-04_143210_Backyard.tif
 sharing/2002/Summer 2002/2002-07-04_143210_Backyard.jpg
 conversion.log
 audit_report.json
 run-state.json

The command line writes both trees by default. The desktop app writes one image per photograph — whichever of its three choices you picked — and the tree it lands in follows the framing, so a whole-frame image goes to archive/ and a cropped one to sharing/.

• archive/ keeps the full frame, lossless — every pixel the camera captured. sharing/ gets the crop, where somebody framed one in the Kodak software at the time.
• Folders keep your own folder names by default, nested under the year; you can also file by year, by year and month, all in one folder, or by a pattern.
• Filenames are {year}-{month}-{day}_{time}_{name} by default, with any part the evidence does not support written as zeros. {name} is required in any pattern you write: in this kind of archive the filename is the only thing a person actually typed, and a pattern without it discards that for good.

There is no capture date in these files. The FlashPix capture-date property is absent from every one of them; the only timestamp present is an import-batch stamp, recording when the photographs reached a computer.

So that stamp goes to DateTimeDigitized, never to DateTimeOriginal. DateTimeOriginal is written only where a date is defensible as a single day — a day-precise folder name, an embedded film-scan date, or a date you supply yourself through the review page. A folder naming a year or a season gives an ordering key and a filename prefix, not a claim. That is why many of your filenames will read 0000-00-00_000000_: the archive being honest.

docs/DATES.md explains all of it, including how to supply the dates you know.

docs/TROUBLESHOOTING.md — SmartScreen, missing ExifTool, filenames full of zeros, wrong time zones, and the rest.

The command line is the whole converter; the window is a front end over it.

python -m venv .venv
.venv/Scripts/python.exe -m pip install -r requirements-dev.txt # Windows
# .venv/bin/python -m pip install -r requirements-dev.txt # POSIX
python -m ruff check .
python -m pytest

Install -r requirements-gui.txt instead if you are working on the desktop front end, building the executable, or running its tests — it pulls in requirements-dev.txt as well as Qt.

Then, at its shortest, a conversion is two commands:

python -m fpx_converter scan ./photos --manifest ./work/manifest.json
python -m fpx_converter convert --manifest ./work/manifest.json --dest ./converted

Every command and flag: docs/CLI.md. ExifTool is a separate install; see step 3 above.

• A .fpx file is an OLE2 compound document holding a resolution pyramid of JPEG-compressed tiles, with the JPEG tables stored apart from the tiles.
• The decoder reads the tile table, splices the tables back onto each tile, stitches, crops to the declared size, and applies the file's own viewing transform — a 90° rotation, a crop, or both.
• Colour space is read per file — NIF RGB or PhotoYCC. A file that declares neither is decoded as RGB and says so, in the log and the audit report.
• ExifTool writes the tags and Pillow reads them back, with defusedxml for the XMP packet. Validating with the tool that wrote proves much less than it appears to.
• Pillow's own FpxImagePlugin is never used in the conversion path: on the reference corpus it opened 39 of 1,265 files and hard-crashed CPython on two.

docs/FORMAT.md is the full format write-up; ARCHITECTURE.md is why the pipeline is shaped this way.

• Tested on all three platforms at tiers 1 and 2. CI runs on windows-latest, ubuntu-latest and macos-latest with Python 3.14, and the CLI has no Windows-only dependencies. Tiers 3 and 4 — the real-archive and eyeball tiers — have only ever run on Windows, because that is where the source archive lives. The desktop app is Windows-only, shipped as a single .exe.
• Python 3.14. pyproject.toml declares >=3.14; 3.14 is what CI runs and what the exact dependency pins are verified against.
• Rotation and rotation-plus-crop geometry has tier-2 cover, but only synthetically. Every rotated file in the source archive contains a person and cannot be published, so a copy of a person-free fixture is given a viewing transform at test time instead — the transform lives in fixed-width float32 fields, which can be rewritten without changing the stream length. Rotation, rotation-plus-crop and axis-aligned crop are all covered this way, including the defect where a rotation carrying a crop once resolved to no crop. What that cannot cover is a photograph whose own embedded thumbnail witnesses its crop, so colour and orientation on a genuinely transformed file remain tier-3 and tier-4 work: a local run, a real archive, human eyes. See tests/fixtures/README.md and docs/TESTING.md.
• Colour and orientation need eyes at least once per variant. "It decoded" is not "it decoded correctly", and this project has the scars to prove it — see docs/PROJECT-HISTORY.md.

Bug reports, questions and pull requests go to the issue tracker. Read CONTRIBUTING.md first — it lists the rules that are not negotiable, chief among them that the source archive is read-only and that nothing personal goes in the repository. Security reports: SECURITY.md.

Apache License 2.0. Copyright Stevie Remich. See LICENSE and NOTICE. Third-party components keep their own licences, listed in THIRD-PARTY-NOTICES.md with the full copyleft texts in LICENSES/; the same notice is inside the application, under Help → Licences. The test fixtures are not covered by the Apache licence — see tests/fixtures/LICENSE.md.

Kodak, FlashPix, DC200 and DC210 are trademarks of Eastman Kodak Company; this project is independent and unaffiliated.

ExifTool by Phil Harvey writes every tag this tool produces. Qt for Python (PySide6) by The Qt Company, used unmodified and linked dynamically, is the window. Pillow, NumPy, olefile and defusedxml do the rest of the heavy lifting.