Why imgstrip?

Most metadata removers delete the tags they recognize and copy the rest of the file untouched. Anything they don't recognize survives.

imgstrip works the other way around. It reads only the pixels and builds a brand-new file from them. By default nothing from the old file is copied over, so there is nothing left to leak. (The opt-in JPEG options --jpeg-lossless, --jpeg-canonical and --scrub-dct keep some of the picture's compressed data; see "Good to know" below.)

$ imgstrip --check holiday.jpg
FOUND    holiday.jpg (JPEG 640x480): 4 item(s)
           [metadata] EXIF: make=NIKON, model=COOLPIX P6000, date=2008:10:22 16:28:39
           [metadata] GPS location: 43.467448, 11.885127
           [hidden]   Embedded EXIF thumbnail: 6702 bytes
           [metadata] XMP: 4000 bytes

$ imgstrip holiday.jpg
cleaned  holiday.jpg -> holiday.clean.jpg
           removed: EXIF, GPS location, Embedded EXIF thumbnail, XMP

$ imgstrip --check holiday.clean.jpg
clean    holiday.clean.jpg (JPEG 640x480)

Your original file is never touched. The clean copy is saved next to it.

🚀 Quick start

You need Python 3.10 or newer.

pip install imgstrip

Prefer to keep it out of your main Python setup? pipx install imgstrip works too.

Then try it:

imgstrip photo.jpg          # writes photo.clean.jpg
imgstrip-web                # or use the point-and-click web app

Want to be sure the install works? Run imgstrip --self-test. It builds 21 booby-trapped test images, cleans them in every mode, and has independent image readers confirm the results.

git clone https://github.com/lloydsato/imgstrip.git
cd imgstrip
pip install .

🔍 How it works

flowchart LR
    A[Your photo] --> B[Decode to raw pixels]
    B --> C[Write a brand-new file]
    C --> D{Check 1: is every part of the new file accounted for?}
    D -->|yes| E{Check 2: do the pixels still match?}
    D -->|no| X[Stop. Nothing is saved]
    E -->|yes| F[Clean copy saved]
    E -->|no| X

If either check fails, nothing is written. Failing safe is the whole point.

(If you're reading this on PyPI and the diagram shows as plain text, the GitHub page renders it.)

🛠 Everyday use

Run imgstrip --help to see every option.

Going further

🖥 Web app

imgstrip-web

Drop photos onto the window and each one shows what it gives away. Click Clean, then download the results one by one, as a ZIP, or save them straight to a folder.

Open Options for the same switches as the command line: Deep scan, Keep JPEG data bit-exact, Scrub least-significant bits, Disrupt camera fingerprint, Sign with Content Credentials, plus the output format, quality and colour profile.

Options that do not apply to the images you added are greyed out, and ones that cannot be done for them are greyed out with a yellow warning triangle and the reason.

• Runs only on your computer (127.0.0.1) and works offline.
• Your originals are never written to disk, and nothing is uploaded anywhere.
• On a shared computer, start it with imgstrip-web --no-browser and open the printed link yourself.

📁 What it handles

Removed: EXIF, GPS, XMP, IPTC, comments, camera serial numbers, embedded thumbnails, depth maps, color profiles that name your device, data tacked on after the image, and colors hidden under transparent pixels.

⚠️ Good to know

No tool can do everything, and it's better to know the limits up front.

• What's in the picture stays. Faces, street signs, reflections and landmarks can reveal more than any tag.
• Your camera's sensor fingerprint stays unless you use --scrub-prnu. Even then, nobody can prove it's gone, and the scrub is not reliable against anyone who holds another photo of the same scene (a thumbnail, a second shot, the original document). Its strength was measured in simulation, not on real cameras, and it costs about 20 dB PSNR against the original; text can shift by several pixels.
• File names can leak dates (IMG_20240517_133700.jpg). Use --random-name (the web app has a Random file names switch).
• --check is a structural check, not proof. It lists what a file's container holds. It does not look inside compressed picture data, so it cannot see data hidden in JPEG image data or in the pixels (use --deep for the pixel tests).
• A clean scan isn't proof. "Not detected" means the tests found nothing, not that nothing is hidden.
• Invisible watermarks built to survive re-saving (such as SynthID) live in the picture itself and may survive.
• JPEGs are re-encoded (quality 95) by default. Use --jpeg-lossless to avoid any quality loss. That mode keeps the compressed picture data exactly as it was, so it doesn't clear steganography hidden in the image data itself, and it is meant for pictures you trust, not as a fix for a file the default mode refused. It does drop tables nothing uses and refuses a file that carries bytes no decoder reads. Re-encoding doesn't reliably clear it either: hiding tools that change zero coefficients mostly survive. --scrub-dct is built for that and destroys payloads read from the DCT coefficients, at a cost of roughly 33 to 39 dB of picture quality against the original. It never claims that nothing is hidden. See how it works.
• -i is not a secure erase. Old copies can linger in backups, cloud sync history and SSD wear-leveling.
• Very grainy JPEGs may be refused by the default mode because the re-encoded copy would drift too far from the original, and imgstrip would rather refuse than save something it couldn't verify. Use --jpeg-canonical for those (same pixels, rebuilt in one canonical form).

📚 Learn more

The full technical reference covers every option, the exact list of what gets removed, how the sensor-fingerprint scrub was measured, optional Content Credentials (C2PA) signing for journalists, and how the project is tested:

➡️ Technical reference

🧪 Development

pip install -e '.[test]'
pytest

📄 License

imgstrip is free software, released under the BSD 3-Clause License.

Versions up to 1.2.0 were released under the GNU Affero General Public License v3.0 or later; 1.2.1 and later are BSD-3-Clause.

Questions, bug reports and odd files that break it are welcome in the issue tracker.