Skip to main content

MagicDispel: wipe the metadata, keep every pixel

Remove private metadata from photos and videos on your own computer, without changing a single pixel.

PyPI Tests Python 3.10+ macOS, Windows and Linux MIT license

A photo carries more than its pixels: where it was taken, when, on which phone, and sometimes hidden extras such as depth maps, portrait mattes and even the serial number of your display. A video may carry a GPS track of every second and its camera's serial number. MagicDispel saves a clean copy with none of that, and no pixel, video frame or sound sample is changed.

Quick start

macOS and Linux:

curl -LsSf https://raw.githubusercontent.com/v1nc3nt-continualab/magicdispel/main/install.sh | sh

Windows (PowerShell):

powershell -ExecutionPolicy ByPass -c "irm https://raw.githubusercontent.com/v1nc3nt-continualab/magicdispel/main/install.ps1 | iex"

Then type magicdispel and a space, drag photos or videos into the terminal, and press Enter. A cleaned copy appears next to each original.

Before cleaning, ExifTool shows the photo's location, phone model and capture time. MagicDispel cleans two files and explains why it refuses a PDF. Afterwards, ExifTool finds none of those fields.

Why MagicDispel

  • It rebuilds instead of deleting. Most tools delete the metadata they know about. MagicDispel writes a new file from only the parts needed to show the image, so metadata it has never heard of is left behind too. The parts it keeps must match their exact layout, so almost nothing can ride along inside them; the few exceptions are listed in the privacy details.
  • Every pixel stays. Compressed image data is copied byte for byte, never recompressed. Color profiles, orientation, transparency, animation and HDR gain maps are kept, and so is the DPI of JPEG, PNG, TIFF and BMP files, so a photo looks the same, in HDR too. A video keeps every frame and every sound sample as they were, with its rotation, HDR, Dolby Vision and spatial video information.
  • It finds what others miss. Depth maps and portrait mattes inside iPhone photos, thumbnails and previews that may show the uncropped original, C2PA manifests, data hidden after the end of the image, the dates and times in file names, the display model and serial number in screenshots' color profiles, and the GPS tracks, face data and camera serial numbers inside videos.
  • It checks before it saves. Every result is read back on its own and compared with the original, frame by frame. ExifTool, if installed, gives an independent second opinion. If anything is off, nothing is saved.
  • It stays on your computer. No uploads, telemetry or network access. Originals are never modified, and existing files are never overwritten. An original keeps all its metadata, so share only the copy.

Compared with the usual one-liner

Many guides suggest exiftool -all=. ExifTool is an excellent metadata editor, and with extra options it can keep color data, but the one-liner and MagicDispel give very different results. Tested on real photos with ExifTool 13.55 on macOS:

Photo exiftool -all= MagicDispel
iPhone portrait-mode HEIC Removes the capture tags but keeps all 10 hidden images of the tested photo: depth map, portrait matte, skin, hair, teeth, glasses and sky mattes, style map and thumbnails, plus a 58 KB Apple property list Removes every hidden image with its bytes and keeps only the HDR gain map. Looks identical
iPhone HDR JPEG Also removes the HDR gain map, the HDR headroom and the color profile: the photo loses its HDR and its colors shift Removes the private data; keeps the gain map, headroom and colors. Looks identical, in SDR and HDR
Mac screenshot Also removes the color profile and the DPI: colors shift, and the image opens at twice its size Removes the display's name, model and serial number from the profile; keeps colors and DPI. Looks identical

"Looks identical" means identical renders in macOS ImageIO and ColorSync: pixels, sRGB, Display P3, HDR, gain maps, orientation and DPI.

For videos, guides suggest exiftool -all= or ffmpeg -map_metadata -1 -c copy. Tested on public sample videos with ExifTool 13.55 and FFmpeg 8.1 on macOS:

Video exiftool -all= ffmpeg -map_metadata -1 -c copy MagicDispel
GoPro HERO8 Keeps the GPS track (231 positions), the camera and lens serial numbers, the firmware, the model and the capture time Removes them, and writes its own name as the encoder Removes them all. Plays identically
iPhone Dolby Vision video Keeps the capture time in the movie and track headers Drops the Dolby Vision configuration, and rewrites the edit list: the clip's length and frame times change Removes the private data, keeps Dolby Vision and the timing. Plays identically
Apple spatial video Keeps the capture time Rewrites the HEVC configuration and shifts the frames Keeps the spatial video information. Plays identically

"Plays identically" means that FFmpeg decodes identical frames, and that macOS AVFoundation sees the same video and sound tracks and shows the same frames.

How it works

  1. Identify the format from the file's contents, not its name; the copy's extension follows the contents too.
  2. Rebuild a new file from an allowlist: the image data and the few fields needed to show it, such as color, orientation and HDR, some of them written afresh. Everything else is left behind.
  3. Verify the result. It must parse on its own, decode to identical frames (for HEIC and videos, every image item and sample is compared byte for byte), and satisfy ExifTool when it is installed. The original must not have changed in the meantime.
  4. Save the copy next to the original under a new name, without ever overwriting a file.

Files that cannot be rebuilt safely, such as damaged files, RAW photos or unknown structures, are refused with an explanation instead of being passed through. Privacy and format details describe every check.

What is kept and removed

Format Kept Removed
JPEG image data, JFIF density, orientation, DPI, color space, ICC profile; HDR gain-map images (MPF), Apple HDR headroom and gain-map XMP other EXIF and XMP, IPTC/Photoshop, comments, C2PA, thumbnails and preview images, maker notes, trailing data
PNG, APNG image data, palette, transparency, color chunks (sRGB, gAMA, cHRM, cICP, HDR), DPI, animation, ICC profile, orientation text, time stamps, C2PA, private chunks, anything after the end
HEIC, HEIF image items and tiles, HDR gain maps (Apple and ISO), alpha, orientation, ICC profile, HDR XMP fields EXIF, other XMP, Apple property lists, depth and calibration, portrait and semantic mattes, style maps, thumbnails, item descriptions and times, unused data
AVIF as HEIF, including animations, which keep only what plays them as HEIF, plus sequence times, names and user data
WebP image data, alpha, animation, ICC profile, orientation other EXIF, XMP, unknown chunks
GIF images, palettes, frame timing, transparency, loop count, ICC profile comments, text overlays, XMP, other extensions
TIFF image data, decoding tags, DPI, orientation, page numbers, ICC profile EXIF and GPS directories, XMP, IPTC, Photoshop, descriptions, private tags, sub-images
BMP converted to lossless PNG with the same pixels, DPI and profile (an embedded one too) everything else; 4-bit run-length files are refused
MP4, MOV video and sound tracks with every sample, decoder settings, rotation, edit lists, color, HDR (HDR10, HLG, Dolby Vision), Apple spatial video, the scene illuminance iPhones show HDR video with, whether a video of 120 fps or more plays at full speed, track language codes location, device, software and dates, timed metadata tracks (GPS, motion, faces, Live Photo data), timecode and chapter tracks, maker data (GoPro serial numbers, Samsung SEF data), unused media data, trailing data
RAW, PDF, stereo photos, audio files not supported; RAW files built on TIFF (DNG, CR2, NEF...) and stereo JPEG and HEIC photos are recognized and refused

ICC profiles keep their color data. Their date becomes a fixed placeholder (2000-01-01), their description Clean, and device and creator fields, and reserved and vendor bits, are cleared. Images up to 268 megapixels are checked, enough for 200-megapixel phone photos; larger ones are refused, and so are animations (GIF, WebP, APNG, AVIF) on a canvas of more than 64 megapixels, or of more than four gigapixels in all their frames. Checking takes memory in proportion: a 200-megapixel JPEG takes about 10 seconds and 1.4 GB, a PNG of that size 3.5 GB, and a 268-megapixel WebP needs about 4 GB. Videos are cleaned in a copy next to the original, mapped rather than read into memory: a 4.7 GB video took about five seconds on an Apple-silicon Mac. Photos are read into memory whole, and a large GIF or PNG animation needs about five times its size. Fragmented and encrypted videos, videos with subtitles, Google's 360-degree videos, and codecs MagicDispel does not know are refused. Apple's 180- and 360-degree videos keep their projection.

File names

A cleaned copy is named after its original with _clean added, but without the dates, times and timestamps that screenshots, phone cameras and chat apps put in names:

Original Cleaned copy
Screenshot 2026-09-23 at 15.14.15.png Screenshot_clean.png
IMG_20240501_123456.jpg IMG_clean.jpg
mmexport1714567890123.jpg mmexport_clean.jpg
IMG_1234.HEIC IMG_1234_clean.HEIC
VID_20240501_123456.mp4 VID_clean.mp4

Recognized are dates with a four-digit year (2026-09-23, 20240501, 2024年5月1日), times with seconds (15.14.15, 123456) and Unix timestamps. The rest of a name stays: other styles of dates and times (23-09-26, May 1, 2024, 15.14 without seconds), words, places, coordinates and people's names. Rename a file first, or use --anonymous, when its name may say too much.

Further copies get _clean_1, _clean_2... --keep-name keeps the original name, dates and times included (_clean is still added), and --anonymous uses a random name such as photo_3f9c...e1.jpg or video_3f9c...e1.mov instead. A name too long for the file system is cut short at its end.

Usage

magicdispel photo.jpg "Screenshot 2026-09-23 at 15.14.15.png" portrait.heic IMG_1234.MOV
magicdispel --anonymous photo.jpg
magicdispel --check
Option
--anonymous name copies photo_<random>.jpg or video_<random>.mov instead of after the original
--keep-name keep the original's name, dates and times included
--check show the version, and whether ExifTool's second check is on
--version show the version
-h, --help show all options

Quote paths containing spaces, or drag the files into the terminal. A folder is not processed: give the files in it (for example *.jpg). A file whose name starts with - goes after --: magicdispel -- -photo.jpg. Exit codes: 0 success or help, 1 one or more files not cleaned, 2 invalid arguments, 130 interrupted. A failed file never stops the rest of a batch.

More ways to install

With uv or pipx and Python 3.10 or newer:

uv tool install magicdispel
pipx install magicdispel

The one-line installers are short scripts you can read first: install.sh and install.ps1. They install uv if needed, which brings its own Python when the system has none that fits. If magicdispel is not found afterwards, open a new terminal window.

To update, run the installer again, or uv tool upgrade magicdispel (pipx upgrade magicdispel). To remove MagicDispel, run uv tool uninstall magicdispel (pipx uninstall magicdispel).

ExifTool is optional. When version 12.73 or newer is installed, MagicDispel uses it to double-check every result: brew install exiftool, sudo apt install libimage-exiftool-perl (older Debian and Ubuntu releases carry a version too old to be used) or winget install --exact --id OliverBetz.ExifTool. If it is installed somewhere unusual, set MAGICDISPEL_EXIFTOOL to its full path. magicdispel --check says whether it is used.

Questions

Does it change my original photo or video? No. MagicDispel only reads the original. The clean copy is a new file next to it, and no existing file is ever overwritten. The original still holds all its metadata and sits next to the copy: share only the _clean file, and remember that a synced or backed-up folder holds both. While a photo or video is being cleaned, its copy is there as magicdispel-<random>.unfinished with the original's extension.

Does anything leave my computer? No. MagicDispel never uses the network. Only the installer downloads anything: uv, then MagicDispel with Pillow and, if the system has no fitting Python, a Python. It also runs uv tool update-shell, which adds uv's folder to your shell's startup files, unless MAGICDISPEL_NO_MODIFY_PATH=1 is set.

Why is a cleaned HEIC or video about as large as the original? HEIC files and videos are cleaned in place, so every offset inside them stays valid: removed data is overwritten with zeros rather than cut out. It is gone, but its space remains.

Why was my file refused? When a file cannot be rebuilt safely, MagicDispel says why and saves nothing rather than guess. The file may be damaged, use a structure MagicDispel does not know, or be a RAW photo: export RAW photos as JPEG or HEIC first.

Can I still edit portrait effects afterwards? No. Depth maps and portrait mattes are removed, so Photos can no longer change the portrait blur or lighting of the copy. Likewise, a cleaned video loses the data only its maker's app reads, such as the Live Photo data that pairs a video with its photo in Photos, or Samsung's slow-motion sections, so those apps may no longer offer such effects. Keep the original if you may want to edit it later.

Does it make me anonymous? No. What the picture shows can still identify people and places: faces, signs, reflections, the view from a window. So can a match with a copy published before, or the account it is shared from. MagicDispel removes metadata; it does not change the picture.

How do I report a problem? Open an issue and fill in the form, without attaching private photos. If a file keeps private data after cleaning, report it privately instead, as SECURITY.md explains.

Development

python -m venv .venv
# Activate the environment for your shell.
python -m pip install -e ".[test]" build
python -m unittest discover -s tests -v
python -m build

The tests need only Pillow. When ExifTool is installed, they also add metadata the way other programs write it and let ExifTool double-check each result. No personal photos or videos are included.

Before and after any change to the cleaning code, run the regression harness over a folder of real sample photos and videos kept outside the repository:

python scripts/make_probes.py ~/magicdispel-corpus    # adds synthetic leak probes
python scripts/regression.py ~/magicdispel-corpus     # saves a run under runs/
python scripts/regression.py ~/magicdispel-corpus --baseline ~/magicdispel-corpus/runs/<run>.json

It checks that every output looks identical to its input (Pillow, and macOS ImageIO/ColorSync when available; for videos, FFmpeg and macOS AVFoundation), that no probe marker survives, and, against a baseline, that no sample changes outcome or gains metadata. Pure refactors should also pass --identical, and --without-exiftool checks the path users without ExifTool take.

python scripts/fuzz.py 10000 --workers 8 --keep ~/fuzz-cases

damages small files of every format, and cleans them: a damaged file must be refused, and never crash the program, take more than eight seconds, or come out as a copy that cleans to something else. It reports what does, and saves those files.

.github/workflows/test.yml runs the tests on macOS, Windows and Linux with Python 3.10, 3.13 and 3.14, with and without ExifTool, once more with the oldest dependencies pyproject.toml allows, and the install scripts on all three. release.yml publishes a tagged version to PyPI once those pass. scripts/screenshot.py renders the images in this README.

License and attribution

MagicDispel is designed by VincentC and MIT licensed. ExifTool, an optional companion, is developed by Phil Harvey and distributed separately under its own license. MagicDispel is not affiliated with ExifTool.

Metadata

Release files for magicdispel 0.2.3

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for magicdispel 0.2.3
File Size Uploaded
magicdispel-0.2.3.tar.gz 217.7 kB Details

Built distribution (wheel)

Table of built distributions (wheels) for magicdispel 0.2.3
File Interpreter ABI Platform
magicdispel-0.2.3-py3-none-any.whl Python 3 none any Details

Total release size: 334.7 kB

Release files / magicdispel-0.2.3.tar.gz

Download URL magicdispel-0.2.3.tar.gz
Size 217.7 kB
Tags Source
SHA-256 checksum
How to use checksums
ef957c8f067127369f326562ae489fb877a8c32c82cac1041dc50c979249263b
BLAKE2b-256 checksum
How to use checksums
8eb556b0d4d6fe4de7f684386b8ceeaf26d8a3a090e3d8dff49b484c085691cf
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.14

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Sep 29, 2026.

Transparency log

Release files / magicdispel-0.2.3-py3-none-any.whl

Download URL magicdispel-0.2.3-py3-none-any.whl
Size 117.0 kB
Tags Python 3
SHA-256 checksum
How to use checksums
7290d8a07852fa2c72ad26897ad5ab1040468362de02f87631d0208c1ca6f38f
BLAKE2b-256 checksum
How to use checksums
d4e556ab4e7d2b8f0e9cd643f96f79834cd5b81bd7dfcbc0da0c73af6ad267b1
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.14

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Sep 29, 2026.

Transparency log

Release history Release notifications | RSS feed

0.2.4

2 release files

This release

0.2.3 This release

2 release files

0.2.2

2 release files

0.2.1

2 release files

0.2.0

2 release files

0.1.5

2 release files

0.1.4

2 release files

0.1.3

2 release files

0.1.2

2 release files

0.1.1

2 release files

0.1.0

2 release files

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page