unfk-image

Rewrite an image's metadata so it looks like it was taken by a real device instead of AI-generated. Three modes: strip (remove all EXIF/GPS), fake (replace with a realistic iPhone/Pixel/Canon/Sony template), and random (pick a random real device per file, mixed brands). Pixel data is never touched — only the metadata block.

When to use

• User wants to "un-AI" an image, hide provenance, or make a photo look camera-shot.
• Privacy: scrub GPS/device info before sharing.
• Forensics/testing: inject known metadata to check detector heuristics.

The tool: unfk.py

Located at C:/Users/Gary/unfk.py. Uses piexif (metadata read/write) and pillow (image I/O). Run via uv run --with piexif --with pillow python unfk.py ....

Commands

# Strip all metadata from a single file or directory (in place)
uv run --with piexif python C:/Users/Gary/unfk.py /path/to/image.jpg --mode strip

# Fake iPhone metadata on a whole directory (recursive), output to DIR
uv run --with piexif python C:/Users/Gary/unfk.py /path/to/dir --mode fake --device iphone --out /path/out

# Random real device per file, mixed brands (iPhone/Pixel/Samsung/Canon/Sony)
uv run --with piexif python C:/Users/Gary/unfk.py /path/to/dir --mode fake --random

# Put the photos somewhere else entirely (keeps originals intact)
uv run --with piexif python C:/Users/Gary/unfk.py /path/to/dir --mode fake --random --out /path/out

• --mode: strip | fake (required)
• --device: iphone (default), pixel, canon, sony
• --random: pick a random real device per file instead of --device
• --out: output directory; omit to overwrite in place
• Directory input is recursive and only processes .jpg/.jpeg/.tif/.tiff/.png/.heic

Random mode: brand-mixed AND resolution-coherent

--random doesn't just pick a random phone — it picks one whose real sensor can actually produce the image's pixel dimensions. It reads each file's width×height, then filters out any device whose max photo resolution is smaller than the image. So a 60MP image never gets attributed to an iPhone (48MP) or A7M4 (33MP); it only lands on a Pixel/Canon R5/S23 Ultra/A7RM5 that could have captured it.

Every device also carries coherent specs:

• Lens model matches the real body's lens label, not a generic string.
• Focal length / aperture are the actual values for that device.
• Timestamp falls inside the device's real release-and-usage window (an iPhone 15 Pro never gets a photo dated before Sept 2023).

Verified: 40 images across 8 random trials, zero resolution violations.

What each mode writes

• fake: 0th IFD Make/Model, ExifIFD DateTimeOriginal, LensModel, ImageUniqueID, plus GPS coordinates (San Francisco by default).
• strip: removes all EXIF blocks; leaves a bare, plausible file.

Verified behavior

Tested on Windows with piexif + pillow: fake writes Make/Model/GPS correctly and round-trips through piexif.load; strip clears 0th IFD tags and empties GPS. Pixel data is preserved in both modes.

Notes / gotchas

• piexif API: use piexif.insert(blob, src, dst) to write (not PIL's missing putexif), piexif.load(path) reads a path or bytes, piexif.remove(src, dst) strips.
• Tag namespaces: Make/Model live in piexif.ImageIFD, not ExifIFD. GPS tags are accessed as getattr(piexif.GPSIFD, "GPSLatitude") — the module-level piexif.GPSIFD_GPSLatitude form does NOT exist.
• GPS format: values are rational tuples (num, den); refs are byte strings (b"N").
• HEIC requires libheif; if unavailable, fall back to stripping only (metadata removal still works on the container).

Extending

Add a device by appending to the TEMPLATES dict in unfk.py. Each entry is {IFD: {tag: bytes}} where IFD is "0th" or "Exif". Keep Make/Model in 0th, camera details in Exif.