image-sanitize

name	image-sanitize
description	Remove all metadata (EXIF, IPTC, XMP, C2PA) and visible AI watermarks from user-provided images. Trigger this skill when a user asks to clean images or remove metadata, content credentials, or AI watermarks. Supports JPEG, PNG, TIFF, HEIC and AVIF files.

This skill accepts one or more image files from the user and returns cleaned versions of those images with metadata removed and visible AI watermarks eliminated. It uses the provided Python script clean_images.py located in this directory. The workflow is as follows:

1. Collect the image files provided by the user. Accept common raster formats such as JPEG, PNG, TIFF, HEIC and AVIF.

2. For each file, run the script:

   python clean_images.py --input <path/to/image1> <path/to/image2> --output-dir <output-dir> --remove-watermark
   

   • The script will strip all EXIF, IPTC, XMP, and C2PA metadata from the images.
   • When the --remove-watermark flag is passed, it attempts to remove visible AI watermarks (e.g., the Gemini sparkle). The current implementation includes a placeholder for watermark removal logic; you can extend it to implement reverse‑alpha blending or integrate third-party libraries.

3. The script writes cleaned files into the specified output directory. Each cleaned file is named cleaned-<original-file-name>.

4. Return the cleaned images to the user.

This skill is designed for Codex and WebGPT agents. It encapsulates a reusable workflow so the agent doesn’t need to rewrite the image-cleaning logic each time. You can extend or replace clean_images.py to improve watermark removal.