New GitHub tool strips AI watermarks from images, raising regulatory concerns
The 'remove-ai-watermarks' library targets markers from major platforms including Google Gemini and Adobe Firefly, though creators warn it cannot erase server-side records or guarantee legal compliance.
A new open-source repository hosted on GitHub, titled 'remove-ai-watermarks', has been released to strip both visible and invisible watermarks from images generated by artificial intelligence models. The tool provides a command-line interface and a library designed to remove markers from platforms including Google Gemini, ChatGPT, DALL-E, Stable Diffusion, Adobe Firefly, and Midjourney. It specifically targets SynthID, C2PA Content Credentials, and EXIF data, aiming to neutralise the provenance signals embedded in AI-generated media.
The utility addresses visible overlays, such as the sparkle logo added by Google Gemini, which is internally codenamed 'Nano Banana'. The tool utilises a Normalised Cross-Correlation detector to locate these markers dynamically, even if the image has been resized or cropped, before cleaning residual artifacts via gradient-masked inpainting. For invisible watermarks, the developers have implemented a diffusion-based regeneration pipeline. Since May 2026, the Stable Diffusion XL (SDXL) model serves as the default engine, chosen for its ability to defeat SynthID v2, which was deployed in October 2025.
To maintain image integrity during the regeneration process, the tool includes a 'Face Protection' mechanism. This feature uses YOLO to detect human faces, extracts them before the diffusion step, and blends them back in with a soft elliptical mask to prevent AI distortion. Additionally, an optional 'Analog Humanizer' can inject film grain and chromatic aberration, a technique intended to defeat AI-generated image classifiers by making the output appear indistinguishable from a photograph of a screen.
The repository notes that while the base installation handles visible watermarks and metadata stripping, removing invisible watermarks requires GPU dependencies for reasonable processing speed. A free web service, raiw.cc, is also available for users who prefer not to install the software locally. The tool parses image layers to remove AI-related fields while preserving standard metadata such as Author, Copyright, and Title, effectively cleaning the file for distribution.
Despite its technical capabilities, the developers issued a stark warning regarding the legal and practical limits of the tool. They clarified that while local markers can be removed, the tool cannot erase server-side records held by providers like Google. SynthID-Image v2 embeds a 136-bit payload believed to encode a user or session identifier, allowing providers to link watermarked files to generating accounts if the original file passed through their systems. The authors stressed that misrepresenting AI-generated content as human-created may violate laws regarding AI provenance, the DMCA, and platform terms of service.
