wiltodelta/remove-ai-watermarks

AI watermark remover. CLI and Python library to strip visible and invisible AI watermarks (Gemini / Nano Banana sparkle, SynthID) and provenance metadata (C2PA, EXIF, IPTC) from images.

What it solves

It addresses the presence of AI-generated identifiers in images, including visible vendor labels (like the Gemini sparkle), invisible pixel watermarks, and embedded metadata (such as C2PA, EXIF, and XMP).

How it works

The tool employs different strategies depending on the target watermark:

  • Visible marks: Detects the label, creates a mask, and fills the area using OpenCV, MI-GAN, or LaMa.
  • Metadata: Uses format-aware stripping to remove provenance signals without recompressing the image.
  • Invisible watermarks: Regenerates the entire image through a diffusion pipeline to disrupt pixel and frequency domain watermarks.

Who it’s for

Users who want to remove provenance signals from images they have generated themselves for lawful use.

Highlights

  • Supports detection and removal of visible marks from providers like Google Gemini, Samsung Galaxy AI, and others.
  • Offers multiple backends for inpainting, including OpenCV, MI-GAN, and LaMa.
  • Provides a high-fidelity diffusion-based mode (qwen-zimage) for complex images.
  • Includes a CLI, Python API, and ComfyUI nodes.