remove-ai-watermarks: a comprehensive utility for stripping visible overlays, invisible frequency-domain watermarks, and AI provenance metadata from images

What it solves

This tool removes both visible and invisible AI-generated watermarks and provenance metadata from images. It is designed to give users autonomy over content they generated themselves, stripping labels from models like Google Gemini, DALL-E, Stable Diffusion, and Adobe Firefly without needing to rely on stock-photo removal tools.

How it works

The project employs different techniques based on the type of watermark:

• Visible Watermarks: For known marks (like the Gemini sparkle or Samsung Galaxy AI labels), it uses reverse-alpha blending. This process recovers original pixels using a captured alpha map rather than guessing via inpainting. For arbitrary logos, it provides a universal region eraser using cv2 or big-LaMa inpainting.
• Invisible Watermarks: To remove patterns like SynthID, StableSignature, and TreeRing, the tool uses diffusion-based regeneration. It encodes the image into latent space, adds noise, and denoises it using SDXL or Qwen-Image pipelines. A Canny ControlNet is used by default to preserve the structure of faces and text during this process.
• Metadata: It strips C2PA manifests, EXIF/XMP labels, and other AI-disclosure tags from various file formats (PNG, JPEG, AVIF, HEIF, MP4, etc.).

Who it’s for

Users who generate AI images and want to remove the platform-specific branding, invisible tracking markers, or metadata tags from their own outputs for a cleaner final result.

Highlights

• Multi-model support: Targets watermarks from Gemini, DALL-E 3, Stable Diffusion, FLUX, Adobe Firefly, and more.
• Structure Preservation: Uses ControlNet to keep text and face structures sharp while removing invisible watermarks.
• Comprehensive Detection: Includes an identify command to report the origin platform and a full inventory of detected watermarks.
• Flexible Pipelines: Offers multiple regeneration pipelines including SDXL and an experimental Qwen-Image (20B) for better text preservation.
• Cross-platform: Works offline via Python or through a cloud-based web interface.