Tool removes watermarks and AI metadata from generated images
Original: Remove AI Watermarks
Source: github.com ↗
Who: Posted by GitHub user janalsncm on Hacker News, linking to a public repository authored by wiltodelta, an independent developer with no further public attribution available.
What's new: The remove-ai-watermarks library is an open-source tool that strips both visible and invisible watermarks from AI-generated images, along with the embedded metadata that causes social platforms like Instagram and X to display "Made with AI" labels. It targets output from Google Gemini, OpenAI's image generators, Stable Diffusion, Adobe Firefly, Midjourney, and others. A free web interface is available at raiw.cc for users who prefer not to install anything locally.
How it works: The tool operates in three distinct layers. For visible watermarks — currently only Google Gemini's sparkle logo — it uses reversal: since the watermark is stamped by mixing the logo with the original at a known opacity, the original pixels can be mathematically recovered. A detector locates the watermark position before reversal, and gradient-masked inpainting cleans up any remaining edge artifacts.
For invisible watermarks — including , , and — the pipeline re-encodes the image through a diffusion model at roughly 1024 pixels, adds a small controlled amount of noise, then denoises it, which destroys the hidden frequency-domain signal while preserving visible content. is the current default backbone for this step after testing showed older pipelines failed against newer SynthID versions. The tool also uses to extract faces before diffusion and blend them back afterward to prevent facial distortion.
The metadata layer strips tags, fields, PNG text chunks, and manifests — the last of which carries cryptographic provenance records used by platforms to verify AI origin.
Caveats: The tool openly acknowledges it cannot yet handle a pixel-level watermark embedded by OpenAI's newest image generator, gpt-image-2, because no public detector exists for it. Diffusion-based regeneration also degrades image quality to some degree, and invisible watermark removal requires a GPU and a roughly 2 GB model download. The library's existence raises straightforward content-authenticity concerns: was specifically designed to make provenance tamper-evident, and this tool demonstrates that stripping it is a one