wiltodelta · 2026-05-18 · major
Remove-AI-Watermarks Hits the HN Front Page — MIT-Licensed CLI Strips Visible Gemini Watermarks Plus Invisible SynthID, C2PA, and EXIF Provenance
Day after OpenAI added SynthID to its image API, a Python CLI for stripping Gemini, SynthID, C2PA, and EXIF watermarks hits the HN front page at 292 points. README ships a regulatory warning for EU/US provenance laws.
An MIT-licensed library lifts every major AI provenance marker — and lands on the HN front page the same week OpenAI adopts SynthID.
Key specs
| License | MIT |
|---|---|
| GitHub stars | 675 |
| Hn points | 292 |
What is it?
A Python CLI and library that targets the four common provenance layers for AI-generated images: Google Gemini's visible logo, Google DeepMind's invisible SynthID pixel signal, C2PA Content Credentials metadata, and EXIF tags. Released v0.4.1 on May 17, hit Hacker News with 292 points on May 19.
How does it work?
For visible Gemini watermarks, a three-stage normalized cross-correlation detector finds the watermark's position and scale, reverses the alpha-blend against a known logo map, and uses gradient-masked inpainting to clean residual edges. For invisible markers, diffusion-based regeneration in pixel or frequency space disrupts the watermark; metadata layers (C2PA, EXIF) are stripped or rewritten.
Why does it matter?
The repo lands the same week OpenAI rolled SynthID and C2PA into its image API and Google expanded watermarking across SynthID and C2PA at I/O. It is the live counter-experiment for that policy push and a stress test for whether AI provenance markers can be enforced. The EU AI Act and several US states criminalise removal-with-intent-to-deceive; the README leads with legal disclaimers.
Who is it for?
Provenance researchers, journalists, policy analysts, red teamers
Try it
pipx install remove-ai-watermarks