Python CLI/library strips visible Gemini sparkle logos, invisible SynthID/StableSignature/TreeRing watermarks, and C2PA/EXIF/XMP metadata from AI-generated images in one command.
Key Takeaways
Visible Gemini “Nano Banana” sparkle is removed via reverse alpha blending in ~0.05s, CPU-only, no model weights needed.
Invisible watermarks (SynthID v1/v2, StableSignature, TreeRing) are defeated by diffusion regeneration using SDXL at ~1024px native resolution; GPU strongly recommended.
Metadata stripping covers C2PA manifests, XMP DigitalSourceType, EXIF, and PNG text chunks – the tags that trigger “Made with AI” labels on Instagram, Facebook, and X.
Analog Humanizer adds film grain and chromatic aberration to bypass AI image classifiers; Smart Face Protection uses YOLO to preserve facial detail through diffusion.
Legal section flags the US COPIED Act (enacted 2025) and EU AI Act Article 50(2) (effective Dec 2026): removing provenance markers with intent to deceive carries legal risk.
Hacker News Comment Review
Commenters challenged the “false-positive AI label” use-case justification: watermark insertion requires the image to have passed through a generator, making false positives on human art unlikely without deliberate AI edits.
A noted caveat: diffusion regeneration at SDXL’s 1024px cap degrades fine detail and cannot cleanly handle 4K outputs from Gemini Nano Banana Pro or GPT Image 2.
Privacy-framing debate emerged: one camp sees watermark removal as a legitimate anti-surveillance stance against session-identifier payloads embedded in SynthID v2’s 136-bit encoding.
Notable Comments
@Tiberium: points out that asking Gemini or ChatGPT to make a minor edit to a real photo can cause the output to test positive for AI watermarks, partially rescuing the false-positive argument.
