Un-banana-ing AI images with reverse-alpha and a diffusion bath

What it does remove-ai-watermarks is a CLI and Python library that scrubs AI-generated images of their origin markers. It reverses known visible watermarks—like the Google Gemini “Nano Banana” sparkle and the Doubao corner text—by subtracting captured alpha maps instead of guessing with inpainting. For invisible frequency-domain marks such as SynthID, StableSignature, and TreeRing, it runs the image through a controlled SDXL diffusion regeneration that breaks the watermark while attempting to preserve the picture. It also strips C2PA manifests, EXIF/XMP tags, PNG text chunks, and the IPTC “Made with AI” label from a wide range of image and video formats.

The interesting bit The visible-watermark engine does not hallucinate replacements; it literally reverses the alpha-blending math using alpha maps extracted from controlled black-background captures. A three-stage normalized-cross-correlation detector finds the mark even after cropping or resizing, and the tool maintains per-service registries for Gemini, Doubao, and Jimeng wordmarks. That is a much more surgical approach than generic inpainting.

Key highlights

• Deterministic visible-watermark removal for Gemini/Nano Banana, Doubao, and Jimeng marks runs on CPU in roughly 0.05 s per image.
• Invisible-watermark removal uses an SDXL diffusion pipeline at native resolution to defeat SynthID v1/v2 and similar schemes; it includes optional YOLO-based face protection to avoid distorting people during regeneration.
• Metadata stripping covers C2PA (including video containers like MP4/MOV and audio via ffmpeg), EXIF, XMP, PNG chunks, and the China TC260 AIGC label.
• The identify command aggregates a surprising breadth of signals—C2PA issuer, forensic watermark vendor, embedded generation parameters, HuggingFace job IDs, and xAI/Grok EXIF signatures—into a single JSON verdict.
• A hosted cloud version at raiw.cc offers the GPU-heavy invisible-watermark removal without local setup.

Caveats

• Defeating the SynthID verifier is not the same as being forensically clean; the README cites recent research that diffusion-based removal leaves detectable traces, and a separate “this image was processed” classifier could still flag the output.
• Some jurisdictions restrict removing AI labels as such, and the authors explicitly disclaim liability for downstream misuse.
• The Jimeng and Doubao marks re-rasterize slightly per image, so reverse-alpha is followed by a thin residual inpaint to clear leftover glyph edges.

Verdict Worth a look if you need to sanitize AI-generated assets for legitimate reuse or audit provenance metadata. Skip it if you are looking for a guaranteed forensic clean-room result; the tool is honest about the traces it leaves.