reverse engineering / synthid / gemini

A GitHub project posted to Hacker News claims to discover, detect, and remove Google’s SynthID AI watermark, offering a pip-installable CLI with settings like “aggressive” and “maximum.” Commenters praise the underlying observations about resolution-dependent carriers and multi-resolution watermark extraction but criticize the repo’s packaging, lack of rigor, missing methodology for the “90% detection rate,” absent before/after images, no CI/tests, and unstable imports. Some note that downscaling/upscaling or black/white canvases may evade the watermark, while others flag ethical concerns about publishing turnkey watermark-stripping tools. The discussion matters because it touches on provenance, model accountability, and the cat-and-mouse dynamics between watermarking and removal in AI-generated content.

Researchers reverse-engineered Google’s SynthID image watermark used by Gemini, discovering a resolution-dependent spectral carrier structure and a model-level fixed phase template. Using black/white reference images and spectral analysis (no proprietary access), they built a detector with ~90% accuracy and developed a V3 multi-resolution spectral codebook bypass that surgically subtracts carriers in the FFT domain. V3 achieves ~75% carrier energy reduction, a 91% drop in phase coherence, and >43 dB PSNR, outperforming crude methods like JPEG or noise transforms. The project highlights that carriers shift with image resolution and that the green channel is dominant, and it requests more black/white Nano Banana Pro outputs to expand its codebook.

Researchers reverse-engineered Google’s SynthID image watermark used by Gemini, revealing a resolution-dependent spectral carrier structure and a consistent phase template tied to model-level keys. Using black/white reference images and spectral analysis, they built a detector with ~90% accuracy and developed a V3 multi-resolution spectral codebook bypass that surgically removes watermark carriers via FFT-domain subtraction, achieving ~91% phase coherence drop, 75% carrier energy reduction, and >43 dB PSNR. The team emphasizes per-resolution fingerprints—carriers shift with image size—so a single-profile codebook cannot generalize; V3 auto-selects profiles or resizes as fallback. This work exposes how invisible watermarks are embedded and demonstrates practical detection and removal tools, raising implications for provenance, content moderation, and watermark robustness.

aloshdenny/reverse-SynthID: reverse engineering Gemini's SynthID detection