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

A research project that reverse-engineers Google's SynthID invisible watermarking system used in Gemini-generated images. Using only signal processing and spectral analysis (no access to the proprietary encoder/decoder), the project discovers the watermark's resolution-dependent carrier frequency structure, builds a detector with 90% accuracy, and develops a multi-resolution spectral bypass (V3) achieving 91% phase coherence drop and 43+ dB PSNR. The V3 approach uses a SpectralCodebook storing per-resolution watermark fingerprints, enabling surgical frequency-bin-level removal on any image size — far outperforming brute-force methods like JPEG compression.