Reverse-Engineering Gemini's SynthID Watermark Detection

This project reverse-engineered the SynthID watermark embedded in Google Gemini images using only signal processing and spectral analysis, without a dedicated encoder or decoder.

The key finding is that the watermark has a resolution-dependent carrier frequency structure. At 1024x1024, low-frequency grids such as (9, 9) and (5, 5) appeared strongly, while at 1536x2816, the carrier was captured in completely different bins such as (768, 704) and (672, 1056).

Images generated by the same Gemini model showed nearly identical phase templates, and the green channel carried the strongest signal. By cross-validating black/white reference images against actual watermarked images to isolate the true carrier, the project claims to have built a detector with about 90% detection accuracy.

On the circumvention side, the V3 multi-resolution SpectralCodebook is the core component.

• It stores separate profiles per resolution
• automatically selects the profile matching the input image's size
• and removes the watermark by directly subtracting the known signal in the FFT domain.

As results, it presents a carrier energy drop of 75.8%, phase coherence drop of 91.4%, PSNR of 43.5 dB, and SSIM of 0.997. It explains that this causes less quality loss than brute-force methods such as JPEG compression or adding noise.