A GitHub repo named reverse‑SynthID exposes a reverse‑engineering of Google’s Gemini image watermark, dubbed SynthID. By applying pure spectral analysis—no access to the encoder—the authors identified a resolution‑dependent carrier‑frequency pattern and built a detector that flags the hidden mark with 90% accuracy. The repository includes CLI tools for building the codebook and applying the bypass, inviting community scrutiny.

The team assembled a multi‑resolution spectral bypass (V3) that subtracts carrier energy directly in the FFT domain. Tests on 88 Gemini images at 1536×2816 resolution show a 75% drop in carrier energy, a 91% reduction in phase coherence, and a PSNR exceeding 43 dB, indicating near‑lossless visual quality after removal. A three‑pass iterative scheme—aggressive, moderate, gentle—captures residual energy, ensuring thorough removal.

Unlike brute‑force JPEG compression or noise injection, V3 relies on a SpectralCodebook that stores per‑resolution carrier fingerprints, auto‑selecting the closest profile at bypass time. Contributors are invited to supply pure black and white reference images from the Nano Banana Pro model to expand the codebook, a step that could tighten detection and improve cross‑resolution robustness. Profiles for 1024×1024 and 1536×2816 resolutions are included, demonstrating the approach’s scalability.