A researcher has successfully reverse-engineered Google's SynthID watermarking system, which embeds invisible markers in all Gemini-generated images. GitHub user aloshdenny documented the effort in a public repository, using only signal processing and spectral analysis to identify the watermark's structure and develop removal methods—without accessing Google's proprietary encoder/decoder systems.
SynthID Watermark Uses Resolution-Dependent Carrier Frequencies
The research revealed that SynthID embeds carrier frequencies at different absolute positions depending on image dimensions. At 1024×1024 resolution, carriers appear at low frequencies like (9,9), while at 1536×2816 they shift to higher frequencies around (768,704). This resolution-dependent structure means a codebook built for one size cannot directly remove watermarks from images of different dimensions.
Phase Template Remains Identical Across All Gemini Images
The watermark's phase template stays consistent across all images generated by the same Gemini model, with cross-image phase coherence exceeding 99.5%. The green channel carries the strongest signal. Researchers validated true carriers through black/white image cross-validation, where phase agreement indicated authentic watermark components versus generation artifacts. As one observer noted, "The SynthID watermark isn't user-specific—it's a fixed, model-level pattern baked into every Gemini image."
V3 System Achieves 91.4% Phase Coherence Drop With Minimal Quality Loss
The project evolved through three bypass generations. V1 used JPEG compression at quality 50 (37 dB PSNR), V2 employed multi-stage transforms combining noise and color modifications (27–37 dB), and V3 implemented multi-resolution spectral codebook subtraction as the current best approach.
The V3 system stores per-resolution watermark fingerprints in a single codebook file. During bypass, it auto-selects the matching resolution profile and performs frequency-bin-level subtraction using direct known-signal removal. Testing on 88 Gemini images at 1536×2816 resolution produced:
- 43.5 dB PSNR
- 0.997 SSIM
- 75.8% carrier energy reduction
- 91.4% phase coherence drop across top-5 carriers
Quality remained consistent across resolutions: at 768×1024 (fallback mode), the system delivered 40.6 dB PSNR with 0.994 SSIM.
Direct Subtraction Method Employs Confidence-Weighted Removal
V3 employs confidence-weighted direct subtraction where each frequency bin's watermark removal equals the estimated watermark magnitude multiplied by phase consistency, cross-validation agreement, removal fraction, and per-channel weights (green=1.0, red=0.85, blue=0.70). A safety cap prevents subtraction from exceeding 90–95% of any frequency bin's energy.
The Hacker News post reached 99 points with 42 comments on April 9, 2026. This work demonstrates that spread-spectrum watermarks, while imperceptible, can be systematically identified and removed through spectral analysis and empirical codebook building.
Key Takeaways
- Google's SynthID watermark uses a fixed, model-level phase template that remains identical across all images from the same Gemini model with 99.5% cross-image coherence
- Watermark carriers shift to different frequency positions depending on image resolution, requiring resolution-specific codebooks for removal
- The V3 removal system achieved 91.4% phase coherence reduction with 43.5 dB PSNR and 0.997 SSIM on 1536×2816 images
- Researchers used only signal processing and spectral analysis without access to Google's proprietary systems
- The green color channel carries the strongest watermark signal, followed by red (0.85) and blue (0.70)