Download
Reverse-SynthID is a research-focused project that analyzes and reverse-engineers Google’s SynthID watermarking system used in AI-generated images. It leverages signal processing and spectral analysis techniques to identify hidden watermark patterns without access to proprietary encoding methods. The project introduces a multi-resolution “SpectralCodebook” that maps watermark characteristics across different image sizes. Using this approach, it can detect SynthID watermarks with high accuracy and selectively reduce or remove them through frequency-domain manipulation. Unlike traditional image degradation methods, it performs targeted, minimally invasive adjustments that preserve image quality. Overall, Reverse-SynthID serves as a technical exploration of AI watermark robustness, detection, and removal strategies.
Features
- Detects SynthID watermarks with ~90% accuracy using spectral and phase analysis.
- Multi-resolution SpectralCodebook enables adaptive watermark handling across image sizes.
- Advanced V3 bypass method reduces watermark signal while maintaining high image quality (43+ dB PSNR).
- Supports CLI and Python workflows for building codebooks, detection, and removal.
- Provides detailed FFT-based analysis tools for studying watermark behavior.
- Research-oriented framework for studying AI watermarking, security, and robustness.