Reverse-synthid is a Python tool for reverse-engineering and bypassing Google's SynthID watermarking system embedded in Gemini-generated images.
The project addresses the challenge of detecting and removing an imperceptible watermark that Google encodes directly into pixel values of images generated by Gemini. Rather than requiring access to the proprietary encoder or decoder, the approach uses signal processing and spectral analysis to identify the watermark's resolution-dependent carrier frequency structure. The tool discovers the diagonal banding pattern that forms the watermark's spatial frequency signature and targets it through multi-resolution spectral analysis. A detector identifies SynthID watermarks by analyzing these frequency characteristics, while a bypass mechanism surgically removes the watermark by reducing carrier energy and phase coherence while maintaining image quality.
Developers should adopt this tool if they need to detect whether an image contains SynthID watermarking or if they require the ability to remove the watermark from Gemini-generated images while preserving visual fidelity. The project includes both command-line utilities and a drag-and-drop desktop application, making it accessible without requiring terminal expertise after initial setup. The tool supports multiple Gemini models and handles images across different resolutions and color backgrounds through per-model profiling and cross-color phase consensus.
The project demonstrates sustained adversarial development across multiple rounds of iterative refinement, with each round addressing newly documented failure modes in the detection system. The codebase incorporates human-in-the-loop calibration that tunes watermark removal strength based on manual detection tallies from the Gemini application. Development has progressed through a unified seven-stage attack pipeline that targets every documented SynthID failure mode simultaneously, achieving visually lossless output across both tested Gemini models.
