aloshdenny/reverse-SynthID

reverse engineering Gemini's SynthID detection

What it solves

This project reverse-engineers Google's SynthID, an invisible watermark embedded in images generated by Google Gemini. It provides tools to detect these watermarks and surgically remove them without introducing visible distortion to the image.

How it works

The project uses signal processing and spectral analysis to identify the watermark's carrier frequency structure. It employs a few different versions of removal techniques:

  • V3 Bypass: Uses a multi-resolution spectral codebook to perform FFT-domain subtraction of the watermark signal.
  • V4 Round-06 Pipeline: A comprehensive 7-stage attack that combines a Stable Diffusion VAE round-trip, elastic deformation (random warp fields), global geometric transforms, resize-squeezing, color-contrast nudges, residual-phase FFT subtraction, and a JPEG compression chain.
  • Cross-Color Consensus: The V4 system builds a codebook by analyzing phase consistency across multiple solid-color backgrounds to isolate the watermark from actual image content.

Who it’s for

Researchers and developers interested in AI watermarking, adversarial attacks on image detection systems, and users who want to remove SynthID watermarks from Gemini-generated images.

Highlights

  • High Accuracy Detection: Identifies SynthID watermarks with 90% accuracy.
  • Multi-Model Support: Includes specific profiles for gemini-3.1-flash-image-preview and nano-banana-pro-preview.
  • Visually Lossless Removal: The Round-06 pipeline is designed to defeat the Gemini app's detector while maintaining high image fidelity.
  • Human-in-the-Loop Calibration: Allows updating the removal strength based on manual detection feedback from the Gemini app.

Related

  • Dispatch
    Google Gemini AI Image Verification

    Google is integrating SynthID watermarking technology into the Gemini app to allow users to verify if images were generated or edited by Google AI.

  • Project
    allenk/GeminiWatermarkTool

    A cross‑platform, portable executable that removes Google Gemini image (and video) watermarks using a deterministic reverse‑alpha‑blending algorithm, with optional GPU‑accelerated FDnCNN denoising, GUI/CLI interfaces, batch processing, and AI‑agent integration via MCP.

  • Project
    wiltodelta/remove-ai-watermarks

    A library and CLI tool to remove visible labels, invisible pixel watermarks, and AI provenance metadata from AI-generated images and videos.

  • Dispatch
    Google DeepMind Content Transparency and Verification Tools Update

    Google DeepMind is expanding SynthID watermarking and C2PA Content Credentials across Search, Gemini, Chrome, and Pixel to help users identify AI-generated and authentic camera-captured content.

  • Dispatch
    SynthID Text Integration in Transformers v4.46.0

    Google DeepMind and Hugging Face have integrated SynthID Text into Transformers v4.46.0, providing a method to apply imperceptible watermarks to AI-generated text for detection via trained classifiers.