Summary

This project, SynthID-Bypass, demonstrates a proof-of-concept bypass for Google's SynthID digital watermarking technology on AI-generated images. It enables the removal of SynthID watermarks from models like Nano Banana Pro, aiming to encourage the development of more robust AI media detection methods by exposing current vulnerabilities.

How It Works

The bypass re-processes watermarked images through a diffusion model with a very low denoising factor. This "re-renders" the image, preserving visual content and structure via ControlNets while replacing embedded watermark noise with new diffusion model noise. Specialized workflows include advanced face detection and inpainting for portraits to maintain fidelity.

Quick Start & Requirements

Requires a functional ComfyUI installation. Users must install specific custom nodes (e.g., ComfyUI Impact Pack, rgthree-comfy) and download several models (e.g., ae.safetensors, z_image_turbo_bf16.safetensors, yolov8n-face.pt). A capable GPU with >16GB VRAM is necessary. Workflows are loaded as .json files onto the ComfyUI canvas. A Discord community is available at https://discord.gg/R7skesPZ.

Highlighted Details

• Successfully demonstrates disruption and removal of Google's SynthID watermark.
• Employs low-denoise regeneration guided by ControlNets for structural integrity.
• Features specialized ComfyUI workflows for portrait images, enhancing face preservation.
• Visualized the watermark pattern, revealing its non-deterministic nature.

Maintenance & Community

Encourages community collaboration via GitHub issues/pull requests for testing and developing better techniques. A Discord server (https://discord.gg/R7skesPZ) is provided for discussion. No specific maintainers or roadmaps are detailed.

Licensing & Compatibility

Released strictly for educational and research purposes. No explicit open-source license is specified, implying potential restrictions on commercial use. Users are asked to cite the repository if used in research.

Limitations & Caveats

Requires significant computational resources (GPU >16GB VRAM) and ComfyUI familiarity. High-resolution images (>4K) may need downscaling or increased denoising. The technique can introduce subtle visual artifacts and may result in detail loss in certain image regions.