cebeuq/Synthid-Bypass

What it solves

This project provides a proof-of-concept method to remove digital watermarks from AI-generated images, specifically targeting Google's SynthID. It addresses the challenge of identifying synthetic media by demonstrating how these watermarks can be bypassed using controlled image reconstruction.

How it works

The project uses a ComfyUI workflow that treats the watermark as low-level image noise. It removes this noise by re-processing the image through a diffusion model while keeping the original composition intact. The process involves:

• Global Redraw: A low-denoise pass using Qwen Image models and a Canny edge map to lock the image structure.
• Adaptive Denoising: A custom node that calculates the appropriate denoise level based on the image resolution.
• Face Reconstruction: A specialized path using Z-Image Turbo and strict segmentation (Ultralytics, MediaPipe, and SAM) to ensure human faces remain virtually identical to the original while still shifting pixels enough to break the watermark detection.

Who it’s for

This tool is designed for AI safety researchers and those studying the robustness of digital watermarking technologies.

Highlights

• Integrated Workflow: Consolidates global reconstruction and face restoration into a single ComfyUI graph.
• Structural Guidance: Uses Canny controlnets to prevent the image from changing its composition during the redraw.
• Adaptive Controls: Includes custom nodes for resolution-aware denoising and adaptive face detailing.
• Research-Based Approach: Grounded in the hypothesis that pixel-space watermarks are fragile to diffusion-based reconstruction attacks.