SynthID-Bypass<span style="background-color:oklab(0.245984 0.00068689 -0.00564504);color:oklab(0.92807 -0.000171393 -0.00221816);font-family:&quot;gg sans&quot;, &quot;Noto Sans&quot;, &quot;Helvetica Neue&quot;, Helvetica, Arial, sans-serif;font-size:16px"> is an AI safety research project demonstrating a vulnerability in Google's SynthID—the invisible digital watermark embedded in images generated by Gemini (Nano Banana Pro) to identify them as AI-generated. --- The Problem:<span style="background-color:oklab(0.245984 0.00068689 -0.00564504);color:oklab(0.92807 -0.000171393 -0.00221816);font-family:&quot;gg sans&quot;, &quot;Noto Sans&quot;, &quot;Helvetica Neue&quot;, Helvetica, Arial, sans-serif;font-size:16px"> Google's SynthID watermark can be removed. The Method:<span style="background-color:oklab(0.245984 0.00068689 -0.00564504);color:oklab(0.92807 -0.000171393 -0.00221816);font-family:&quot;gg sans&quot;, &quot;Noto Sans&quot;, &quot;Helvetica Neue&quot;, Helvetica, Arial, sans-serif;font-size:16px"> Two ComfyUI workflows that "re-render" watermarked images through a diffusion model pipeline using: <ul><li>Low-denoise regeneration</li><li>ControlNet structural guidance</li><li>Multi-pass processing</li></ul><span style="background-color:oklab(0.245984 0.00068689 -0.00564504);color:oklab(0.92807 -0.000171393 -0.00221816);font-family:&quot;gg sans&quot;, &quot;Noto Sans&quot;, &quot;Helvetica Neue&quot;, Helvetica, Arial, sans-serif;font-size:16px"> The Result:<span style="background-color:oklab(0.245984 0.00068689 -0.00564504);color:oklab(0.92807 -0.000171393 -0.00221816);font-family:&quot;gg sans&quot;, &quot;Noto Sans&quot;, &quot;Helvetica Neue&quot;, Helvetica, Arial, sans-serif;font-size:16px"> Processed images pass Google's own SynthID detection as "not AI-generated" while preserving visual content. --- Why Release This?<span style="background-color:oklab(0.245984 0.00068689 -0.00564504);color:oklab(0.92807 -0.000171393 -0.00221816);font-family:&quot;gg sans&quot;, &quot;Noto Sans&quot;, &quot;Helvetica Neue&quot;, Helvetica, Arial, sans-serif;font-size:16px"> This is responsible disclosure research. By exposing the weakness, we aim to: <ol><li>Highlight limitations in current AI content authentication</li><li>Encourage development of more robust watermarking techniques</li><li>Invite the community to create bypass-resistant watermarks</li></ol>

Files with SHA256: 97088a9fcb13841e21abedecbdc4ef5bc527c75362678ede16e91723aa7a2db9