I’ve been doing AI safety research on the robustness of digital watermarking for AI images, focusing on Google DeepMind’s SynthID (as used in Nano Banana Pro).
In my testing, I found that diffusion-based post-processing can disrupt SynthID in a way that makes common detection checks fail, while largely preserving the image’s visible content. I’ve documented before/after examples and detection screenshots showing the watermark being detected pre-processing and not detected after.
Why share this?
This is a responsible disclosure project. The goal is to move the conversation forward on how we can build truly robust watermarking that can't be scrubbed away by simple re-diffusion. I’m calling on the community to test these workflows and help develop more resilient detection methods.
Repo (writeup + artifacts): https://github.com/00quebec/Synthid-Bypass
Try the bypass for free: https://discord.gg/5mT7DyZu
I'd love to hear your thoughts!
[link] [comments]