Hacker News Debates a Repo Claiming to Reverse Gemini's SynthID

What happened

A Hacker News submission that reached 100 points and 42 comments highlighted reverse-SynthID, a repository that claims to reverse-engineer Google's SynthID watermarking on Gemini-generated images. The project says it works without access to Google's proprietary encoder or decoder, relying instead on signal processing and spectral analysis. The author frames the work as a three-step result: discover the watermark structure, build a detector, and create a multi-resolution bypass that reduces the watermark while preserving image quality.

The repo's headline numbers are what made it spread quickly. It claims 90% detection accuracy, roughly a 75% drop in carrier energy, more than a 91% drop in phase coherence, and 43+ dB PSNR for the strongest bypass path. Another notable claim is that SynthID appears resolution-dependent, with different carrier positions at different image sizes, so the bypass needs a resolution-specific codebook rather than one universal fingerprint. That makes the project sound less like crude denoising and more like targeted watermark removal in the frequency domain.

Why Hacker News pushed back

The discussion was split between people who saw this as useful red teaming and people who questioned the validation. Supporters argued that if a watermark can be removed this easily, it was never a strong provenance mechanism to begin with, and public breakage is exactly how weak designs get exposed. Critics focused on a narrower technical point: the repo evaluates removal against its own detector, not against Google's official SynthID detector. On Hacker News, that missing ground truth was the main reason some readers called the result incomplete rather than conclusive.

The thread also broadened into a larger argument about provenance systems. Several commenters argued that invisible watermarks are too fragile to serve as a reliable authenticity layer once motivated users can route around them with alternate models, transformations, or dedicated removal tools. Others suggested that if provenance matters, it may need to look more like signed capture pipelines than platform-specific hidden marks inside pixels. In other words, the repo became a prompt for a wider discussion about what watermarking can realistically guarantee.

For Insights readers, the important part is not just whether this specific repository fully proves its case. It is that reverse-engineering pressure is already arriving for closed image stacks, and the community is treating watermark robustness as something that must survive adversarial testing, not marketing claims. Original discussion: Hacker News. Original source: GitHub repo.