SynthID vs C2PA: The Two AI Watermarks Explained

People often think they’ve removed an AI watermark when they’ve only removed half of it. That’s because there are two completely different things at play — SynthID and C2PA — and they hide in different places.

SynthID: in the pixels

SynthIDis Google DeepMind’s watermark. It’s an invisible pattern baked into the pixels themselves, designed to survive cropping, compression and re-saving. You can’t strip it by editing metadata — the only way to disrupt it is to rebuild the image so the pattern is replaced.

C2PA: in the metadata

C2PA “Content Credentials”is a signed record of the image’s origin stored in the file’s metadata, alongside EXIF/XMP tags. It’s attached by OpenAI, Google and others. Because it’s metadata, it’s stripped losslessly — the pixels don’t change at all.

Why the difference matters

• Removing C2PA does not touch SynthID, and vice versa.
• Metadata removal is pixel-perfect; hidden-mark removal returns a faithful full-resolution copy.
• A Gemini image can carry both — plus a visible sparkle — so a full clean means handling all three.

Unmarkr shows you which of these a file carries and removes each in the right way. Start with the free scan, or read how SynthID works.