WatermarkWipe
What Is SynthID? Google's Invisible AI Watermark Explained
Every image Gemini creates carries two watermarks. One is the small star in the corner that you can see. The other is SynthID, an invisible signal woven through the entire image. Here is what it does and why it matters.
Last updated: 12 June 2026 · 4 min read
The short version
- SynthID is invisible. It does not change what the image looks like to a human.
- It is everywhere in the image, not just the corner. The signal is embedded across every region, so cropping or resizing usually does not remove it.
- Only Google can verify it. A matching detector model is required to read the signal back out.
- WatermarkWipe does not remove it. It only takes off the visible star.
What SynthID is, in plain language
SynthID is a watermarking system developed by Google DeepMind. When Gemini generates an image, the model is steered to produce pixels that carry a faint, structured pattern. The pattern is too subtle for the human eye to see, but a paired detector model can scan a suspect image and tell, with statistical confidence, whether it came from Gemini.
You can think of SynthID like a barely-audible tone hidden under a song. You will not hear it during normal listening, but a microphone tuned to that exact frequency can pick it up. With images, the "tone" is spread across many pixels and many frequencies so that ordinary edits do not destroy it.
How is it different from the visible star?
The two watermarks have very different jobs:
| Property | Visible star | SynthID |
|---|---|---|
| Visible to humans | Yes | No |
| Where it sits | Bottom-right corner only | Spread across the whole image |
| Purpose | At-a-glance "this is AI" | Programmatic detection |
| Survives cropping? | Lost if the corner is cut | Usually survives |
| Survives resizing / compression? | Yes | Designed to, within limits |
| Removed by WatermarkWipe? | Yes | No |
Can SynthID be removed?
Not reliably, and that is by design. Because the signal is spread across the image and across many visual frequencies, simple operations like resizing, JPEG compression or filtering only weaken it rather than erase it. Heavily destructive edits such as strong noise, big crops or repeated re-encoding can degrade it to the point where detection becomes uncertain, but that also visibly damages the image.
WatermarkWipe does not attempt to remove SynthID and does not claim to. If a website, platform or checker uses SynthID to flag AI imagery, a WatermarkWipe-cleaned image will still be flagged. That is the right outcome: hiding the visible corner star for personal projects is a cosmetic fix; pretending an AI image is a real photograph is something else entirely.
Why this matters for AI transparency
Visible watermarks are easy to remove (cropping is sometimes enough) and easy to fake. Invisible watermarks like SynthID are the harder, more durable signal. They give platforms a way to flag AI content even when the visible label is gone. Whether you agree with Google's approach or not, the practical reality is that two layers exist, and removing only the visible one is not the same as making an image undetectable.
So what does WatermarkWipe actually do?
It removes the visible star, cleanly. That is the whole job. Because the star is a known shape blended in a known way at a known position, the blend can be reversed mathematically, with no AI inpainting, no pixel hallucination and no quality loss. SynthID remains untouched.
If you want the step-by-step, see the how-to guide. If you just want to clean an image now, the tool is one click away.