How does SynthID work?
SynthID is Google DeepMind's invisible watermark for AI-generated images. It hides a signal in the pixels while the image is being created — imperceptible to you, readable by Google's detector, and built to survive everyday edits.
When one of Google's models generates an image, SynthID adjusts pixel values across the whole frame into a subtle, structured pattern. Each adjustment is far too small for the human eye to notice, and the pattern is woven into the picture's own content rather than stamped on top of it. The result: an image that looks identical to an unwatermarked one, but that a matching detector can identify as AI-generated with high confidence.
Embedding: the watermark goes in during generation
Traditional watermarks are added after the fact — a logo in the corner, or a metadata tag in the file. SynthID works differently: one neural network embeds the watermark as part of the image-generation process itself, and a second network is trained to detect it. Because the two are trained together, the embedder learns to hide the signal exactly where the detector can most reliably find it, while an adversarial constraint keeps the change below the threshold of human vision.
Detection: reading the mark back
Detection runs the suspect image through the paired detector network, which reports how strongly the watermark pattern is present. There is no key you can extract by staring at pixels — the "key" is the trained detector itself, which is why checking an image means asking Google's tooling rather than running a local script. Google has been rolling out a SynthID Detector portal for exactly that.
What SynthID survives — and what it doesn't
Robustness is the whole point of pixel-level watermarking: the signal is spread redundantly across the frame, so no single edit removes it cleanly. Publicly reported behaviour looks like this:
| Edit | Watermark outcome |
|---|---|
| Save as compressed JPEG | Typically survives |
| Resize | Typically survives |
| Color filters, brightness changes | Typically survives |
| Screenshot of the image | Often survives, not guaranteed |
| Moderate cropping | Often survives, confidence drops |
| Heavy repainting or inpainting | Signal weakens with replaced pixels |
| Regenerating the image with another model | Mark does not carry over |
| Stripping file metadata | No effect — the mark is in the pixels |
That last row is the key difference from provenance metadata such as C2PA content credentials: metadata dies the moment a platform re-encodes the upload, while a pixel-domain watermark rides along inside the image itself.
Can a SynthID watermark be removed?
Not with the one-click edits that defeat visible logos and metadata. Because the signal hides in the same frequencies as the image content, removing it reliably means degrading the image — or transforming it so heavily that it is effectively a different picture. Whether a dedicated removal tool should exist is a live debate about provenance, consent, and who controls the label on a piece of media.
That debate is exactly what this site measures.
Vote: should an AI watermark remover exist?Frequently asked questions
Is SynthID visible in the image?
No. The watermark is a pattern of tiny pixel-value adjustments spread across the whole image, tuned to stay below what the human eye can notice. Only a detector holding the matching model can read it.
Does SynthID survive cropping or compression?
Usually, yes. The signal is spread redundantly, so it typically survives JPEG compression, resizing, moderate cropping, and filters. Heavy edits that replace most of the pixels can weaken it below detection.
Which Google products add a SynthID watermark?
Output from Google's generative models, including images from Imagen and Gemini. DeepMind has extended the technique family to AI-generated audio, video, and text.
Can I check whether an image has a SynthID watermark?
Detection requires Google's detector — the watermark can only be read by the model that knows how it was embedded. Google has been rolling out a SynthID Detector portal for uploading and checking content.
Is SynthID the same as C2PA content credentials?
No. C2PA credentials are signed metadata attached alongside the file and are stripped by re-encoding. SynthID lives inside the pixels, so it survives metadata removal. They're complementary provenance layers.