What Is SynthID?

A signal embedded in content

Google describes SynthID as a technology for watermarking and identifying generated content. For images, the invisible signal is embedded in the content at generation time. Looking at a preview, finding EXIF or reading a filename cannot confirm its presence.

An export is not a verdict

The watermark is designed to withstand common edits, including lossy compression. A smaller file or a copy without metadata can still carry it. Our experimental export changes sampling and encoding; it is not a validated watermark removal engine. A visually cleaner copy therefore remains unverified.

Use metadata as a control

A supported container report can show camera fields, editing information or PNG prompts. Metadata cleanup removes recognized blocks while retaining image data. Keep that output as a control when testing other transformations. An AI probability score from a visual classifier is not interchangeable with a watermark detector.

Test the original and the copy

Obtain a positive baseline with an official verification surface before evaluating a transformation. Test the exported file under the same conditions, record both outcomes and keep the files separate. If the source was never positive, the pair cannot demonstrate removal. One result also cannot establish future platform behavior.

Google DeepMind: SynthID ↗

Preserve evidence

Save the report, actual settings and detector record together. A SHA-256 hash identifies file bytes; it does not identify a watermark. Review text detail, gradients, color and transparency before delivering an experimental copy. Keep an original with provenance when the receiving workflow requires it.