Gemini & Nano Banana image detector

Google’s three labels

Google has taken a layered approach to marking AI images from Gemini, Imagen and the Nano Banana editing model. Each layer survives different things.

SynthID — in the pixels

SynthID is an invisible watermark woven into the image while it is generated. It is built to survive cropping, resizing, colour changes and compression. The trade-off is that it can only be read by Google: there is no public specification and no open detector. When someone online claims their tool “detects SynthID”, they almost always mean they detect the metadata described below.

IPTC digital source type — in the metadata

Google also writes the industry-standard IPTC field that declares an image as created by a trained algorithm. Photo apps and social networks that read this field may show a label such as “AI info” or “Made with Google AI”. GPTTrace reads it from XMP and reports it as proof.

Content Credentials — signed metadata

Google is a steering member of the C2PA coalition and has been adding signed Content Credentials to images from its products. Where they exist, GPTTrace shows the signer and the recorded actions.

Metadata can be stripped by a screenshot or an upload; SynthID cannot easily be removed but cannot be read by third parties. That gap is where GPTTrace’s neural classifier does its work.

What else to look for

Images generated in the consumer Gemini app have carried a small visible sparkle mark in a corner, which people often crop off — a corner that looks oddly trimmed is worth noticing. Nano Banana edits are often realistic portraits of real people placed in new settings, figurine-style renders or restored and colourised old photos. For an edited real photo, the classifier may only partly react, because most of the pixels came from a camera.

Edits versus generations

Much of what people do with Nano Banana is editing rather than generating: changing an outfit, adding a person to a group photo, removing a crowd, restoring an old print. Google records that distinction in metadata — “composite with trained algorithmic media” for edits — and GPTTrace reports it separately, because an edited real photo and a fully invented scene are different kinds of problem. A restored family photograph is not a fake; a real person placed somewhere they never were may be. When the metadata is gone, crop to the part of the picture that matters and check that region on its own.

For anything important, combine a GPTTrace check with Google’s own SynthID verification, and look for the image’s original source.