Gemini and Nano Banana images: the visible mark, SynthID and what each one means

By The NeutronField editorsPublished 2026-11-02Updated 2026-11-02

Checked on 2026-11-02 with c2pa-rs (c2pa-python) and a JPEG segment reader

Short answer

Every image Google's Gemini models generate carries an invisible SynthID watermark in its pixels, Google says. Our Gemini API file also carried a Google-signed content credential that records that watermark, and nothing else: no EXIF or XMP fields and no prompt. The visible sparkle is a separate, app-only mark that depends on your plan.

Evidence order

Strongest first

Signed content credentialGoogle signed it, and it records how the file was made.Present
Invisible watermarkSynthID, recorded in the credential; only Google's detector reads it.Present
Metadata fieldsNo EXIF, XMP or IPTC block outside the credential.Absent

Status on the test file in frame 12, as the Gemini API returned it.

Our file test

One file from the Gemini API, read field by field before any resizing.

Frame 12, the test file: a potter in a denim apron shapes a bowl on the wheel beside a window — ▸ 12NF 01 · TEST FILE

What frame 12 carries

Signal	Yes	No
Content credential	marked	not marked
Invisible watermark	marked	not marked
Metadata fields	not marked	marked
Prompt saved in file	not marked	marked

Checked on 2026-11-02 with c2pa-rs (c2pa-python) and a JPEG segment reader. The watermark row rests on Google's signed record in the file.

Fields read from frame 12, a 1264 by 848 JPEG returned by gemini-3.1-flash-image on 2026-11-02.
Field	Present	Value
C2PA manifest	Yes	Valid; signed by Google Media Processing Services (Google LLC)
Claim generator	Yes	Google C2PA Core Generator Library
Actions recorded	Yes	Created by Google Generative AI; Applied imperceptible SynthID watermark
Digital source type	Yes	trainedAlgorithmicMedia (IPTC)
EXIF, XMP, IPTC blocks	No	None outside the credential
Prompt text	No	Not stored anywhere in the file

Raw output excerpt · c2pa-rs

claim_generator_info: Google C2PA Core Generator Library
signature: Es256, Google LLC, Google Media Processing Services, 2026-11-02T04:28:12Z
c2pa.created: Created by Google Generative AI.
c2pa.edited: Applied imperceptible SynthID watermark.
digitalSourceType: trainedAlgorithmicMedia
validation_state: Valid

The visible mark

The mark you can see is a small Gemini sparkle in the corner of the picture. It is added by the Gemini app, not by the model, and it is the least reliable of the four records: anyone can crop it out, and an image without it may still have come from Google.

When Google launched Nano Banana Pro in November 2025 it said the sparkle would stay on images made by free and Google AI Pro users, and come off for Google AI Ultra subscribers and in the Google AI Studio developer tool (Google, Introducing Nano Banana Pro). App settings have changed often since, so check what your own download shows. Our frame 12 came from the Gemini API and has no visible mark at all.

That split matters if you are deciding what to say about an image. A missing sparkle tells you only which product made the file and on which plan. It does not tell you a person drew it.

The invisible watermark

SynthID is Google DeepMind's watermark. It is written into the pixels as the image is generated, it does not change how the image looks, and Google says it is designed to stand up to cropping, filters and lossy compression (Google DeepMind, SynthID). For developers the rule is stated plainly: "All generated images include a SynthID watermark" (Gemini API docs, image generation).

You cannot see SynthID or read it with a metadata tool. What we could read was Google's own signed statement about it: the content credential in frame 12 lists two actions, "Created by Google Generative AI." and "Applied imperceptible SynthID watermark." That is why our record marks the watermark as present: the evidence is Google's own signed record, not a detector run.

The two records behave differently once the file moves. When we resized frame 12 and saved it as WebP for this site, the content credential disappeared completely; the WebP copies on this page carry no manifest. The watermark is meant to survive that kind of change, though Google's help page notes that after many alterations it may no longer be detected (Gemini Apps Help, verify AI-generated content). Google has also said it now adds C2PA credentials to Nano Banana Pro images in the Gemini app, Vertex AI and Google Ads (Google, AI image verification in the Gemini app); our API file from the faster Nano Banana 2 model carried one too. We repeated the test with Nano Banana Pro (gemini-3-pro-image) through the API on the same day: same signer, the same two actions, and a valid credential. For the same test on OpenAI's images, see what ChatGPT puts in your images.

Diagram of the layers of an image file: pixels, invisible watermark, metadata fields, content credential and saved prompt — Frame 20. The layers of one file, from the pixels out. A Gemini file fills layers 2 and 4.

Checking a Google image

If you made the image yourself, keep the file Gemini gave you. The original is the only copy that still has the credential, and it is what you would point to if anyone asks how the picture was made.

Start from the original download, not a screenshot or a copy saved from a chat or social feed; those are re-encoded and lose the credential.
Open a content credential viewer, such as the free one from the Content Authenticity Initiative, and load the file. A Gemini file shows Google LLC as the signer and lists the actions above.
To check for SynthID, upload the image to the Gemini app and ask whether it was created or edited by Google AI. Google allows about ten image checks in any 24 hours, and the app also reads Content Credentials on the web and on Android (Gemini Apps Help).
Note the date and the result next to the file, the same way we date every record on this site. Tools and app rules change, and a dated note is what makes the check worth something later.

What this can and cannot prove

A valid Google credential shows that Google signed a statement about the file and that the pixels have not changed since. Detecting SynthID shows that Google's models made or edited all or part of the image. Neither tells you who wrote the prompt, whether a person edited the picture afterwards in another tool, or whether a version without these records exists elsewhere.

The absence of every record proves the least. A screenshot of a Gemini image carries no credential, and a heavily altered copy may lose the watermark too. For other generators, start at the Made with AI hub, and for the four layers in general, what your AI image carries.

A missing label does not mean an image is human-made.

Sources

Gemini API docs: Image generation (accessed 2026-11-02)ai.google.dev
Google: Introducing Nano Banana Pro, 2025-11-20 (accessed 2026-11-02)blog.google
Google: How we're bringing AI image verification to the Gemini app, 2025-11-20 (accessed 2026-11-02)blog.google
Gemini Apps Help: Verify AI-generated images, videos, and audio (accessed 2026-11-02)support.google.com
Google DeepMind: SynthID (accessed 2026-11-02)deepmind.google

Questions

Do all Nano Banana images have a watermark?

Yes, an invisible one. Google's developer documentation says all images generated through the Gemini API include a SynthID watermark, and Google says the same of media from its consumer tools. The visible sparkle is separate and appears only on some Gemini app images.

Is the Gemini sparkle the only watermark?

No. The sparkle is a visible mark the Gemini app adds on some plans. Every image also carries SynthID, an invisible watermark in the pixels, and files we tested also carried a signed content credential from Google.

Do Gemini API images have a visible watermark?

Our file from the Gemini API on 2026-11-02 had none, and Google said at the Nano Banana Pro launch that the visible mark comes off in the Google AI Studio developer tool. The invisible SynthID watermark is still there.

Does resizing a Gemini image remove the watermark?

Resizing removed the content credential from our copy completely. SynthID is designed to survive resizing, recolouring and compression, though Google says that after many alterations it may no longer be detected.

How can I tell if an image came from Gemini?

Upload it to the Gemini app and ask whether it was created or edited by Google AI; the app checks for SynthID and reads content credentials. If you have the original file, a content credential viewer shows Google's signed record.