Skip to content
perspicaxby avancer.ai
Guide · Made with Gemini?

How to tell if an image was made with Gemini

Google labels every image its models make, in more than one way. Here's how to read those labels, and what to do when a copy has lost them.

By Perspicax research team, avancer.ai7 min read
In short

Can you tell if Gemini made an image?

Usually, if you have a good copy. Google adds SynthID, an invisible watermark, to every image its models make, and the Gemini app can check for it. Copies often lose the visible sparkle and the signed label. Perspicax checks the pixels instead: it caught 97.8–100.0% of Gemini images in testing, including Nano Banana Pro and Imagen 4.

On this page
  1. What Google puts in its images
  2. Step 1: ask Google's own tools
  3. Step 2: look for the sparkle, but don't rely on it
  4. Step 3: look at the picture
  5. Step 4: trace where it came from
  6. Step 5: check the pixels with Perspicax
  7. What survives a screenshot, a crop or a share
  8. Check an image
  9. Questions

What Google puts in its images

The pictures people call "Gemini images" come from Google's image models. Nano Banana makes the images in the Gemini app, and also in AI Mode in Search, AI Studio, Flow and Google Ads. Imagen 4 reaches people through the Gemini API and Vertex AI. So an image made with Google's models may never have been near the Gemini app itself.

Google marks these images in up to three ways:

  • SynthID: an invisible watermark that Google adds to every image its models make. It sits in the pixels rather than in the file's metadata, so it is designed to survive re-saves and screenshots that strip other labels.
  • The sparkle: a small visible Gemini mark in a corner of images from the Gemini app. Since August 2026 people can choose to hide it.
  • Content Credentials (C2PA): a signed record in the file's metadata saying which tool made the image. Nano Banana Pro images carry it in the Gemini app, Vertex AI and Google Ads.

Each signal fails in its own way. The sparkle is the easiest to remove. The signed label goes with the first re-save: we made a photo with Google's model ourselves, and one re-save erased its label. SynthID lasts longer, but you need Google's own tools to read it.

Step 1: ask Google's own tools

Because Google watermarks its own images, Google can tell you whether its watermark is there. Three free checks, two of them Google's own:

  1. Ask the Gemini app

    Upload the image and ask "Was this created with Google AI?". The app looks for SynthID and explains what it found. You need to be signed in, and there is a daily limit on checks.

  2. Open "About this image"

    In Google Lens, or with Circle to Search on Android, open the image's "About this image" panel. It can show whether Google's watermark was found, what the file's metadata says and where else the image appears online.

  3. Read the label in the file

    If you have the original file, drop it into our Content Credentials checker. A valid record signed by Google names the tool that made the image.

Step 2: look for the sparkle, but don't rely on it

When you see the Gemini sparkle in the corner of an image nobody has edited, it is a strong hint. On its own, though, it is a weak test, in both directions:

  • Since August 2026, people can turn the visible mark off before they share an image.
  • A crop removes it in a second, and many reposts are cropped to fit a feed anyway.
  • Images made through Google's API, Vertex AI or apps built on Google's models may never have had it.
  • Anyone can paste a sparkle onto a real photo, so finding one isn't proof either.

Samsung phones add their own sparkle mark to photos changed with Galaxy AI's Generative Edit. That marks an AI edit to a real photo, not a picture Google generated. If that is your question, see how to tell if a photo was edited with AI.

Step 3: look at the picture

Nano Banana Pro and Imagen 4 make some of the most convincing photos of any generator. The tells that worked a year or two ago rarely help now, so look for the things a single prompt struggles to keep consistent:

Still worth checking

  • A story that is too complete: every element of a scene present and perfectly placed, as if it were an advert for itself.
  • Background people and objects: faces in a crowd that repeat, or things far from the subject that merge into each other.
  • Brand names and signs: usually spelt right now, but a shop or product that doesn't exist, or a real one in the wrong typeface, is worth a search.
  • Consistency across a set: "holiday photos" where the same person's face, clothes or room change slightly from one shot to the next.

No longer reliable

  • Hands and fingers: Google's newest models get them right.
  • Garbled text: Nano Banana Pro writes legible, correct text, even in small print.
  • Plastic skin and a glossy finish: asked for a phone snapshot, Google's models add grain, blur and the small flaws of a real camera.
  • Looking for the sparkle alone: it can be hidden, cropped or added.
Grilled kebabs on thin flatbread with fresh mint and a bowl of tomatoes, a real phone photoReal photo
A slice of kunafa lifted from a round tray, cheese stretching, with a glass of tea beside it, made with AIAI-generated
Two food photos. The AI one has steam, a cheese pull and a hand holding the server, and none of it is wrong.Real photo: Behnam9395, CC0, via Wikimedia Commons.

The AI half of this pair was made with Nano Banana Pro, the Google model behind many of Gemini's best images. Look as closely as you like: the steam, the stretching cheese and the hand holding the server all hold up. That is why a close look is a first filter, not an answer.

Step 4: trace where it came from

Nano Banana images spread fast because they are fun to make: figurines, retro portraits, selfies with celebrities, made-up holiday snaps. Many are posted with no mention of AI, and some are reposted as real. Finding the first post often settles the question:

  • Run a reverse image search and sort the results by date.
  • Look at the account that posted it first: does it share other Gemini creations, prompts or a link to a prompt collection?
  • If the picture claims to show a real event, look for the same moment photographed by other people.
  • Search for any text in the image, such as a sign or a slogan, together with the place it claims to show.

Step 5: check the pixels with Perspicax

Perspicax doesn't need Google's watermark or the signed label. Proven computer vision techniques and TrueEye 2.3, our state-of-the-art detector, examine every region of the image, so it works on screenshots, crops and re-saved copies as well as on originals.

97.8–100.0%of Gemini images (Nano Banana, Imagen) caught in testing

Measured the way Perspicax answers: an inconclusive answer counts as a miss. See every number

In testing, it caught 97.8–100.0% of Gemini images: 97.8% for Nano Banana Pro, 99.3% for Nano Banana 2.1 and 100.0% for Imagen 4. Nano Banana Pro was the hardest of the three.

Three possible answers

Likely AI-generatedThe pixels match images made by AI generators.
InconclusiveThe evidence was not strong enough either way: check further.
Likely realThe pixels match photos taken with a camera.

It tells you whether an image was fully generated by AI, not which model made it. For the model-by-model results and the labels each Google tool adds, see our Gemini image detector.

What survives a screenshot, a crop or a share

Most Gemini images reach you second- or third-hand. Here is what each step on the way usually does to Google's signals:

  • A screenshot: the signed label is gone. The sparkle stays if it was in the frame. SynthID is designed to survive.
  • A crop: the sparkle usually goes, and the label goes with the re-save. SynthID is designed to survive light edits.
  • An upload to a social or messaging app: the label is usually stripped and the picture is compressed. The sparkle stays only if nobody hid or cropped it.
  • Heavy editing in another app: anything can change, and no watermark is impossible to weaken.

The one thing that always travels with the picture is the picture itself. That is why Perspicax checks the pixels: it gives you an answer even when every label has been lost on the way.

Check an image now

Drop, paste or link a photo. Perspicax checks every region of it with computer vision and TrueEye 2.3, and answers likely AI-generated, likely real or inconclusive.

Paste an image or a link anywhere

Your first 3 scans are free. By scanning, you accept our Terms. Images and results are kept until you delete them (Privacy).

Questions people ask

Does Gemini add a watermark to its images?

Yes. Every image from Google's models carries SynthID, an invisible watermark, and Nano Banana Pro images in the Gemini app also carry Content Credentials. The visible sparkle is optional and easy to crop.

How do I check if an image was made with Gemini?

Ask the Gemini app "Was this created with Google AI?", or open About this image in Google Lens. Both look for Google's watermark. To check any copy, label or not, use a detector that reads the pixels, such as Perspicax.

Is the Gemini sparkle enough to tell?

No. Since August 2026 people can hide it, a crop removes it, and anyone can paste one onto a real photo. Treat it as a hint, not a verdict.

Is Nano Banana the same as Gemini?

Nano Banana is the family of Google image models behind the Gemini app's pictures. The same models also run in AI Mode in Search, AI Studio, Flow and Google Ads.

If Google's check finds nothing, is the image real?

No. Google's tools look for Google's watermark only, so an image from another company's model passes them. Check the pixels as well.

Which Gemini model is hardest to detect?

In our tests, Nano Banana Pro: Perspicax caught 97.8% of its images, against 99.3% for Nano Banana 2.1 and 100.0% for Imagen 4.

Related guides

Keep checking with the guide for your situation.

All guides

Before you believe it, check it.

Your first 3 scans are free, no account needed. A free account gives you 5 scans every day.

Scan an image