Nano Banana Image Detector: How to Tell If Gemini Made That Picture

Nano Banana images carry SynthID in the pixels and C2PA in the file. How to check a Gemini image and what a negative result really means.

Google marks Nano Banana output two ways: SynthID, an imperceptible watermark baked into the pixels, and C2PA metadata attached to the file. You can check either in the Gemini app itself. The catch is that the visible sparkle mark became optional in August 2026, so its absence now proves nothing at all.

Of every major generator, Google has done the most here. It still is not enough on its own.

Key Takeaways

• Nano Banana Pro is Gemini 3 Pro Image. The nickname stuck, and both names refer to the same Google model family.
• SynthID sits in the pixels, not the file metadata, so it survives a screenshot or a metadata strip that would destroy C2PA.
• C2PA metadata has been embedded in Nano Banana Pro output across the Gemini app, Vertex AI and Google Ads since the week of 20 November 2025.
• Over 20 billion pieces of content have been watermarked with SynthID since Google introduced it in 2023.
• The visible sparkle watermark is now optional. Google made it toggleable on 14 August 2026, while SynthID and C2PA stay embedded either way.

What Google actually attaches

Three separate layers, and people collapse them into one constantly.

The visible sparkle mark is a small overlay in the corner of the image. It was the thing most people meant by "the Gemini watermark", and as of August 2026 users can switch it off. That change matters more than it sounds, and we will come back to it.

SynthID is an imperceptible signal embedded directly into the image data. Google describes it as watermarking technology that embeds signals into AI-generated content without changing how the image looks to you. Because it lives in the pixels, it survives things that destroy file metadata: a screenshot, a re-save, a platform recompression. You can read the SynthID technical overview for how the approach works.

C2PA Content Credentials are cryptographically signed metadata attached to the file itself, recording what generated the image and when. This is the industry standard defined by the C2PA specification, and Google began embedding it in Nano Banana Pro output in late November 2025.

Two markers with opposite failure modes, which is exactly the right design. Metadata is fragile but readable by any compliant tool. Pixel watermarks are durable but only readable by whoever holds the detector.

How to check an image in the Gemini app

This is the part most people do not know exists, and it is genuinely useful.

Google shipped verification inside Gemini itself. Upload an image and ask a plain question: "Was this created with Google AI?" or "Is this AI-generated?" Gemini checks for the SynthID watermark and tells you what it finds, with context about the origin. Google's own announcement of the verification feature walks through it.

That is a real first-party capability and it puts Google ahead of every rival on transparency. xAI, by contrast, publishes no detector at all for Grok, which we covered in our Grok Aurora image detector guide.

But read the question carefully, because the scope is narrow.

The limit nobody says out loud

Gemini's check answers one question: was this made with Google AI?

It does not tell you whether an image came from Midjourney, DALL-E, Stable Diffusion, Flux, Grok or any of the dozens of smaller models. A "no SynthID detected" result is not a verdict of authenticity. It means Google did not find its own watermark, which is exactly what you would expect from a Midjourney render, a Photoshop composite, or an ordinary photograph.

Google has said it intends to support C2PA credentials from outside its ecosystem and to extend verification to video and audio. Until that ships broadly, the tool is a Google-only check.

So the practical reading of a negative result is "not Google, origin unknown". That is a much weaker statement than most people take away from it.

Why the optional watermark changed the maths

Before August 2026, a visible sparkle was decent evidence of Gemini origin, and its absence was mild evidence against.

Now the sparkle is a user preference. Anyone can turn it off. The invisible layers remain, which is the responsible part of Google's decision, but the signal most people can actually see with their own eyes has been removed from the equation.

The consequence: you can no longer eyeball a Gemini image and conclude anything from a clean corner. You have to run an actual check.

That shift matters because it mirrors what is happening everywhere. Visible marks are the easiest to remove and the first to go. Provenance metadata survives until a platform recompresses the file, which happens constantly. We went through that failure chain in detail in whether removing C2PA makes an image undetectable.

The three-layer check that actually works

Here is the sequence worth running on a suspect image, in this order.

Start with pixel analysis, not metadata. This is the layer that survives screenshots, crops, reposts and recompression. A detector reading the image data does not care whether the file arrived with credentials attached.

Run it through our free AI image detector, which gives you a confidence assessment plus a region heatmap marking which parts of the image drove the result. It reads C2PA provenance in the same pass when the file carries it, and analysis runs privately. Worth being precise about the cost: one image scan uses 1,000 credits, the entire starting grant on a free account, so a new account covers a single image. Paid plans start at $7.99 a month. What you are not paying for is the heatmap and the provenance reading, which most rivals reserve for subscribers.

The heatmap matters more than any single number on Gemini images specifically, because Nano Banana is heavily used for editing real photographs rather than generating from scratch. A real photo with a generated background produces a heatmap concentrated on the background. A fully synthetic scene lights up broadly. A percentage cannot tell those apart, and they mean completely different things. We covered that distinction in detecting AI-modified images.

Then check SynthID via Gemini if you suspect Google origin specifically. A positive is strong evidence. A negative narrows nothing.

Then read provenance last. If the original file reached you intact, C2PA credentials are useful corroboration. If the image came through a social platform, assume the manifest is gone and do not read anything into its absence.

Visual tells in Nano Banana output

Google's model is strong on lighting and skin, which is why it caught on. The weak points cluster predictably.

Text rendering remains imperfect on signage, labels and packaging, though better than most rivals. Look at small print under a crop.

Hands in motion and partially occluded fingers still break down, particularly where a hand overlaps an object.

Edited regions are the real tell with Nano Banana, because editing is what it is used for. Look for lighting that does not agree across the frame, a background whose depth of field does not match the subject, and edges where a replaced element meets the original photo.

Our general checklist in how to tell if an image is AI-generated applies here too, and the provenance side is covered in C2PA Content Credentials and how to verify them.

FAQ

Is Nano Banana the same as Gemini?

Nano Banana Pro is the nickname for Gemini 3 Pro Image, Google's image generation and editing model. Same thing, different name.

Does Nano Banana watermark every image?

Google embeds SynthID in the pixels and C2PA metadata in the file. The visible sparkle mark has been optional since 14 August 2026.

How do I check if an image came from Gemini?

Upload it to the Gemini app and ask whether it was created with Google AI. Gemini checks for the SynthID watermark.

If Gemini says no SynthID, is the image real?

No. That result only means it was not made with Google AI. It says nothing about Midjourney, DALL-E, Stable Diffusion or any other generator.

Does a screenshot remove the watermark?

A screenshot destroys C2PA metadata but SynthID lives in the pixels, so it can survive. Heavy editing and aggressive recompression degrade it.

Before you trust the image

Google has built the most serious provenance stack of any major generator, and the Gemini verification tool is worth using. Just remember what it answers, which is a Google-only question in a world with dozens of generators.

Check the pixels first at GPTOne and read the heatmap, then let provenance corroborate.

Meta description: Nano Banana images carry SynthID in the pixels and C2PA in the file. How to check a Gemini image and what a negative result really means.