How to tell if an image is Nano Banana (Google)
"Nano Banana" is Google's nickname for its Gemini image generation and editing models. As of August 2026, Google's own API documentation applies the name to four models: gemini-2.5-flash-image (the original Nano Banana, released August 2025, now listed as legacy), gemini-3-pro-image (Nano Banana Pro), gemini-3.1-flash-image (Nano Banana 2) and gemini-3.1-flash-lite-image (Nano Banana 2 Lite). This page covers what these models leave in the file, what their output tends to look like, and where each of those signals stops working.
Visual tells
A visible Gemini sparkle in a corner. Google states that images created by users on the free and Google AI Pro tiers carry a visible Gemini sparkle watermark, that it is removed for Google AI Ultra subscribers, and that Google AI Studio produces output without it. Finding the sparkle is a strong signal. Not finding it tells you nothing, since three common routes produce images without it.
Short text renders correctly while long text falls apart. Unlike older generators, these models usually get headlines, labels and product names right. Reviewers consistently report that paragraphs and body copy still warp, wander off the baseline, or resolve into non-words. A correct headline sitting above a malformed paragraph in the same image is a characteristic combination.
Smoothed, low-texture skin in portraits. Faces tend toward even skin with few visible pores, uniform lighting and near-symmetrical features, the "plastic skin" look. Users reported this getting more pronounced in Gemini app output during 2026, alongside complaints that faces read as older than requested.
Detail that does not survive magnification at 2K and 4K. Where the high-resolution output comes from upscaling rather than native generation, out-of-focus highlights show stair-step and zipper artefacts, and fine texture reads as invented detail rather than real structure. Reviewers report imperfections becoming obvious at 200 to 300 percent zoom on otherwise convincing images.
Identity drift between an edited image and its source. These models are used heavily to edit real photographs, and Google's own support forum carries user reports that the subject does not stay fully consistent with the original. If you can obtain the source photo, compare face geometry, hairline, teeth and jewellery: small unexplained shifts point to a generative edit rather than a conventional one.
Dimensions that land on Google's published grid. Google documents output at 512px, 1K, 2K and 4K in the aspect ratios 1:1, 3:2, 2:3, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9 and 21:9. Exact squares such as 1024x1024, 2048x2048 and 4096x4096 are consistent with this family. Many other tools use the same round numbers, so this narrows the field rather than identifying a generator.
Metadata & provenance
SynthID, always, in the pixels. Google's Gemini API documentation states plainly: "All generated images include a SynthID watermark." SynthID is imperceptible and is embedded in the pixel data rather than in metadata, so it survives some re-encoding. Google says the watermark may survive re-scaling, re-colouring and compression, but also says that "after many alterations the watermark won't be detected." No independent party can read SynthID. The watermark pattern and the detector model are not published, and there is no open specification comparable to C2PA. Google verifies SynthID through three channels it controls: asking the Gemini app "Is this AI generated?", the waitlisted SynthID Detector portal, and a Content Detection API previewed to selected partners. wedetect cannot check SynthID, and neither can any other independent detector. C2PA Content Credentials, in the file. Google Cloud's documentation states that images created or modified by its supported Gemini image models, which include Nano Banana and Nano Banana Pro, automatically receive Content Credentials digitally signed by Google LLC. The C2PA manifest is embedded metadata recording the tool used and that the content is AI-generated. Any open C2PA viewer can read it and verify the signature. Google's Gemini app reads Content Credentials version 2.2 and later. EXIF and XMP. We have not verified any EXIF or XMP generator tag written by these models, so do not treat an empty EXIF block as meaningful either way. Provenance here is carried by SynthID in the pixels and C2PA in the metadata, not by a conventional "Software" field. Absence of provenance is not evidence of anything. Google's own help page states that if a SynthID watermark is not detected, it means the image was not created or edited by Google AI, but it could have been created by other AI systems. C2PA manifests are ordinary metadata: a screenshot, a re-save, most social platform upload pipelines, or any metadata stripper will remove them, and tools that specifically target the sparkle, SynthID and C2PA together are publicly distributed. Read a clean file as "no information", never as "not AI".
How well WeDetect handles Nano Banana (Google)
wedetect has no published accuracy figure for Nano Banana specifically, and this page does not claim one. Our measured figures are roughly 89% overall across modern generators and roughly 85% recall on Midjourney, DALL-E, SDXL and Flux. On Stable Diffusion 3.5, measured recall is roughly 62%. That spread is the honest picture: performance is not uniform across generators, and we have no measurement placing Nano Banana at the top of the range rather than the bottom. wedetect works on pixel-level statistics alone. It cannot read SynthID, because Google does not publish the detector, and it cannot restore metadata that has already been stripped. Every result is a probability, not a verdict, and it should be weighed alongside a Content Credentials check and your own reading of the image.
Check an image now
Drop the image in below — free, no sign-up, deleted after analysis.
Drag & drop, paste, or click to upload
PNG · JPG · WEBP · HEIC — your image is analyzed privately and deleted
This is an estimate, not proof — the detector can be wrong, producing both false positives and false negatives. Don't rely on it as your only source; verify anything important. wedetect.ai is not liable for decisions made from these results. See our Terms.
Frequently asked questions
How can I tell if an image was made with Nano Banana?
There is no single check that settles it. Three things are worth doing in order. First, upload the file to the Gemini app and ask whether it was made with Google AI: that queries SynthID and C2PA, which only Google can read. Second, open the file in any C2PA viewer to see whether a Google-signed Content Credentials manifest survived. Third, look at the image itself for the tells above, particularly long text, skin texture and detail under magnification. If the file has been screenshotted, re-saved or run through a social platform, the metadata route usually returns nothing, and you are left with pixel analysis, which is probabilistic.
Does Nano Banana add a watermark to images?
Yes, in three separate senses. Every generated image includes SynthID, an invisible watermark embedded in the pixels. Images from Google's supported models also carry C2PA Content Credentials signed by Google LLC. Separately, Google applies a visible Gemini sparkle to images made by users on the free and Google AI Pro tiers, removes it for Google AI Ultra subscribers, and does not apply it to Google AI Studio output. So an image with no visible sparkle can still be Nano Banana output.
Can the Nano Banana watermark be removed?
The visible sparkle and the C2PA metadata can both be removed, and tools that do it are freely available. SynthID is harder to remove because it lives in the pixels rather than the metadata, but Google itself states that after many alterations the watermark will not be detected. The practical consequence for verification is that a file with no detectable provenance is not evidence that an image is real. Treat missing provenance as an absence of information.
Can wedetect tell me if an image came from Nano Banana specifically?
No. wedetect returns a probability that an image is AI-generated. It does not attribute images to a named model, and we do not publish a per-generator figure for Nano Banana. Our published numbers are roughly 89% overall on modern generators and roughly 85% recall on Midjourney, DALL-E, SDXL and Flux, while measured recall on Stable Diffusion 3.5 is roughly 62%. For a Google-specific answer, the Gemini app's own verification is the only tool that can read SynthID.
Other generators: Midjourney · DALL·E 3 · GPT-4o · Stable Diffusion · Flux · Grok · Firefly · Gemini · SDXL · GPT Image 2 · Flux 2 · SD 3.5 · Sora · all generators.