How to Tell If an Image Was Made With Nano Banana (Gemini)

Every fake image in an ImageGuesser puzzle comes from Google's Nano Banana family: the original Nano Banana, Nano Banana Pro and Nano Banana 2. I chose them because they are among the hardest image models to catch by eye, and because they are what most people now meet in the wild. Nano Banana is the image model built into the Gemini app, which made it one of the most widely used image generators in the world almost overnight. If you are looking at a suspicious image today, there is a good chance it came from one of these models. Here is how to check, starting with the methods that give a real answer.

Start with SynthID, because it is the only check that can actually confirm a Gemini image. Google DeepMind embeds an invisible SynthID watermark into every image its models generate. It is woven into the pixels rather than stored as metadata, so it survives screenshots, resizing and moderate compression. You can check it yourself: upload the image to the Gemini app and ask whether it was created with Google AI. Gemini looks for the watermark and tells you whether it found one. A positive result is about as strong as evidence gets in this field. A negative result is much weaker: it only means the image was probably not made or edited with Google's tools. It says nothing about Midjourney, Flux, OpenAI's models or any of the open-source generators, and a heavily cropped, repainted or re-generated image can lose the signal.

Second, look for the visible watermark and the metadata. Images downloaded from the consumer Gemini app usually carry a small sparkle icon in a corner. It is a useful hint when present, but it is trivially cropped out, so its absence means nothing. Google has also started attaching C2PA Content Credentials to images from its newer models, which you can inspect with a Content Credentials viewer. The catch is the same one I described in the verification article: social platforms strip metadata on upload, so by the time an image is going viral, the credentials are usually gone. Treat metadata as a bonus when it survives, not as something to expect.

If the provenance checks come back empty, you are down to your eyes, and this is where Nano Banana is harder than the models people trained on in 2023. The classic tells have mostly been fixed. Hands are usually right. Faces are symmetrical without looking uncanny. Text is the biggest change: Nano Banana Pro renders headlines, signs and labels cleanly, so "the writing is gibberish" no longer works for the main text in an image. What still breaks is the text nobody asked for. Small print on a bottle in the background, a street sign three blocks away, the spines of books on a shelf: the model spends its effort on what the prompt describes and improvises the rest.

The second thing I look for is a lighting setup that is too good. Nano Banana images often look professionally lit even when the scene would not allow it: a soft, flattering glow on a face in a dim bar, a sunset that lights the subject from the front, shadows that are present but gentle in a way harsh midday sun never is. Real photographs taken by ordinary people are full of bad lighting. An everyday scene that looks like it had a lighting crew is worth a second look.

Third, check the edges of the frame and anything that repeats. The model is strongest at the center of the image, where the subject is, and weakest at the borders and in patterns that need counting. Fence posts that change spacing, tiles that stop lining up, a crowd where the faces near the edge melt into each other, a window grid with one pane that is subtly the wrong shape. These are the same accounting failures I described in the diffusion article, just pushed further out of the spotlight.

Finally, ask what kind of photo this is supposed to be. Nano Banana images default to a clean, well-composed look: sharp focus where it matters, pleasing colors, a subject placed thoughtfully in the frame. Real snapshots are messier. Someone half out of frame, a thumb in the corner, motion blur, a cluttered background nobody bothered to tidy. A fake can imitate this if prompted, but by default the model produces the photo a good photographer would have taken, and life rarely cooperates that well.

None of these visual tells is proof on its own, and it would be wrong to call an image fake because the lighting looks nice. The order I use is: SynthID check in Gemini, then visible watermark and Content Credentials, then a reverse image search to find where the image first appeared, and only then the pixel-level details. If you want to train the last step, the daily puzzle is built for exactly that. Every fake in it was made with a Nano Banana model, and after you guess, you can see which one.