Early Nano Banana Image Editing Examples and a Hands-On Test

In August 2025, a model known as Nano Banana appeared in LM Arena’s image battles. I look at examples shared by others and try a few edits myself, focusing on how well it changes one part of an image while preserving the rest.

One example asks the model to add a third bag of dog food to a shopping cart. The result keeps much of the scene and packaging consistent, although I notice some imperfections in the dog illustration.

A generated selfie combines a young Michael Jackson and Billie Eilish. I find the image convincing, which also illustrates that a realistic composite is not evidence that the depicted event occurred.

Another example overlays a 3D-looking mesh on a photograph of Tom Holland. The mesh appears to follow the clothing and pose, but the picture alone does not prove that a correct underlying 3D model was recovered.

Several examples repair and colorize damaged or faded photographs. The results can be visually appealing, but inferred colors and reconstructed details should not be assumed to record the original scene accurately.

At the time, people encountered Nano Banana through random matchups in LM Arena. I interpret a banana emoji posted by Google’s Logan Kilpatrick as a clue that Google may be behind it; this was an inference at the time of the video.

After a sponsored segment for Chatbase, I show a four-panel sports montage that follows a reference style. Other examples colorize an image, imagine the reverse view of a scene and create an advertisement; unseen parts of a scene are generated guesses.

One composite combines separate images of a man, a woman, a dog and a car. Details such as the chef’s clothing carry through, illustrating multi-image composition, although not every property of the source car is easy to verify.

An example puts Tim Cook with a rumored phone, but I find the handset awkwardly shaped and oversized. Another replaces Batman with Superman while preserving the cat; the change is not perfect, but the unchanged element remains consistent.

A hat is added to a portrait, and another instruction turns a stack of books upright between bookends. The book example exposes a limitation: some names and lettering become unclear or altered after the edit.

A comparison with GPT Image 1 gives two different visual styles for the same prompt. I prefer the more realistic-looking result in this example, but one selected output is not a general benchmark.

A product-placement example compares FLUX.1 Kontext with Nano Banana on inserting a particular beer bottle into a character’s hand. In the displayed outputs, Nano Banana retains the bottle more closely, while the FLUX example produces a glass and a problematic hand.

Other examples combine public figures in a generated beach scene and place a chair beside a lamp. In the latter, I like the light and patterned shadows; these are selected demonstrations rather than proof of consistent performance on every prompt.

At the time of recording, the way to try it was to choose image generation and battle mode in LM Arena, then wait for the service to select Nano Banana randomly. Availability and access methods described here reflect August 2025.

For my own test, I ask to remove the background from a portrait and place me in space. Compared with the Gemini 2.0 Flash image shown beside it, I prefer the Nano Banana result because it preserves my face more closely.

I then add a space helmet and ask for slightly tinted glass, comparing Nano Banana with Qwen Image Edit. For the tint, I actually prefer the Qwen example, showing that the result depends on the particular edit.

Finally I ask for a giant banana chasing me. The result is funny but does not look much like me, despite carrying some features forward. These early examples make me eager to test the model further.