What Is Nano Banana? Google’s image editing model explained

It was not a product name. An unlabelled model appeared in public head-to-head evaluation arenas, where people compare anonymous outputs without knowing which system produced them, under the placeholder "nano banana". It won a great deal, people noticed, and the guessing about who was behind it ran for weeks before Google confirmed it.

Google then did something unusual and kept the nickname, which is why a serious piece of infrastructure is publicly known by a joke handle. The formal identity is a Gemini image model; the searched-for name is the banana.

That history is worth knowing because it explains why the term behaves so oddly in search. Interest arrived before any marketing did, so the questions people ask are the ones you ask about a rumour — what is it, who made it, is it real — rather than the ones you ask about a product.

Most image models are built around generating a picture from a description. Nano Banana is built around changing a picture you already have, and the hard part of that is everything you did not ask it to change.

Ask a general-purpose model to put the person in this photograph in a red coat and you will usually get a person in a red coat who is not quite the same person — a slightly different jaw, different eye spacing, a face that reads as a sibling. Identity survives one edit and degrades over several. The thing Nano Banana is built to hold onto is that identity, across a run of successive instructions.

The interaction model follows from that. You are not writing one elaborate prompt and accepting what returns; you are having a conversation in which each turn adjusts the last result. Move the light. Now make it evening. Now put him outside. That loop is the product.

Anything where the same subject has to appear more than once. Product photography in several settings, a character across a sequence of panels, a set of variations on one portrait, a mock-up that has to show one room in four different colourways.

It is also unusually forgiving for people who do not want to learn prompt syntax. Because the instruction is an edit rather than a specification, plain sentences work, and the failure mode of a vague request is a small change rather than an unrecognisable picture.

Where it is a poorer fit is the blank-page case. When there is no source image and no subject to preserve, the consistency machinery has nothing to do, and models built for text-to-image composition tend to give you more control over framing, optics and style from a standing start.

Against Midjourney, the divide is aesthetic authorship. Midjourney applies a strong house style and is generally used to make something striking; Nano Banana is trying to leave the picture alone except where you asked. One is a stylist, the other is a retoucher.

Against the open models — Stable Diffusion and FLUX — the divide is where the control lives. Open weights let people bolt on ControlNet, LoRAs and inpainting pipelines to get precise control at the cost of assembling the pipeline. Nano Banana puts a narrower slice of that control behind a sentence, which is faster to reach and harder to push past.

Against gpt-image-1, the two are closer than either is to anything else: both are closed, instruction-following, conversationally driven models attached to a large assistant. The practical difference most people report is which one holds a face better over several turns, and that is worth testing on your own material rather than taking from a leaderboard.

Nothing commercially — we are not affiliated with Google and do not resell its models. This page exists because people ask what the thing is and the honest answer is short, and a directory that only described the models we happen to like would not be a directory.

It is worth reading alongside the editing pages here, because the vocabulary transfers. What Nano Banana does conversationally, the open ecosystem does with named techniques: inpainting for a change inside a mask, outpainting for a change beyond the frame, denoise strength for how far a result is allowed to drift from the source. Knowing those words makes it much clearer what any editor is and is not doing.