Nano banana photo editing: what it is and what it is for

What Nano Banana actually is

Half the confusion around this name comes from the sheer number of sites selling "Nano Banana" as if it were a product you sign up to. It is not. Nano Banana is the family name Google uses for the image models in the Gemini 3 stack, the part of the assistant that turns a description plus your uploaded pictures into a new image.

You reach it three ways, all of them Google's: the Gemini app and website, Google AI Studio for a developer-facing version, and increasingly inside other Google surfaces such as Search and Google Photos. Anything else charging you for "Nano Banana access" is reselling API calls at a markup.

The name started as an unlabelled model that appeared in public testing arenas in 2025 and was good enough that people noticed before Google confirmed what it was. Google kept the nickname. That history is why the naming is doubled: every model has a banana name for people and a Gemini identifier for developers, and the two are used interchangeably in coverage.

The important consequence for anyone editing photos: you are using a general-purpose generative image model, not a photo editor with an AI feature bolted on. Everything below follows from that.

The four models, and which one you are using

All four accept an image plus text and return an edited image. They differ in speed, fidelity and how many references they will take at once.

In the Gemini app you are almost certainly on Nano Banana 2 unless you pay for a Google AI plan and explicitly regenerate with Pro. If a comparison you have read says "Nano Banana cannot do text", check the date: that was true of the 2.5 Flash original and is no longer true of the current model.

What it is genuinely good at

Words inside pictures. The standout. Google built the current model specifically to render accurate, legible text for mockups, greeting cards and marketing layouts, and it will translate text that already exists inside an image. No other mainstream model is as dependable here.

Keeping a subject consistent. Give it several photos of the same person, pet or product and ask for a new scene, and the subject survives. Google's stated figures are character resemblance across up to five characters and object fidelity for up to fourteen objects, which matches what the thing does in practice for one or two rounds.

Native aspect ratios. Ten supported shapes including 4:5 and 9:16 means you ask for the format you need rather than cropping a square afterwards and losing the composition.

Local changes that leave the composition alone. Changing a wall colour, swapping a jacket, altering the weather outside a window. These are the edits where a generative model does something a slider cannot, and where the result usually survives inspection.

Speed. The Flash family is fast enough to iterate properly, which is worth more than a few points of quality when you are converging on a result by trial and error.

What it is bad at

Leaving your photograph alone. This is not a defect so much as a category error people keep making. There is no version of this where your original pixels come back untouched with one object removed. The file you get is drawn from scratch, informed by yours. Photographic grain, sensor noise and skin texture are re-invented every time.

Faces over multiple rounds. One edit of a portrait usually holds. By the fourth or fifth, the face has quietly become a slightly prettier, slightly more symmetrical, slightly less specific version of the person. The drift is invisible step to step and obvious against the original.

Real, identifiable people. Safety filtering around identity modification tightened noticeably through 2026, and refusals are inconsistent enough that the same prompt can pass and fail on different days.

Anything that must be true. Restoring a document, a number plate, a price tag, a face in a low-quality frame. The model will produce something confident and legible that is simply not what was there. Google itself points at Nano Banana Pro for tasks needing factual accuracy, which is an honest admission that the fast model optimises for plausibility.

Unmarked output. SynthID is embedded in everything. If you need a file with no provenance marking, this is the wrong tool.

Editing a real photo against generating a new one

It is worth separating the two things people mean by "photo editing" here, because the model is excellent at one and structurally unsuited to the other.

Generating from a photo means using your image as a starting point for something new: your dog on a beach it has never visited, your product on a marble worktop, your portrait in a different jacket. The photograph is a reference. Nobody expects the result to be a record of anything. This is where Nano Banana earns its reputation and where it is genuinely the best free option available.

Editing a photograph means the result is still supposed to be that photograph, with one thing different. A stranger removed from the background of your holiday shot. A bin gone from the driveway. Here the generative approach has a real cost: you get back a lower-resolution, re-drawn version of a moment you actually recorded, and the parts you did not ask to change have changed anyway.

For that second job, a conventional editor with a patch or generative-fill tool operating on your own file is usually better: Google Photos' own Magic Eraser, Apple's Clean Up, or Snapseed's healing brush. They are less clever and they respect the original, which for a photo of real people in a real place is the point.

Five jobs worth using it for, and three that are not

Worth it:

1. Mockups with text. A poster, a menu, a card, a label. The one job where this model beats everything else outright.
2. Product shots in invented settings. One clean cut-out plus a described scene, repeated across a catalogue with the product staying recognisable.
3. Reformatting for a platform. Turning one image into 4:5, 9:16 and 16:9 versions with sensible content in the new space rather than a brutal crop.
4. Trying an idea before committing. What would this room look like in green, what would this shoot look like at sunset. Cheap answers to expensive questions.
5. Translating text in an image. A menu, a sign, a packaging panel, rendered back in place in another language.

Not worth it:

1. Restoring a damaged family photograph you care about. It will invent faces of people you knew. Use a restoration tool that works at low strength, and keep the scan.
2. Retouching a portrait. The skin comes back smooth and generic. A retouch tool that separates tone from texture keeps the person.
3. Anything evidential. Insurance, identification, a dispute, a document. Generative enhancement has been excluded from court proceedings for exactly this reason, and a plausible invention is worse than a blurry truth.