Nano Banana is Google’s image editor that works the way you talk: upload a photo, describe the change in plain language, and get back the same photo with that change made. It lives in the Gemini app, it keeps the same person looking like themselves across edits, and it powered more than 5 billion creations in its first weeks. This guide explains what it is and how to edit with it.
The short version is that Nano Banana, officially the Gemini 2.5 Flash Image model, was built from the ground up to process text and images at the same time. Instead of only generating images from a text prompt, it understands an existing image and works on it. It also remembers what it just created within a conversation, so edits build on each other instead of starting from scratch every time.
What Nano Banana Is
Nano Banana is the Gemini app’s image generation and editing model, built by Google DeepMind. The name started as a codename that went viral before the official reveal, and it stuck: Google itself now uses Nano Banana for the model in the Gemini app, while developers know the same model as Gemini 2.5 Flash Image.
Its core idea is native multimodality. Traditional image tools generate from text alone, or edit with masks, brushes and layers. Nano Banana processes the image and your words together, so it can understand what is already in the picture, apply your instruction to it, and use Gemini’s knowledge of the world to fill in the blanks sensibly.
Google’s app has a long history of understanding images, from identifying what the camera sees, which our guide to how Google Lens and Ask Photos work covers, to editing them. Nano Banana is the editing half of that story: not just recognising a photo, but remaking it on instruction.
How to Edit a Photo in the Gemini App
Open the Gemini app and start a conversation with your photo. Upload the image you want to change, then describe the change in everyday language: change the background, swap an outfit, restyle the photo, remove an object. You do not need masks, layers or editing vocabulary; the instruction is the edit.
Be specific about the one thing you want changed, and let the model handle the rest. “Change the sofa from grey to deep green” works better than “make the living room nicer,” because the model has to guess what “nicer” means, and its guess may rearrange things you wanted untouched.
After the first edit, keep talking. Nano Banana does not treat each new request as a blank slate: it understands what it just created, so you can refine the result with follow-up instructions instead of starting over. That conversation is the workflow, and it is what separates this kind of editing from single-shot generation.
When a result goes wrong, refine it in the same thread rather than starting a new chat. A new chat loses the context the model built up, including the version you liked. A follow-up message keeps that version and corrects only the part that failed, which is almost always cheaper than regenerating from scratch.
Keep the Same Person Across Edits
Nano Banana’s key strength, and the reason it went viral, is consistency. The model can reuse the same characters while altering their outfits, poses, lighting or the entire scene, or render them from different angles, all while preserving their likeness.
That matters most with people you know well. As Gemini’s product manager puts it, subtle flaws make a difference with familiar faces: a depiction that is close but not quite the same feels off. The model’s progress, in his words, is from something that looks like your AI distant cousin to images that look like you.
The most popular demonstration of this is turning a photo into a figurine: one prompt takes a selfie and returns the same person as a collectible figure, face intact. The same principle powers professional uses, from better headshots for profiles to seeing yourself in different outfits or professions while still looking like yourself.
Animals get the same treatment. Google’s product team talks about photos of people and even animals looking consistently like themselves, which is why pet portraits survive restyling: the same dog in a new scene is still recognisably the same dog, not a generic replacement.
Change One Detail Without Touching the Rest
Google calls the second strength pixel-perfect editing: altering a specific element within an image without affecting the rest of the scene. Change the colour of a sofa to brainstorm home decor ideas, and the room around it stays put. Fix a sign to say something different, and the street behind it does not change.
Google’s own example is almost comically small: the prompt “the dog’s mouth is closed” takes a photo of a dog with its head out of a car window, tongue out, and returns the same dog with its mouth closed. Nothing else moves. That precision is the point: small, controlled changes are where most real photo editing lives.
The same precision works on graphics. Upload a mockup of an interface and change the colour of a button, or adjust a logo’s size, without distorting the surrounding layout. For anyone who has watched a whole design shift because of one tweak, that restraint is the feature.
Edit Step by Step in One Conversation
Because the model remembers previous images in a single conversation, you can break a complex task into smaller steps. Google’s example is decorating an empty room one piece of furniture at a time: add the sofa, check it, add the lamp, check it, adjust the rug, and so on.
This step-by-step approach is also the fix for vague instructions. Instead of asking for a finished room in one prompt and hoping, you build it in stages, correcting course at each step. Each generation becomes the starting point for the next, and the conversation holds the context together.
The practical rule is one change per message. When a message asks for five changes at once, a failure in any one of them costs you the other four. When each message asks for one change, you keep the four that worked and redo only the one that did not.
Combine Up to Three Images
Nano Banana can combine up to three images to create something new. Merge disparate photo elements, blend objects, colours and textures, or take a pattern from one image and apply it to an object in another: Google’s example puts one pattern on a wall and another on the same woman’s dress, from three input images.
The same combining power restores the past. Nano Banana can restore and colourise old black-and-white photos with an understanding of historical context, turning a faded family photograph into a colour image that still looks like the original, not like a new photo of strangers.
It can also go the other way, from rough to real: turn a simple sketch into a realistic scene, or annotate real-world objects, using its knowledge of the world to interpret a vague starting point and fill in the details contextually rather than randomly.
Five Edits Worth Trying First
Start with a headshot. Upload a portrait and ask for a cleaner background and better light, keeping the face unchanged. It is the fastest way to feel the model’s restraint: a good result changes the surroundings and leaves the person alone.
Next, try the figurine. Turn a selfie into a collectible figure with one prompt, and compare the face with the original. This is the consistency test in its purest form, and it shows immediately whether the likeness survived the transformation.
Then try a decor edit. Take a photo of a room and change one thing, the sofa’s colour, the curtains, the rug, and watch the rest of the room stay still. This is pixel-perfect editing in action, and it is the closest thing to a traditional photo editor’s job.
Fourth, restore something old. Scan a black-and-white family photo and ask for a colourised restoration. The model uses its historical context to choose plausible colours, and the result is a photograph that looks like the original, only in colour.
Finally, combine images. Upload a portrait, a pattern and a location, and ask for the person in that pattern in that place. Three inputs, one output, and the clearest demonstration that Nano Banana works with images as materials, not just as subjects.
Prompting That Works
First, name what must stay. “Keep the same woman, the same dress and the same pose; change only the background to a beach at sunset” protects the parts you care about and frees the model to change the one part you named. Prompts that only describe the destination leave the model free to redesign everything.
Second, describe the change concretely. Colours, objects, places and actions are good prompt material; moods and adjectives alone are not. “Deep green velvet sofa” beats “nicer sofa” every time, because the model does not have to invent what you meant.
Third, iterate in the open. If an edit drifts, say what drifted: “the face changed, keep the original face and only change the jacket.” The model’s memory of the conversation is an asset, but only if you tell it which parts of the last result were right and which were wrong.
Fourth, label your inputs when combining images. “Use the woman from the first image, the pattern from the second, and the beach from the third” gives the model a job for each picture. Without labels, it has to guess which element came from where, and guesses are where blends go wrong.
Fifth, remember that every edit is a new generation, not a retouch of the original file. Fine detail can shift slightly with each pass, which is why the one-change-per-message rule matters: fewer changes per pass means fewer details that can drift.
From Edits to Projects
Nano Banana is integrated into Canvas in the Gemini app and into Google AI Studio, so image ideas can become small apps. One of the first such apps was an internal project that became PictureMe: upload one photo, choose a theme like an ’80s Mall Shoot or Pro Headshots, and see yourself transformed across six different images in that style.
That template idea generalises. A set of headshots in different styles, a product photographed in different settings, a character shown across a story: all of these are multi-image projects where consistency is the whole job, and Nano Banana’s memory across generations is what makes them feasible without a design team.
It also covers pure creative play. Google notes people using Nano Banana for everything from better headshots for professional profiles to out-of-this-world sci-fi photos, and both ends of that range use the same skills: one anchor image, one change at a time, and the likeness carried through.
For creators, the practical project is a series, not a single image. Plan the series first, generate the anchor image, and then derive the variations from it in conversation, keeping the anchor in the thread so every variation refers back to the same source.
Before You Share an Edit
Check the result the way an editor would: the face, the hands, the small details, and the edges where the change meets the original. Generated images are most convincing at a glance, and the flaws, when they appear, hide in exactly those places.
Be honest about what the image is. An edited photo is not the original photograph, and a restyled portrait is not a portrait sitting. For personal projects that distinction barely matters; for anything presented as evidence, documentation or journalism, it matters completely, and the edit should be disclosed.
Finally, keep the original. Nano Banana returns a new image, and the thread that made it can be long. If you ever need the unedited photo again, or want to try a different direction from the same starting point, the original file is the only clean way back.
What Nano Banana Cannot Do Yet
First, vagueness is still the enemy. The model interprets vague instructions by applying logic to fill in the blanks, which is a strength when the blanks are small and a surprise when they are large. A prompt that leaves the whole scene undescribed will get a scene the model invented, not the one in your head.
Second, consistency is strong, not perfect. The model preserves likeness across edits and angles, but “looks like you” is not “is a photograph of you.” For identity-critical uses, check the face, the hands and the small details before publishing, because those are where generated images still slip.
Third, it works on images, not measurements. It knows what rooms and objects look like, but it has never measured your room. A restyled interior is a visualisation, not a floor plan, and it should be treated as inspiration rather than specification.
Background
Nano Banana arrived in the Gemini app in late August 2025 as Gemini 2.5 Flash Image, and by early October it had powered more than 5 billion creations. Its appeal is simple: editing by conversation, with the same faces surviving the conversation. The workflow it introduced, describe the change, check the result, refine in the same thread, is the durable part. Browse the subjects we cover on our topics index.
Sources: Google Blog: “Nano Banana tips”; Google Firebase: “Gemini image docs”.
0 Comments