Nano Banana is the nickname for Google's image model, and it does two jobs: it generates a new image from a text prompt and it edits an image you give it. A prompt that works names the subject, the one change you want, what must stay the same, the style and the output format, in that order. This page gives you a template for edits and one for new images, the style recipes that hold up and the places where it breaks; the free guide adds use cases, tips and a mini-course.
What Nano Banana does
Two jobs, and they need different prompts.
Editing is the one people share. You upload a photo and describe a change: swap the background, change the outfit, remove the person in the corner, put the product on a marble counter. The photo stays recognisably itself and the change lands.
Generating is the familiar one. You describe an image and get one back. It gets interesting when you combine the two: generate a scene, then edit it until it is right.
The jobs I use it for: product shots on new backgrounds, thumbnails with a consistent look, headshots cleaned up for a website, mockups in a real setting, and a character kept consistent across a set.
The prompt template for an edit
An edit prompt has five parts. The labels are mine; the model reads plain sentences.
Image: what I uploaded, in one line
Change: the one thing to alter
Keep: what must stay identical (face, pose, logo, framing, colours)
Style: how the change should look, in a few words
Output: aspect ratio and anything about size or format
Filled in:
Image: a photo of a white sneaker on a grey studio floor.
Change: place the sneaker on a wet city street at night.
Keep: the sneaker exactly as it is, same angle, same laces, same scale.
Style: photoreal, neon reflections in the puddles, cool tones.
Output: square, product-photo framing, nothing else added.
The Keep line is the one that matters. Without it the model treats everything as fair game and you get a different shoe on a different street. One change per prompt: if you want the street and a new colourway, that is two prompts, and the output of the first becomes the upload for the second.
The prompt template for a new image
A generation prompt reads like a photographer's shot list.
Subject: who or what, with two physical details
Setting: where, and what else is in frame
Light: the source, its direction, the time of day, the mood
Style: photograph, illustration, render, or a named look
Composition: framing, angle, what is in focus
Output: aspect ratio
Filled in:
A ceramic pour-over coffee set, sand coloured, on a linen cloth. Setting: a
sunlit kitchen counter, a window out of focus behind. Light: soft morning
light from the right. Style: editorial product photograph. Composition:
45-degree angle, the kettle in focus, the cup soft. Output: 4:5 portrait.
Write the nouns first, then the light, then the style, and keep the feeling words ("cosy", "premium") for last. They do less than a named light source.
Style recipes that hold up
A style recipe is the Style line plus the Light line, kept together for reuse. Five I come back to:
- Studio product: "plain white backdrop, soft box light from above, slight shadow under the object, photoreal."
- Editorial portrait: "35mm film look, window light from the side, shallow depth of field, muted colours."
- Flat illustration: "flat vector illustration, four-colour palette, no gradients, no outlines."
- Interface mockup: "a laptop on a desk, screen showing the uploaded design, soft office light."
- Film still: "anamorphic widescreen, teal and orange grade, one practical light in frame, slight grain."
Keep a file of recipes that worked and paste them verbatim. The words that produced a look you liked are worth more than any list online, including this one.
Where Nano Banana breaks
| The limit |
What you see |
The workaround |
| Text in the image |
Lettering that drifts, misspells or blurs |
Leave text out and add it in a design tool afterwards |
| Hands and fingers |
Extra fingers, odd grips |
Crop hands out of frame, or describe them holding a named object |
| Exact counts |
Ask for seven apples, get six or nine |
Say "a small pile", or generate and fix the count in a second pass |
| Two changes at once |
One change lands, the other half-lands |
One change per prompt, chained |
| Identity across many edits |
The face or product drifts after several rounds |
Edit from the original upload each time, never from an edit |
| Real people and brands |
Refusals or altered results |
Use your own photos and your own marks |
Tips that save generations
Four habits that cut my failed runs.
Say what to keep before you say what to change. Re-upload the original for every new edit, because editing an edit compounds small drifts until the product is a cousin of itself. Write style in nouns: "35mm film" beats "cinematic", and "soft box" beats "nice lighting". Save the prompt with the image, because you will want the same look next week and you will have forgotten the exact wording.
What is inside the Nano Banana guide
The free download has four parts.
- Nano Banana Prompt Guide: the two templates above with worked examples.
- Prompt Templates, Use Cases, Limitations, Tips and Tricks: fill-in templates for the jobs above and the limits in detail.
- Nano Banana Mini-Course: short lessons from a first edit to a repeatable process.
- Lifetime access and updates: every revision as the model changes.
Drop your email in the form and it is yours.
If you work in ChatGPT too, the ChatGPT images guide applies the same idea as a JSON template, and the ChatGPT image prompts post has ready prompts. For the assistant itself, start with the Gemini guide, and the Gemini prompts in the library cover the text side.
