# How do you edit photos with Nano Banana?

Date: 2025-09-06 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/edit-photos-nano-banana/

**Answer:** Upload a photo to the Gemini app, describe one change in plain words, and keep refining the result over several turns, one change at a time. Nano Banana, Google's Gemini 2.5 Flash Image model, is built to change what you ask for while leaving the rest of the picture, and the people in it, as they were.

Nano Banana is the nickname of Gemini 2.5 Flash Image, the image model Google released on August 26 in the Gemini app and for developers. Before the launch it appeared anonymously on a public comparison site under that name, and the name stuck. What sets it apart is editing: you give it an existing photo and describe a change, and it alters that part while keeping the rest recognizable.

Earlier image tools often redrew the whole picture when you asked for a small change, so a person's face could drift with each edit. Google says its work on this model concentrated on keeping people and pets looking like themselves from one image to the next.

## What you can ask it to do

Google describes four kinds of edits, and they cover most everyday needs:

- **Targeted changes:** blur a background, remove a stain from a shirt, take a stranger out of a holiday photo, or add color to a black-and-white picture.
- **Costume or location changes:** put the same person or pet in a new outfit or setting.
- **Blending photos:** combine two or more images, for example you and your dog, into one new scene.
- **Style transfer:** take the color or pattern of one image and apply it to an object in another.

The edits are written as ordinary sentences. There are no brushes, masks or layers to learn.

## A worked multi-step edit

Here is a realistic sequence for preparing a photo of a living room before repainting. Each step is a separate message in the same Gemini chat, with the photo uploaded at the start.

```prompt
1. Here is a photo of my living room. Remove the boxes on the floor and the cables behind the TV. Keep everything else exactly as it is.
```

```prompt
2. Now paint the back wall a soft sage green. Keep the lighting, the furniture and the window unchanged.
```

```prompt
3. Add a tall wooden bookshelf to the left of the window, filled with books and a few plants.
```

```prompt
4. Make one more version with the wall in warm off-white instead of green, everything else the same.
```

Three habits make this work. First, change one thing per message, so you can tell which instruction caused a problem. Second, name what should stay the same; "keep everything else exactly as it is" is a useful sentence to repeat. Third, if a step goes wrong, go back to the last good image rather than piling corrections on a bad one. You can download any version you like, and Google notes that a finished image can also be uploaded back into Gemini to turn it into a short video.

Developers can do the same through the Gemini API. This is the pattern from Google's announcement, using the preview model name published there:

```python
from google import genai
from PIL import Image
from io import BytesIO

client = genai.Client()   # reads your Gemini API key from the environment
prompt = "Paint the back wall a soft sage green. Keep everything else unchanged."
image = Image.open("living_room.png")

response = client.models.generate_content(
    model="gemini-2.5-flash-image-preview",
    contents=[prompt, image],
)
for part in response.candidates[0].content.parts:
    if part.text is not None:
        print(part.text)
    elif part.inline_data is not None:
        Image.open(BytesIO(part.inline_data.data)).save("living_room_v2.png")
```

To chain edits in code, open the saved image and send it with the next instruction. Google lists the price at $0.039 per generated image.

> **Try it:** Pick a photo with some clutter in it. In three separate messages, remove one object, change one color and add one item, each time asking Gemini to keep everything else the same. Compare the first and last images side by side.

## Watermarks and responsible use

Every image made or edited in the Gemini app carries a visible watermark and an invisible one called **SynthID**. SynthID is a signal hidden in the pixels that identifies the picture as AI-generated or AI-edited, even when you can't see it. Images made through the API carry SynthID too. If you share an edited photo, the honest move is to say it was edited with AI, especially if it shows real people or places.

Google's terms forbid some uses outright, including sexual images of people made without their consent. A good personal rule is simpler still: only edit photos of people who would be comfortable seeing the result.

## How do you keep a face consistent across edits?

Start from a clear, well-lit photo where the face is visible from the front. The model can only preserve what it can see, so a blurry or side-on picture gives it less to hold on to.

Say what must not change. A line such as "Keep her face, hairstyle and glasses exactly the same; change only the jacket" tells the model where its freedom ends. Repeat that line in later steps, because each new instruction is a fresh chance for small drift.

Keep the number of steps low for portraits. Several small changes in a row can add up, so for important photos, return to the original image and apply your final combination of changes in one or two messages instead of eight. When you place the same person in different scenes, use the same source photo each time rather than a previous edit.

Finally, check the result as a friend would: eyes, smile, skin tone and any distinctive features. Google says it is still improving consistency and fine details, so a close look before you share is part of the job.

## Sources

- Image editing in Gemini just got a major upgrade: https://blog.google/products/gemini/updated-image-editing-model/
- Introducing Gemini 2.5 Flash Image, our state-of-the-art image model: https://developers.googleblog.com/en/introducing-gemini-2-5-flash-image/
- Google Gemini's AI image model gets a 'bananas' upgrade: https://techcrunch.com/2025/08/26/google-geminis-ai-image-model-gets-a-bananas-upgrade/