---
title: Gemini 3 Pro image (Nano Banana Pro) early preview
url: https://calvin.my/posts/gemini-3-pro-image-nano-banana-pro-early-preview
published: 2025-11-23
updated: 2026-09-16
category: AI
tags:
- gemini
- Nano Banana
- Image
summary: An early API review of Google’s Nano Banana Pro image model examines generation, reference-guided output, conversational editing, mixed realistic and illustrative scenes, and handwriting-style math solutions. The preview supports selectable resolutions up to 4K and as many as 14 reference images. Tests produced strong results, but requests took two to four minutes. Developers must handle mandatory reasoning signatures, higher token use, large signature storage, and sizable 4K files, favoring background streaming for long-running jobs.
---

# Gemini 3 Pro image (Nano Banana Pro) early preview

Google has recently added [Nano Banana Pro](https://deepmind.google/models/gemini-image/pro/) to its offerings. This article performs several tests on the API itself and the new capabilities of the model.

Related articles:

1. [Gemini "Nano Banana" image editing capability](../posts/gemini-nano-banana-image-editing-capability)
2. [Gemini "Nano Banana" generally available](../posts/gemini-nano-banana-generally-available)

* * *

## API Updates

1. The model name is `gemini-3-pro-preview`. You can use it with the generateContent or streamGenerateContent endpoints.
2. This model enforces a mandatory thinking process. The content parts would return the `thoughtSignature`, which must be stored. Conversation could not continue without the signature being attached as part of the input.
3. It is possible to specify the image output size, such as 1K, 2K, or 4K. Sample:

   ```json
   {
      ......
       "generationConfig": {
           "maxOutputTokens": 32000,
           "imageConfig": {
               "aspectRatio": "16:9",
               "imageSize": "1K"
           },
           "thinkingConfig": {
               "thinkingBudget": 3000
           }
       }
   }
   ```

* * *

## Test #1 - 4K Image

1. The first test is a simple, straightforward request to generate an image based on a given prompt. The streaming takes 3 minutes to complete, and a JPEG image is received.

   ![](https://camy-pub.s3.ap-southeast-1.amazonaws.com/19c476b4-86fa-40d5-91be-f2ae895e3f70.png)

2. Check the image's metadata, and we can see it produced a 4K image.

   ![](https://camy-pub.s3.ap-southeast-1.amazonaws.com/0646a9de-c5d6-4d6d-8ef6-c0913dd5db27.png)

* * *

## Test #2 - With reference input images

1. The model allows up to [14 input images](https://ai.google.dev/gemini-api/docs/image-generation?batch=file#use-14-images). In this test, we provide several images as a reference and observe how they are being used in the output.

   ![](https://camy-pub.s3.ap-southeast-1.amazonaws.com/6ad6061b-fd3d-40b8-9dea-2ee18c851656.png)

2. The output looks great! Especially since it is done in 4K and a 16:9 ratio.

   _(This is a scaled-down version of the original output.)_

   ![](https://camy-pub.s3.ap-southeast-1.amazonaws.com/af47fd91-d1c3-4efe-aba5-0c814287974b.jpg)

* * *

## Test #3 - Conversation & Image editing

1. In this test, we first generate an image.

   ```markup
   Create a teddy bear image, sit behind the glass in a toy store. street view: modern tokyo. 
   slightly cloudy. only a few ppl on the streets. no cars.
   ```

   ![](https://camy-pub.s3.ap-southeast-1.amazonaws.com/1e4afc75-e5a9-423c-b473-592a427970e4.jpg)

2. Then request an edit to it.
   ```markup
   add some Christmas decoration around the bear. keep the street unchanged.
   ```

   ![](https://camy-pub.s3.ap-southeast-1.amazonaws.com/aabbc2c3-50fd-489d-9029-1e892187997d.jpg)

## Test #4 - Combination of realistic & imaginary/illustrative

1. In the last test, we asked the model to generate an image of a computer screen showing a shooting game. 

   ![](https://camy-pub.s3.ap-southeast-1.amazonaws.com/a2b3b1db-fc85-4451-be5f-e977e38c8d48.jpg)

* * *

## Test #5 Combining Math & Handwriting Imitation

1. Given a photo and a prompt below:
   ```markup
   Solve this equation and write the steps with similar handwriting on the paper.
   ```

   ![](https://camy-pub.s3.ap-southeast-1.amazonaws.com/0d411c4a-179c-4c53-8859-7b5339106a0d.png)

2. It returns a solution with the correct answer, written in a similar handwriting style.

   ![](https://camy-pub.s3.ap-southeast-1.amazonaws.com/1298d684-a3bf-46dd-9a02-575651fdac6a.jpg)

* * *

## Things to take note of

1. Unlike the original Nano Banana, the new Pro model takes a much longer time (E.g., 2 to 4 minutes) to complete the request. It is recommended to use the streaming endpoint in a background process.
2. If you have a restriction on file size (either firewall, server, or app level), remember to increase the value. The 4K image is huge; for example, a 16:9 image is around 8MB each.
3. The Pro model utilises Gemini 3 Pro thinking capability. This is mandatory, and it increases the usage of tokens.
4. The size of thoughtSignature can be huge. We have seen a range of hundreds of KB to a few MB. If your use case does not require a conversation, you should discard them without saving the signature.