# Nano Banana Pro prompting: a guide to controlled image edits

> Published 2026-04-02, updated 2026-09-05 · https://www.promptzone.com/klaus_kamau/nano-banana-pro-lightweight-ai-art-generation-unveiled-2f9c

---
title: "Nano Banana Pro prompting: a guide to controlled image edits"
published: true
description: "Write Nano Banana Pro prompts that separate subject, layout, text, and editing constraints, then review generated assets against a concrete brief."
tags: ai, imagegeneration, generativeai, prompts
---
Nano Banana Pro is Google DeepMind's Gemini 3 Pro Image model for image generation and editing. You access it through hosted Google products or the Gemini API, where the current model identifier is `gemini-3-pro-image`. Google provides no open weights, so prompting it means directing a hosted image service. [Developer announcement][prodev] and [model documentation][pro].

## What are the key facts about Nano Banana Pro?

| Developer | Released | Type | Size or parameters | License and access | Where it runs |
| --- | --- | --- | --- | --- | --- |
| Google DeepMind. [Announcement][prolaunch] | November 20, 2025. [Announcement][prolaunch] | Image generation and editing from text and visual references. [Model page][pro] | Not published in the model documentation. [Model page][pro] | Hosted access under Google's API terms; no open weights. [Terms][terms] and [access][prodev] | Google's services, including Gemini API, AI Studio, and Vertex AI. [Developer access][prodev] |

A useful prompt is a brief you can evaluate. Before generating, identify the subject, the required change, and the details that must remain recognizable. Then decide what would make the result unsuitable. This turns editing from open-ended exploration into a task with a clear acceptance decision.

The sibling [Nano Banana Pro developer setup guide](/dalia_bernard/nano-banana-pro-googles-new-ai-tool-for-developers-517l) covers the integration. Here, the focus is the written instruction and the review of what it produces.

## What should a Nano Banana Pro prompt specify?

Google documents Pro's ability to combine visual references, change lighting and camera treatment, and render text in generated designs. Its prompting guide asks users to specify the subject, composition, action, location, style, and intended edit. Treat those as separate parts of a brief rather than a collection of decorative keywords. [Capabilities][prodev] and [prompting guidance][prompting].

Consider a ceramic-vase photograph. Your subject description might identify the vase's silhouette and glaze. Your edit might move it into a different room. Your preservation requirement might keep the rim and painted pattern recognizable. The intended result is now concrete enough to compare with the reference.

For a design containing words, write the required copy separately before describing typography and placement. Make clear which words belong in the image and which sentences merely instruct the model. Then review the rendered copy against the original text instead of reading it from memory.

For a composition with several references, assign each reference a role in your own brief. One may describe the subject, another the setting, and another the desired palette. Avoid asking a reviewer to infer which part of each source was meant to survive in the output.

## What can a detailed Nano Banana Pro prompt still get wrong?

Google's Pro guidance says that small text, spelling, factual diagrams, localization, complex blending, and character consistency can still fail. A detailed prompt improves the clarity of your request; it does not eliminate the need to review the image. [Published limitations][prompting].

Treat a preservation instruction as a requirement to inspect, not a guarantee. For the vase example, compare the rim, outline, decoration, and orientation before praising the new room. If an essential product feature has changed, mark that requirement as failed even when the surrounding scene is attractive.

Google's image-generation guide notes that the number of requested outputs may not be followed exactly. It also explains that generated images include SynthID. Plan around the actual response and keep the image's generated origin clear in your asset records. [Image-generation limitations][images].

The editing example below sends text and an image as request inputs to the hosted model. It does not change model weights; keep your reference assets and revised prompts as the record of the editing process. [Image-editing request format][images].

## How do you prompt Nano Banana Pro to edit a reference image?

Start with a written brief before opening an interface. Use the following original example as a structure to adapt to your own reference image:

> Edit the supplied photograph of a ceramic vase. Place it on a narrow oak shelf against a pale wall. Preserve the vase's outline, painted pattern, rim shape, and orientation. Use soft light from the left. Leave open wall space above the vase for a heading that will be added later.

For API access, create a key using Google's instructions, set `GEMINI_API_KEY`, and install `google-genai` with `python -m pip install google-genai`. [SDK installation][libraries]. Google's current guide shows sending base64 image content alongside text in an Interactions request. This example applies that pattern to the Pro model, explicitly requests PNG output, and saves the returned image. [Key setup][key], [image-editing API][images], and [model identifier][pro].

```python
import base64
from pathlib import Path
from google import genai

result = genai.Client().interactions.create(
    model="gemini-3-pro-image",
    input=[
        {"type": "text", "text": "Put this vase on an oak shelf. Preserve its outline, rim, and painted pattern."},
        {"type": "image", "mime_type": "image/png",
         "data": base64.b64encode(Path("vase.png").read_bytes()).decode()},
    ],
    response_format={"type": "image", "mime_type": "image/png"},
)
if result.output_image is None:
    raise RuntimeError("No image returned")
Path("vase-on-shelf.png").write_bytes(
    base64.b64decode(result.output_image.data)
)
```

Use a reference you can inspect clearly, and keep it alongside the output. Review the required edit first, then the preserved features, then lighting and background details. Write down the first failed requirement instead of responding with a general request to “make it better.”

If the rim changes, make the next instruction about that feature. If the product is correct but too small in the frame, make the next instruction about placement and scale. Keep successful aspects in the preservation portion of the revised brief. This is a suggested iteration method you can apply to your own task.

For text, begin with a short piece of approved copy and a simple placement requirement. Inspect each rendered word, punctuation mark, and line break. If the task depends on dense or exact typography, decide in advance which parts you will accept as generated artwork and which parts require separate finishing.

For factual diagrams, write the underlying information before requesting the illustration. Review the visual relationships as well as the labels: the correct words in the wrong arrangement can still fail the brief. Google's guidance specifically calls for verifying data-driven visuals. [Factual-accuracy guidance][prompting].

## How should you compare Pro and Nano Banana 2 prompts?

Nano Banana 2 is a real alternative for the same broad class of generative image tasks. Google's model page describes Gemini 3.1 Flash Image as a counterpart focused on efficient generation and conversational editing, while Pro's page emphasizes complex design and product mockups. These descriptions suggest tasks to test, rather than establishing a guaranteed winner. [Nano Banana 2][flash31] and [Nano Banana Pro][pro].

| Task | Comparison to run |
| --- | --- |
| Change a simple background | Give both models the same reference and preservation brief. |
| Build a complex product composition | Compare whether each required object and layout instruction survives. |
| Add approved text | Check exact copy and placement before judging visual style. |

For more prompt ideas to adapt, browse [AI prompts on PromptZone](/ai-prompts). Keep the written brief usable independently of the interface so a change of tool does not obscure what the image must achieve.

## What else should you know about Nano Banana Pro prompts?

### How do I preserve a product while changing its background?

Describe the requested background change and name the product features that must remain recognizable. Compare those features with the reference after generation, since Google documents limits in complex edits and consistency. [Limitations][prompting].

### Can Nano Banana Pro put readable text in an image?

Google documents text rendering as a Pro capability, while warning about spelling and small-text errors. Supply approved copy and inspect the result character by character. [Capability][prodev] and [limitations][prompting].

### Is a reference-image edit the same as fine-tuning?

The documented workflow sends reference content in an image-generation request. Google supplies this model through hosted access, without open weights for local training. [Image-editing API][images] and [access][prodev].

### How many prompt details should I include?

Include the details needed to judge the image: subject, required changes, preservation constraints, and intended layout. Remove instructions that conflict with those requirements, then revise the brief according to the specific failure you observe.

## Sources

Primary sources checked on September 5, 2026.

- [Google developer guide to the Nano Banana Pro launch][prodev]
- [Google model documentation for Gemini 3 Pro Image][pro]
- [Google announcement of Nano Banana Pro][prolaunch]
- [Google Gemini API terms][terms]
- [Google prompting guidance and limitations for Nano Banana Pro][prompting]
- [Google Gemini API image-generation documentation][images]
- [Google documentation for Gemini API keys][key]
- [Google model documentation for Gemini 3.1 Flash Image][flash31]
- [Google Gemini API SDK installation documentation][libraries]

[prodev]: https://blog.google/innovation-and-ai/technology/developers-tools/gemini-3-pro-image-developers/
[pro]: https://ai.google.dev/gemini-api/docs/models/gemini-3-pro-image
[prolaunch]: https://blog.google/innovation-and-ai/products/nano-banana-pro/
[terms]: https://ai.google.dev/gemini-api/terms
[prompting]: https://blog.google/products-and-platforms/products/gemini/prompting-tips-nano-banana-pro/
[images]: https://ai.google.dev/gemini-api/docs/image-generation
[key]: https://ai.google.dev/gemini-api/docs/api-key
[flash31]: https://ai.google.dev/gemini-api/docs/models/gemini-3.1-flash-image

[libraries]: https://ai.google.dev/gemini-api/docs/libraries

<!-- pz-related-guides -->
## Related guides on PromptZone

- [Best SDXL Models in 2026](/tara_suzuki/best-sdxl-models-in-2026-realistic-anime-and-all-purpose-checkpoints-116)
- [ComfyUI 2026: The Complete Guide](/tomas_novak/comfyui-2026-the-complete-guide-to-power-user-ai-image-generation-1g17)
- [AI Model Releases Timeline](/ai-model-releases)