Nano Banana Pro is Google DeepMind's Gemini 3 Pro Image model for generating and editing images from text and visual references. Developers access it through Google's hosted Gemini API, AI Studio, or Vertex AI; Google provides no open weights for self-hosting. Its current documented API identifier is gemini-3-pro-image. Developer announcement and model documentation.
What are the key facts about Nano Banana Pro?
| Developer | Released | Type | Size or parameters | License and access | Where it runs |
|---|---|---|---|---|---|
| Google DeepMind. Announcement | November 20, 2025, initial release. Announcement | Image generation and editing with text and image inputs and outputs. Model page | Not published in the model documentation. Model page | Hosted service under Google's API terms; no open weights. Terms and developer access | Google's infrastructure through Gemini API, AI Studio, and Vertex AI. Developer announcement |
For an application, the first design decision is the request's role. Is it producing a draft for someone to choose, editing an approved reference, or supplying an image for an existing layout? Write that role down before choosing configuration fields. It determines what a successful response must contain and who reviews it.
The sibling Nano Banana Pro prompting guide focuses on the creative brief. This page focuses on reaching the model, handling its response, and keeping the integration understandable.
What can developers build with Nano Banana Pro?
Google's developer announcement documents composition controls, reference-image blending, text rendering, and search-grounded visual generation. Its examples include combining product imagery and logos into mockups. These capabilities make Pro a candidate for applications whose inputs include both written instructions and existing visual assets. Developer capabilities.
The current model page lists image and text as inputs and outputs, with search grounding and thinking supported. Your integration should be designed around image creation and the specific output your application needs. Supported capabilities.
A useful application brief might ask for a catalog illustration with a product centered and room for a heading. Keep the product reference, placement instruction, and desired output format visible in the request-building code. Reviewers should be able to trace an unwanted result back to the input that requested it.
Separate the creation of the image from the decision to use it. Design a review state in which the result can be inspected, accepted, or rejected. This is a suggested application pattern, not a claim that the model exposes a built-in approval system.
What limitations should a Nano Banana Pro integration handle?
Google's Pro prompting guidance identifies small text, spelling, factual visuals, translation, complex edits, and character features as remaining sources of error. A response can contain a usable image file while still failing the business requirement. Validate visible content as well as the success of the request. Published limitations.
The model page does not publish a parameter count. Hosted API access also supplies no basis for an asserted local VRAM requirement. Do not turn undocumented model architecture into deployment requirements for the client application. Model documentation and developer access.
Google's API image guide says the requested number of images may not be followed exactly. Its examples demonstrate saving output_image for an image response and iterating through content blocks for interleaved text and images. Match your response parser to the output shape you request. Response handling.
For a first integration, request one image with an explicit purpose. Decide what the application should show if there is no image, if the request fails, or if a reviewer rejects the content. Keep these states separate so an operational error does not look like a creative rejection.
How do you call the Nano Banana Pro API?
Create a Gemini API key in Google AI Studio for the project you intend to use. Follow Google's billing setup for that project and place the key in the GEMINI_API_KEY environment variable. Google's documentation covers both project-level key management and the billing relationship. API-key setup and billing.
Install Google's SDK with python -m pip install google-genai. SDK installation. The following example uses the current image guide's Interactions API, selects the stable Pro identifier, requests a PNG image response, and writes the returned bytes. It is a minimal integration example, not an end-to-end production application. Image API guide and Pro model identifier.
import base64
from pathlib import Path
from google import genai
result = genai.Client().interactions.create(
model="gemini-3-pro-image",
input="Create a catalog illustration of a desk lamp with space for a heading.",
response_format={"type": "image", "mime_type": "image/png", "image_size": "2K"},
)
if result.output_image is None:
raise RuntimeError("The response did not contain an image")
Path("catalog-draft.png").write_bytes(
base64.b64decode(result.output_image.data)
)
Run one request and inspect both the saved file and the returned response. Keep the selected model, output configuration, prompt, and a reference to the result in your own application record. Record the actual request settings instead of relying on what a user remembers selecting in the interface.
For editing, Google's guide shows passing a text instruction together with image content. Use that documented input shape and keep reference assets identifiable in your application. If several images have different roles, describe those roles explicitly in the prompt so the intended composition can be reviewed. Image-input examples.
If the task requires current information, review Google's search-grounding documentation for image generation. Pro's model page lists that capability, but your acceptance criteria should still require checking the visual's facts. Search support is an input to the process, not a substitute for reviewing the result. Capabilities and factual limitations.
Before integrating a larger queue, agree on rejection and retry behavior. Keep a rejected visual distinct from a failed request, retain enough context to reproduce the intended task, and make a later model change visible in the application record. These are implementation choices to adapt to your workload.
How does Nano Banana Pro compare with Nano Banana 2?
Nano Banana 2 is a real alternative using the same broader Gemini developer ecosystem. Google's model page describes it as Gemini 3.1 Flash Image and positions it for efficient image generation and conversational editing. Pro's page emphasizes complex graphic design and product mockups. Nano Banana 2 and Nano Banana Pro.
| Integration decision | Suggested evaluation |
|---|---|
| General image creation | Test Nano Banana 2 against a representative brief. Model description |
| Complex product composition | Test Pro with the actual reference assets. Model description |
Those are reasons to run a comparison, not measured rankings from this article. Evaluate acceptance rate and review effort on the same workload. The ComfyUI guide covers the broader workflow if a visual interface is the next integration step.
What else should developers know about Nano Banana Pro?
What model ID should a developer use for Nano Banana Pro?
The current model page lists gemini-3-pro-image as stable. Check that page when updating an integration that was built from an older preview example. Model documentation.
Is Nano Banana Pro an open-source model?
Google provides Pro through hosted services and does not provide open weights. The API's client code and documentation do not constitute a model-weight release. Developer access.
Can Pro generate and edit images?
Yes: Google's developer announcement covers both generation and reference-based editing. Its model documentation lists text and image inputs and outputs. Developer announcement and model page.
Should I automatically publish every returned image?
Plan a review step appropriate to the task, because Google documents errors in text, factual visuals, and edits. A returned image should be checked against the requirements before it is accepted for use. Limitations.
Sources
Primary sources checked on September 5, 2026.
- Google developer guide to the Nano Banana Pro launch
- Google model documentation for Gemini 3 Pro Image
- Google announcement of Nano Banana Pro
- Google Gemini API terms
- Google prompting guidance and limitations for Nano Banana Pro
- Google Gemini API image-generation documentation
- Google documentation for Gemini API keys
- Google Gemini API billing documentation
- Google model documentation for Gemini 3.1 Flash Image
- Google Gemini API SDK installation documentation
Top comments (0)