Nano Banana guide: Google image generation and capabilities

Nano Banana is Google DeepMind's image generation and editing model, officially named Gemini 2.5 Flash Image. You access it through Google's hosted developer services; its original release also brought the model to the Gemini app. Google provides no open weights for local inference. Model documentation and launch announcement.

What are the key facts about the original Nano Banana?

The original Nano Banana uses the stable identifier gemini-2.5-flash-image. Google's model page marks its earlier preview identifier as deprecated. Use the stable identifier when recording experiments so another reader can tell which model produced an image. Version documentation.

For browser access to the Pro variant, read the sibling Nano Banana Pro online guide.

What can the original Nano Banana do?

Google introduced the model for changes described in ordinary language: altering a photograph, combining references, and retaining a recognizable subject across edits. The launch examples include changes to backgrounds and clothing, plus combining a product with a different setting. These are documented capabilities, rather than independently measured success rates. Developer announcement.

Choose a capability before writing the test prompt. For generation, describe an image that does not require a reference. For editing, supply a photograph and specify a change. For image blending, explain which part of each reference belongs in the final composition. Google demonstrates these different request types in its release examples. Capability examples.

Make the acceptance question match the task. A new illustration needs the requested subject and layout. An edit also needs the important reference details to remain recognizable. A blend needs the intended elements from each source. Keep these requirements in your brief so a visually appealing result does not obscure a missing element.

For a dedicated reference-photo workflow, use the sibling Nano Banana photo-transformation guide.

What are Nano Banana's image-generation limits?

Google's release documentation identifies long passages of rendered text, consistent characters, and factual visual details as areas needing improvement. A convincing picture is therefore insufficient evidence that a label, diagram, or depicted object is accurate. Inspect the details that matter to the intended use. Documented limitations.

The model page does not publish a parameter count, and it does not offer downloadable weights. A number in the nickname cannot establish model size. Likewise, a local computer used to send an API request does not become the machine performing Google's model inference. Model documentation and hosted access.

Google also documents that requested image counts may not be followed exactly and that Gemini-generated images carry SynthID. Build your review around the outputs actually returned. Do not assume that a written request for several alternatives guarantees that many usable files. Image-generation limitations.

Before reviewing an image, decide what would make it unacceptable. For a product illustration, a missing feature may matter more than an attractive background. Keep required details separate from optional styling when deciding whether to accept the result.

How do you generate an image with the original Nano Banana?

For a developer-controlled test, use the stable model through the Gemini API. Create an API key in Google AI Studio and configure it as GEMINI_API_KEY; Google's key documentation explains project selection and environment-variable handling. Keep that key in your execution environment. API-key setup.

Install Google's Python SDK with python -m pip install google-genai. This text-to-image example uses the SDK's documented image-response configuration and the original model's stable identifier. The prompt is an original example, and the code saves image bytes with the extension indicated by their MIME type. SDK request patterns, response types, and stable model.

import mimetypes
from pathlib import Path
from google import genai
from google.genai import types

response = genai.Client().models.generate_content(
    model="gemini-2.5-flash-image",
    contents="Create an illustration of a blue mug on a pale stone table.",
    config=types.GenerateContentConfig(response_modalities=["TEXT", "IMAGE"]),
)
for index, part in enumerate(response.parts or []):
    blob = part.inline_data
    if blob and blob.data and (blob.mime_type or "").startswith("image/"):
        suffix = mimetypes.guess_extension(blob.mime_type)
        if suffix is None:
            raise ValueError("Unknown image MIME type")
        Path(f"mug-{index}{suffix}").write_bytes(blob.data)
    elif part.text:
        print(part.text)

Read any returned text as well as inspecting the saved image. If the request returns no image, inspect the response before repeating it. Keep the prompt, model identifier, and accepted result together so you can reconstruct what you asked for.

Review the subject, color, setting, and layout against the generation prompt. If the mug is the wrong color, make the next request address that requirement. Keep the first result so you can compare the revision without relying on memory.

Once the image meets the brief, make a copy for delivery and keep the accepted source image. Record any further crop or layout changes separately. This makes it easier to explain whether a difference came from a generation request or from your own finishing work.

How does the original Nano Banana compare with Pro?

Nano Banana Pro is the relevant alternative when your task needs more elaborate compositions. Google's Pro announcement documents expanded reference-image blending, text rendering, and creative controls. That feature list provides reasons to test Pro; it does not establish that every Pro output will satisfy your brief. Pro announcement.

For workflow planning, use the ComfyUI guide. Compare complete tasks: the input preparation, requested edit, review effort, and usable result. A model name alone does not tell you how much work remains before delivery.

What else should you know about the original Nano Banana?

Is Nano Banana 1 a one-billion-parameter model?

Google identifies the original model as Gemini 2.5 Flash Image and does not publish its parameter count on the model page. “Nano Banana 1” is a way to distinguish the original release here, not a documented model-size specification. Model documentation.

Can I download Nano Banana weights?

Google's documented access is through hosted services, and no open weights are provided. Use the API or a supported Google interface to generate images. Developer access.

Can it edit an existing photograph?

Yes: Google documents natural-language image edits and combinations of reference images. Define the desired change and the details to preserve, then inspect the returned image against both requirements. Editing examples.

Does a successful generation prove the content is correct?

No: Google identifies factual visual detail and text rendering as limitations. Review labels, shapes, and other consequential details before accepting the image. Limitations.

Sources

Primary sources checked on September 5, 2026.

• Google announcement of the original Gemini image-editing update
• Google developer announcement of Gemini 2.5 Flash Image
• Google model documentation for Gemini 2.5 Flash Image
• Google Gemini API terms
• Google Gemini API image-generation documentation
• Google documentation for Gemini API keys
• Google announcement of Nano Banana Pro
• Google Gen AI Python SDK README
• Google Gen AI Python SDK response and image types

Related guides on PromptZone

• Best SDXL Models in 2026
• ComfyUI 2026: The Complete Guide
• AI Model Releases Timeline