What Is Nano Banana?
Nano Banana is Google's hosted Gemini image model for generation and editing. Nano Banana 2 is Gemini 3.1 Flash Image, and the weights stay in Google's cloud. You can call it from a local Python script, or self-host a separate open-weight alternative such as Qwen-Image-Edit-2511, HiDream-O1-Image, or HunyuanImage 3.0. This independent overview keeps those two routes separate.
For a Nano Banana 2 local model or how to run Nano Banana locally, use the install guide. It covers the API client only: Google has not released a downloadable Nano Banana 2 checkpoint. Short answers, including whether Nano Banana is open source, are on the FAQ.
Model selection checked : Google now recommends Nano Banana 2.1 for new projects. Nano Banana 2 examples here remain explicitly versioned; a newer API model does not change the cloud/local distinction.
Try Nano Banana in LMArena
Validate text‑guided image editing in minutes. Try real prompts and compare outputs.
Choose a Local or Cloud Route
| Your requirement | Where inference runs | Next step |
|---|---|---|
| Use Nano Banana without a local GPU | Google's cloud; your computer runs the API client | Set up the Python client |
| Keep inference on your own hardware | A separate open-weight model on your machine | Read the local AI image generator guide |
| Run open weights without owning suitable GPUs | A rented machine, not your local computer | Plan a cloud GPU deployment |
For privacy, hardware, cost and maintenance trade-offs, read cloud APIs versus local image models. A rental is still remote infrastructure, not an offline solution.
Can you download Nano Banana?
No. Nano Banana is Google's hosted Gemini image model. There are no downloadable model weights, no official desktop app or .exe, and it cannot run offline.
You can use it in the Gemini app or on the web, in Google AI Studio, or call the Gemini API from a local client or script. The quickstart is the getting-started path, and the local guide covers that API client. You can also download the images you generate. Those steps still need a network connection, because the model stays in Google's cloud.
If you need weights that actually run on your own machine, choose a separate open-weight image model. That is not a Nano Banana download.
Local Models & Nano Banana 2
Nano Banana is a cloud image generation and editing route. This site separately covers self-hosted alternatives and links to their official weights, licenses, and setup instructions. Compare the routes without confusing a local API client with a locally running model.
Benchmarks & Evidence
Our first-party comparisons publish inputs and settings where they are available. The broader leaderboard is a dated snapshot from Artificial Analysis and is labeled as third-party data. Treat model rankings as changing evidence, not permanent product claims.
Inspect a disclosed editing example or learn how to interpret image-model Elo scores.
Why use this guide
Spatially Targeted Edits
Use text and image inputs to request changes such as background replacement or object removal. “Targeted” describes the edit area; it does not mean the Google model runs on-device.
Character Consistency
Inspect character consistency and scene preservation with disclosed inputs and prompts. Treat outputs without complete run metadata as examples, not rankings.
Evidence Disclosure
We publish inputs and settings that were actually retained, and explicitly identify missing model versions, parameters, masks, or selection records.
Choose the Correct Route
Use a lightweight Python client for Google's cloud model, or follow an official repository for true offline inference. Hardware needs vary dramatically by open model.
Real‑world use cases
Social profile updates
Request outfit, background, or lighting changes, then check identity details across multiple outputs.
Product photography
Test background replacement, props, and targeted retouching while reviewing product shape, labels, and lighting.
Marketing creatives
Iterate banners and posters, then verify subject pose, typography, trademarks, and visual artifacts.
Content creation
Draft visual variants for blogs and social media while keeping a human quality and rights check in the workflow.
How We Handle Evidence
Primary sources first
Model identity, licensing, release status, and setup commands should link to the provider's documentation, repository, or model card.
Dated measurements
Rankings, pricing, and availability change. We date snapshots and avoid presenting volatile values as permanent facts.
No invented testimonials
We publish attributable user feedback or reproducible tests—not anonymous marketing quotes presented as customer evidence.