Generate or edit images from plain instructions in the browser, with up to eight reference images, two model tiers reaching 4K, and credits metered at two per image.
Used this tool? Rate it
Used this tool? Rate it
Nano Banana at banana-nano.org is a browser workspace for generating and editing images, operated by SupaFusion. It is a front end onto Google's Gemini image models rather than the maker of those models: the site states that its standard tier runs on Gemini 2.5 Flash Image and its Pro tier on Gemini 3 Pro Image.
That distinction determines what the site actually controls. Image quality, subject handling and watermarking come from the underlying model; what this operator sets is the interface, the credit system, the plan structure, the upload handling and the terms — and those are what differ between one such site and another.
The workspace itself is deliberately narrow. Two modes, a prompt box, a model selector, an aspect-ratio control and an input area that takes up to eight images. There is no canvas, no layer stack and no selection tool; every change is expressed as an instruction in words. That constraint is the product's actual proposition: it trades the precision of a conventional editor for the ability to describe a change and have the model work out where it applies.
| Capability | Standard | Pro |
|---|---|---|
| Resolution | Up to 1K | Up to 4K |
| Text rendering | Basic | Advanced and accurate |
| Studio controls | Not available | Available |
| Localization | Not available | Available |
| Subject consistency | Standard | Up to 5 people |
| Real-world knowledge | Good | Advanced with grounding |
The gap is widest where it usually matters most in commercial work: legible text in the image and output resolution. If a design needs readable words rendered inside the picture, or an asset destined for print rather than screen, the comparison puts that squarely on the Pro side. The subject-consistency line points the same way — a single figure is within the standard tier's stated scope, while a group scene with several recognisable people is described only for Pro.
One inconsistency is worth flagging. The tier table names the Pro engine Gemini 3 Pro Image, while a banner higher on the same page calls it Gemini Image 3. The two labels are not written identically, so anyone matching this against a model's official documentation should verify which is meant rather than assume they are interchangeable.
The site publishes four worked examples with their prompts, and they map to distinct jobs. Garment transfer puts clothing from a reference onto a model to show the fit — the e-commerce case, where the product exists as a flat photograph and the requirement is to see it worn. Multi-image composition replaces an object in one picture with an object from another: the advertising case, a held item swapped for the product being sold.
Background replacement changes the setting behind a subject while keeping the subject intact, useful for product shots taken against whatever was available. Style conversion turns a photograph into a named illustration style with a one-line prompt — the cheapest of the four to attempt and the one most exposed to questions about the referenced style's provenance.
What connects them is that all four are single-instruction edits on existing images rather than generation from nothing. That is where an instruction-driven interface with strong subject consistency earns its place over a general image generator: the value is in changing one thing about a picture that already exists while everything else survives the edit intact, which is precisely the operation a text-to-image generator handles worst.
It is a poor fit for anyone needing a free evaluation, since the pricing page lists four paid tiers and no free tier. It is also unsuited to automated use: the terms prohibit robots, spiders, scraping and equivalent access, and grant only a personal-use licence that bars service bureau use. Teams with strict data-handling requirements should read the privacy section below before uploading client material.
The service runs in the browser. Uploading, prompting, tier selection and refinement all happen on the site, and no desktop application or documented public API is presented.
The privacy policy names Google Play Services and App Store Services among the third parties that may collect identifying information, alongside Google Analytics and Google Ads. No application for this product is presented on the site itself, so the browser is the actual surface.
Purchasing happens off-platform. Orders are processed by external resellers acting as merchant of record depending on the buyer's location, and those resellers — not the site operator — handle customer service enquiries and returns. That is a meaningful platform fact rather than a billing detail: it determines who answers when something goes wrong with a payment, and it means the entity on your card statement is not the entity running the site.
Four paid tiers are published, all priced as annual-billing equivalents against a higher reference price, and all marked as billed annually with cancellation available at any time.
| Plan | Monthly equivalent | Annual total | Credits per year | Credits per month | Images per month |
|---|---|---|---|---|---|
| Starter | 4.90 USD | 58.80 USD | 1200 | 100 | Up to 50 |
| Pro | 9.90 USD | 118.80 USD | 3600 | 300 | Up to 150 |
| Ultra | 19.90 USD | 238.80 USD | 9600 | 800 | Up to 400 |
| Max | 59.90 USD | 718.80 USD | 33600 | 2800 | Up to 1400 |
The credit-to-image ratio is fixed at two to one on every tier, so an image is metered the same way whichever plan you buy. Moving up buys volume, not a better rate — which makes the choice a straightforward capacity question rather than a value calculation.
The feature lists reinforce that: premium quality, priority processing, ad-free use, watermark-free output and email support appear identically on all four. Only support level varies, with Ultra adding priority support and Max adding dedicated support.
Three commercial terms sit behind the table, all from the document that names another product, so treat them as published rather than settled. Refunds are written as discretionary — fees may be fully or partially refundable in specific scenarios and the operator decides who qualifies. Prices may change without notice for new customers, while existing subscribers keep the fee agreed at signup. Listed prices exclude VAT.
Because this is an interface onto a model rather than a model, the real comparison is between access routes rather than between image models. The operator named in these terms is not the maker of the underlying engine, so what this route actually decides is the credit metering, the upload ceiling, the payment channel and the support boundary — the model behaviour itself is the model owner's.
An analysis published by a company that sells API relay access itself, frames it similarly: the best alternative is usually a different access route rather than a different image model, and first-party routes retain capabilities a relay may not reproduce. The question is not whether this site produces different pictures than another route to the same model family — it is whether its credit meter, terms and support suit you better than going direct, weighed against the inconsistencies documented below.
It is a third-party interface. The site is operated by SupaFusion and states that it runs on Google's Gemini image models — Gemini 2.5 Flash Image on the standard tier and Gemini 3 Pro Image on Pro. The models are Google's; the interface, credits, plans and terms are the operator's.
No. The pricing page lists four paid plans and no free option, so there is no documented way to test output quality at no cost before subscribing.
The input area provides eight slots and accepts PNG, JPG, JPEG, WebP, HEIC and AVIF. Note that the multi-image fusion description mentions up to three images, so the count that actually composes may be lower than the number of slots.
Resolution up to 1K versus up to 4K, basic versus advanced text rendering, studio controls and localization available only on Pro, subject consistency described as standard versus up to five people, and good versus advanced grounded real-world knowledge.
At two credits per image on every tier. Starter gives 100 credits and up to 50 images a month, Pro 300 and 150, Ultra 800 and 400, Max 2800 and 1400 — the same ratio throughout.
Possibly, but not by right as written. They state fees may be fully or partially refundable in specific scenarios and that the operator determines eligibility — in a document that names another product. Payments run through external resellers acting as merchant of record, who also handle returns.
The privacy policy says submitted content is stored temporarily during processing and may be retained for a limited time to improve the service. No retention period is given, so the duration cannot be determined from the published policy.
Not as the terms are written. They prohibit robots, spiders, page scraping, deep links and equivalent automated or manual processes for accessing or copying any part of the service, and state the licence covers personal use only — in that misnamed document.
