Nanabanana2.run is an independent web platform that calls Google's Nano Banana 2 image model through its API. It offers 4K output, text rendering, character consistency and one-time credit packs.
Used this tool? Rate it
Used this tool? Rate it
Nanabanana2.run is a commercial web application that generates and edits images from text prompts or reference pictures. The single most important thing to understand before spending a cent here is the ownership question, because the product name invites confusion.
"Nano Banana 2" is not a name this website invented. It is Google's own nickname for a real image model. Google's DeepMind model card for Gemini 3.1 Flash Image lists that model publicly, and Google's own developer documentation uses the Nano Banana 2 label alongside the formal name. The model reached general availability in late February 2026 and is offered to outside developers through the Gemini API and Google AI Studio.
Nanabanana2.run is one of those outside developers. Its Terms of Service describe the platform as a service "that creates images from text prompts or reference images using Google's nanobanana 2 (Gemini 3.1 Flash Image Preview) model." The same document, in its disclaimers section, states plainly: "Our Services rely on Google's nanobanana 2 (Gemini 3.1 Flash Image Preview) API and other third-party services."
That single sentence settles the category. This is a third-party wrapper: an independent company buys access to somebody else's model, puts a web interface and a payment system in front of it, and resells generation capacity as prepaid credits. That is a legitimate and common business model. It is also a model with a specific risk profile that buyers should understand, because the value you receive depends on a supply relationship you cannot see or verify.
To its credit, this site does something many wrapper sites refuse to do: it names its upstream supplier, repeatedly and in writing, in documents that carry legal weight. Four separate sections of the Terms and two sections of the Privacy Policy identify the Google model by name. Many competing sites in this space describe their engine only in vague phrases like "our advanced architecture." Naming the supplier is a meaningful transparency signal, and it is the main reason this platform reads as a straightforward reseller rather than something more evasive.
Being transparent about your supplier is not the same as being clear about your relationship to that supplier, and this is where the site falls short.
Across the homepage, the pricing page, the Terms of Service and the Privacy Policy, a full-text search for the phrases "not affiliated," "unaffiliated," "independent," "trademark," "endorsed" and "disclaimer" returns nothing at all. There is no statement anywhere telling a visitor that this website is not operated by, sponsored by, or connected to Google.
Meanwhile the homepage says the product is "powered by Google's AI image generation technology," and the Nano Banana Pro page describes a "professional AI image generation model powered by Google." The Google name is used freely as a quality signal, while the absence of any relationship disclaimer leaves the nature of that relationship unstated.
So let this documentation state it, since the site does not: Nanabanana2.run is not Google. It is not an official Google product, service, or channel. Google's names and model names appearing on the site refer to a supplier whose interface the site calls, not to a partnership, endorsement, or corporate connection. Nothing found during research indicates any formal affiliation exists, and nothing on the site claims one in explicit terms either. The gap is one of omission rather than false assertion, but for a paying customer the practical effect is the same: a casual visitor could easily leave believing they bought from Google.
If you want Google's model from Google directly, the official routes are the Gemini app, Google AI Studio and the Gemini API. Those require a Google sign-in and are billed by Google. This site is an alternative front end, not a substitute identity.
The second disclosure gap is larger, and it deserves more weight than the affiliation issue.
The Terms of Service and Privacy Policy, both dated 11 November 2025, do not name a company. There is no legal entity, no registration number, no incorporation jurisdiction, and no business address in either document. The Terms refer to the operator only as "Nanabanana2.run" and as "we," "our," or "us." The sole contact channel is an email address.
The governing-law clause makes this concrete. Rather than naming a jurisdiction, it says the Terms "are governed by and construed in accordance with applicable laws," and that disputes will be resolved "through good faith negotiations or, if necessary, through binding arbitration or appropriate legal proceedings." Applicable to where? Arbitration under which body, seated in which country, under whose procedural rules? None of it is specified.
This matters in a practical, unglamorous way. If a payment is taken and the service stops working, if your stored images vanish, or if you dispute a charge, your written recourse is an email address and a clause that identifies no forum. Compare this to the payment side, which is fully specified: Stripe is named as the processor in both documents. The company is comfortable naming its payment vendor and its model vendor, but not itself.
For a small one-time purchase this may be an acceptable risk. For a business planning to build a production workflow on this platform, or for anyone in a regulated industry that requires a named counterparty and a defined jurisdiction, this is a genuine blocker. Treat the missing entity as the single most important limitation on this page.
Setting the disclosure questions aside, the functional description is reasonably specific and internally consistent, which makes it easier to evaluate than most sites of this type.
Text-to-image and image-to-image. The workspace exposes both modes directly on the homepage, and the Terms list both among the services. The prompt field on the main workspace accepts up to 20,000 characters, and the Nano Banana Pro page shows a 3,000-character field for that model, so long and heavily specified prompts are supported.
Text rendering inside images. This is the headline claim and the reason the product exists as a distinct offering. The marketing focuses on legible labels, mathematical notation, multilingual captions and dense infographic text, all of which were historically weak spots for image models. The underlying capability is genuine at the model level: Google's own model card for Gemini 3.1 Flash Image cites text rendering for posters and intricate diagrams, along with multilingual text rendering, as characteristics of the model.
Resolution up to 4K. The interface offers 1K, 2K and 4K options, with PNG and JPEG output formats. Google's model card confirms 4K output as a model capability.
Character and subject consistency. All three paid tiers list character consistency as an included feature. The Nano Banana Pro page goes further, claiming you can combine up to fourteen reference images in one output while holding as many as five characters consistent.
Batch generation. The workspace offers a choice of one, two, three or four images per run, and all three tiers list batch generation.
A multi-model catalogue on one balance. The platform is not limited to the Nano Banana family. It lists Nano Banana 2 Lite, Nano Banana 2, Nano Banana Pro, GPT Image 2, GPT-4o Image, Qwen Image 2, Seedream 4.0, 4.5 and 5.0 Lite, Flux Kontext Pro and Max, Flux 2 Pro and Flex, Z Image, and uni-1, plus the video models Veo 3 and Happy Horse 1.0. The pricing page states these share a single credit balance.
Packaged tools. Beyond raw generation, the site groups presets into named tools: an AI portrait generator, a commercial and product visual generator, a cover generator, a tattoo concept generator, a hairstyle changer, and a watermark remover.
A note on that last item, because it is unusual enough to flag. The watermark remover is described on the homepage as a tool to "Remove Gemini watermarks in batches." Google applies SynthID, an invisible provenance watermark, to imagery its models generate, and Google states this marking is designed to survive compression and re-uploading. A tool explicitly aimed at stripping provenance marks from AI imagery sits awkwardly beside the platform's own dependence on that same supplier, and anyone using generated media in regulated, journalistic or disclosure-sensitive contexts should think carefully before using such a feature.
Three separate internal inconsistencies turned up during research. None of them is hidden in fine print; each comes from comparing two of the site's own published pages. They are listed here because a buyer benefits more from knowing where the documentation disagrees with itself than from a tidy summary that papers over it.
Contradiction one: how many credits an image costs. The Terms of Service, section 5.1, states that "Standard image generation: approximately 4 credits per image." The live pricing page publishes a detailed per-model rate table that says something different: Nano Banana 2 costs 5 credits at 1K, 8 at 2K and 12 at 4K, while Nano Banana Pro costs 6 credits at 1K and 2K and 12 at 4K. Only the separate, older "Nano Banana" model is listed at 4 credits.
The pricing table is the one that matches the plan arithmetic. A 1,000-credit pack divided by 5 gives the 200 images the page advertises for Nano Banana 2 at 1K; divided by 6 it gives roughly the 166 Pro images quoted; divided by 12 it gives the 83 4K images quoted. The Terms figure of 4 credits does not reconcile with any advertised capacity. Treat the pricing page table as authoritative, verify the live quote in the workspace before a large run, and be aware that the legal document understates the cost of a standard image by roughly 25 to 50 percent depending on the model you pick.
Contradiction two: whether your content trains their models. This one affects your rights rather than your wallet, and it is the more serious of the two.
The Terms of Service, section 8.1, commits that "We do not use your content for purposes other than providing our Services unless you explicitly consent." That is an opt-in promise: nothing beyond service delivery without your active agreement.
The Privacy Policy, section 3.2, lists among the standard purposes for processing your information: "Training and enhancing our AI models (you may opt-out by contacting us)." That is an opt-out default: training happens unless you write in and ask them to stop.
These cannot both be operative. One says consent must be given before secondary use; the other says secondary use is the default and you must actively refuse it. Section 7.2 of the Terms adds a third shade, granting the operator a licence to process inputs "solely for the purpose of generating images and improving our Services," where "improving our Services" is exactly the ambiguity the other two clauses disagree about.
If your prompts or uploaded reference images are confidential — unreleased product photography, client artwork, internal campaign material — do not rely on the more favourable clause. Assume the opt-out reading applies, email the operator to opt out explicitly, and keep the reply.
Contradiction three: which model actually runs. The Terms and Privacy Policy consistently identify the engine as Google's Gemini 3.1 Flash Image Preview. The Nano Banana Pro product page instead says that tier is "Built on Gemini 3 Pro Image architecture." Those are different models within Google's lineup. Since the platform sells the Pro tier at a higher credit rate on the strength of superior quality, the discrepancy is worth noting rather than assuming the pages are loosely worded.
Pricing is public, does not require an account to view, and is unusually well documented for this category. What follows is what the pricing page published at the time of research; promotional pricing changes, so confirm current figures before purchase.
The model is prepaid credits sold in one-time packs, not a recurring subscription. Three tiers were listed, each showing a headline price alongside a higher struck-through price:
| Pack | Advertised price | Listed original | Credits | Approx. images at 1K (Nano Banana 2) | Approx. 4K images |
|---|---|---|---|---|---|
| Basic | $19.90 | $39.90 | 1,000 | up to 200 | up to 83 |
| Pro | $49.90 | $99.90 | 3,300 | up to 660 | up to 275 |
| Max | $99.90 | $199.90 | 7,500 | up to 1,500 | up to 625 |
Per-model credit costs published on the same page: Z Image 1 credit per image; Nano Banana 3 credits; the older Nano Banana 4 credits; GPT-4o Image, Qwen Image 2 and Flux Kontext Pro 5 credits; Nano Banana 2 five credits at 1K, eight at 2K, twelve at 4K; Nano Banana Pro six credits at 1K and 2K, twelve at 4K; Flux Kontext Max 8 credits; GPT Image 2 and Seedream 4.0 and 4.5 in a 3 to 10 credit range; Seedream 5.0 Lite and uni-1 in a 5 to 10 range; Flux 2 Pro 5 to 7; Flux 2 Flex 14 to 24.
Several practical points follow from this structure.
Credits are stated not to expire, in the plan features, the pricing FAQ and the Terms, which is genuinely more forgiving than a monthly subscription that resets unused allowance.
The advertised per-image prices on the homepage — twelve cents on Basic, nine on Pro, eight on Max — are computed against the Pro model at lower resolution. Work primarily in 4K and the effective cost roughly doubles, because 4K costs twelve credits against six. Budget from the rate table and your actual resolution, not from the headline figure.
The wide ranges on some models are unavoidable rather than evasive, since those families price by output settings. But it means a spend estimate for Flux 2 Flex could be off by a factor of nearly two, so test at small volume before committing a large balance to an unfamiliar model.
The discount framing deserves scepticism. A "50% off" banner with a countdown timer showing under two hours remaining appeared on both the homepage and the pricing page during research. Perpetual countdown timers that reset are a common conversion tactic. Treat the lower number as the ordinary price and the struck-through figure as marketing.
Refund terms are contradictory in tone across pages, and the stricter reading should govern. The homepage FAQ says flatly that "we are unable to offer refunds" and advises testing with free credits first. The Terms are marginally softer, saying refunds for consumed credits "may not be available" while inviting case-by-case enquiries by email, and confirming unused credits stay valid. Assume consumed credits are non-refundable.
There is a free entry path. The homepage FAQ references testing with available free credits, and the workspace is visible without signing in, though generation requires an account.
Reasonable fits. Independent creators, marketers and small studios who want to try several current image models against one prepaid balance without opening accounts with Google, OpenAI, Black Forest Labs and ByteDance separately. The consolidation is the honest value proposition here, and for occasional users it is a real convenience. Educators and technical writers producing labelled diagrams may find the text-rendering focus useful. Anyone wanting a small, capped, one-time spend rather than a subscription is well served by non-expiring credit packs.
Poor fits. Businesses that require a named legal counterparty, a defined jurisdiction, a data processing agreement or a service level commitment cannot get any of those here. Anyone handling confidential or client-owned reference imagery should be deterred by the unresolved training contradiction and the permanent storage default. Teams needing guaranteed model versioning or uptime should note the Terms explicitly disclaim responsibility for interruptions or changes to the upstream service. Developers who need programmatic access should go to the Gemini API directly, since no public API is documented here. And anyone who simply wants Google's model with Google's own terms and support should use Google's official channels.
Open the site and use the workspace before creating an account, since the model selector, resolution options, format choices and batch controls are all visible unattended and tell you a lot.
Read the credit rate table on the pricing page and calculate cost for the model and resolution you will actually use. Do not budget from the headline per-image figure.
If you intend to use free credits, spend them on your hardest realistic prompt rather than an easy one. Text-heavy layouts, small type and multilingual captions are where these models differentiate, so test the thing you actually need.
Before uploading anything sensitive, email the operator to opt out of model training and keep the response. Given the contradiction between the two documents, an explicit written opt-out is the only defensible position.
Buy the smallest pack first. Non-expiring credits mean there is no penalty for starting small and topping up once the service has proven itself.
Download and archive everything you intend to keep. The site advertises permanent download links, but your archive should not depend on a platform with no identifiable operator remaining online indefinitely.
Verify the licensing position yourself if the work is commercial. The site grants generous rights, but those rights are only as strong as the upstream terms allow.
Ownership of outputs. Terms section 7.2 is unusually generous and clearly written. Users retain full ownership of generated images, may use them for any commercial or personal purpose, and may modify, distribute or sell them. All three paid tiers list commercial use rights, and "no watermark" appears in every tier's feature list.
The limits on that grant. A reseller can only pass along rights it holds from its supplier, and the Terms require compliance with "Google's content policies." The no-watermark claim should be read as meaning the platform does not stamp a visible logo on your output. It is not a statement about invisible provenance marking, which is applied at the model layer by the upstream provider and is not something a reseller controls. If provenance matters for your use case, verify independently rather than relying on a feature bullet.
Uploads and responsibility. Section 7.3 puts responsibility for the rights in anything you upload entirely on you.
Retention. Both documents state that uploaded images, prompts and generated images are stored permanently in your account unless you delete them. Permanent storage is the default, deletion is manual, and account deletion requires emailing the operator.
Data collected and shared. Account details, payment metadata, prompts, uploads, generated outputs, device information, IP address and approximate location, and usage analytics. Payment data is held by Stripe. Named categories of recipients include Stripe, cloud storage providers, Google Analytics, support tooling and email providers. The policy states personal information is not sold, and notes data may transfer to a new entity in a merger or acquisition.
Age and content rules. Users must be at least 18. Prohibited content includes sexual material, content depicting minors inappropriately, harassment, IP infringement, malware and fraud. The operator may remove content or suspend accounts at its discretion.
This platform sits in a crowded namespace where many similarly named domains are unrelated to each other and to Google. During research, searches for this product surfaced at least four distinct unaffiliated domains with near-identical names. Industry reporting has also documented that the Nano Banana trend attracted fake apps and lookalike sites harvesting personal and financial data. Being able to check a site yourself is more durable than trusting any single review.
Read the Terms before the marketing. Whether a supplier is named, and whether a company is named, tells you more in two minutes than any landing page.
Search the whole site for "not affiliated." Absence is not proof of bad faith, but its presence is a strong good-faith signal.
Look for a legal entity and a governing jurisdiction. A named company and a named forum are the baseline for recourse.
Cross-check the credit rates against the plan maths, exactly as done above. Documentation that does not reconcile with its own arithmetic tells you how carefully it is maintained.
Compare the Terms against the Privacy Policy on training and retention specifically. Contradictions cluster there.
Confirm what the official channel actually costs. Sometimes the wrapper is genuinely cheaper or more convenient; sometimes it is not.
The operator is unidentified, with no company name, registration, jurisdiction or address, and no governing-law forum. This is the most significant limitation on this page.
No affiliation disclaimer exists anywhere on the site, while Google's name is used prominently as a quality signal.
The Terms and the Privacy Policy contradict each other on model training, one requiring consent and the other assuming it.
The Terms understate credit costs relative to the live pricing table.
The Terms and the Pro model page name different upstream models.
Uploaded and generated content is retained permanently by default.
The platform is entirely dependent on an upstream supplier and explicitly disclaims responsibility for interruptions or changes to that supplier's service, which means access, pricing and quality can shift outside the operator's control.
No independent verification exists. There is no coverage of this specific domain on G2, Capterra, Trustpilot or Product Hunt, and no reporting from established technology media. Everything above that concerns the platform itself comes from the operator's own pages. That is not evidence of wrongdoing, but it does mean no third party has scrutinised the service, and it is the reason this document leans so heavily on internal consistency checks.
Marketing urgency devices, including a short countdown timer paired with a large percentage discount, are present.
Refunds on consumed credits should be assumed unavailable.
No public API or documented programmatic access was found, so this is a browser-based product.
Quality claims for the underlying model come from the supplier's documentation, not from independent benchmarking of this platform's implementation. A wrapper can constrain resolution, apply its own pre-processing, or route to a cheaper model tier, and none of that is externally auditable.
No. It is an independent third-party website that calls Google's image model through the Gemini API. Its own Terms of Service list Google's model under "Third-Party Services." No affiliation, sponsorship or endorsement by Google is indicated anywhere, and the site carries no disclaimer clarifying this. If you want Google's model directly, use the Gemini app, Google AI Studio or the Gemini API.
No, and this is worth being precise about. Nano Banana 2 is Google's own nickname for Gemini 3.1 Flash Image, used in Google's developer documentation and DeepMind materials. The model is real and is offered to third-party developers. What is unclear is not the model's identity but the website's relationship to Google, which the site never states.
Per the live pricing table, Nano Banana 2 costs 5 credits at 1K, 8 at 2K and 12 at 4K; Nano Banana Pro costs 6 credits at 1K and 2K and 12 at 4K. Other models range from 1 credit to 24. Packs were $19.90 for 1,000 credits, $49.90 for 3,300 and $99.90 for 7,500. Note that the Terms state roughly 4 credits per standard image, which does not match the pricing table or the advertised image capacities; the pricing table is the figure that reconciles.
The plan listings, the pricing FAQ and the Terms all state credits do not expire. This is one of the more customer-friendly aspects of the offering, since a one-time pack carries no monthly reset.
The Terms grant full ownership of generated images and explicitly permit commercial use, modification, distribution and sale, and all paid tiers list commercial use rights. Bear in mind that a reseller can only pass on rights it holds upstream, and the Terms also require compliance with Google's content policies. For high-stakes commercial work, verify the position independently.
The documentation contradicts itself. The Terms say content is not used beyond providing the service without explicit consent. The Privacy Policy lists training and enhancing AI models as a purpose, with opt-out by contacting support. Assume the opt-out reading, email the operator to opt out, and keep the reply. Avoid uploading confidential material until you have written confirmation.
Permanently, by default. Both documents state uploaded images, prompts and generated images are stored in your account indefinitely unless you delete them yourself. You can delete content from the dashboard, and full account deletion requires emailing the operator.
Assume not, for credits you have already spent. The homepage FAQ states refunds are not offered and recommends testing with free credits first. The Terms are slightly softer, saying refunds for consumed credits may not be available while allowing case-by-case email enquiries, and unused credits remain valid. Start with the smallest pack.
An email address is the only support channel published. There is no company name, no registered address, no phone number and no named legal jurisdiction in either policy document, and the governing-law clause does not identify a forum or arbitration body. Weigh this before making a large purchase or building a business process on the platform.
Nothing found in this research indicates the site is fraudulent, and there is a real, well-documented product with a named upstream supplier and Stripe handling payments, which means card details are processed by a major provider rather than the site itself. The concerns are transparency and recourse rather than evidence of misconduct: no identifiable operator, no jurisdiction, no third-party reviews of this specific domain, and self-contradictory policies. A small one-time purchase carries limited exposure. Committing business-critical work or confidential material is a different proposition, and for that the absence of a named counterparty should give you pause.
Note: All platform-specific details above were taken from the operator's own public pages at the time of research and reflect what those pages stated then. Pricing, credit rates, model availability and policy wording can change without notice. Verify current terms on the site before purchasing.
