Policy document
Content moderation & risk profile
How generated content is filtered on BananaBanana, what our own rules prohibit, and how we enforce them. Written for payment providers, partners and anyone reviewing the service.
Last updated: 16 August 2026
A pay-per-use generation tool
BananaBanana is an API and web app for generating images, video and speech. Customers top up a prepaid balance and are charged per generation, from $0.03 per image. There is no subscription, no user-to-user messaging, no social feed, no public gallery of user content and no marketplace.
The only content that exists on the platform is what a customer generates for themselves and downloads. Generated files are deleted automatically after 30 days.
We do not operate a model — we resell Google's
Every generation runs on Google's hosted models through Vertex AI:
| Capability | Model | Provider |
|---|---|---|
| Images | Nano Banana 2 / 2 Lite / Pro (Gemini image models) | |
| Video | Veo 3.1, Veo 3.1 Fast, Gemini Omni Flash | |
| Speech | Gemini 3.1 Flash TTS |
We train nothing, host no weights, and run no open-weight or uncensored model. This is the single most important fact about our risk profile: the content boundary is enforced by Google inside their own infrastructure, before any file reaches us. A refused request returns an error instead of a file — we never receive the rejected content, and could not disable that behaviour if we wanted to.
Our platform therefore inherits Google's prohibited-content boundary in full, including the categories payment providers care about most: sexual content, any sexualised depiction of minors, and real identifiable public figures are refused by the upstream model on every attempt.
Two stages, both enforced by Google
Stage 1 — request
Checked before anything is rendered
Configurable through Vertex AI's documented safetySettings and personGeneration parameters. This stage produces most false positives on legitimate work — swimwear and sportswear, medical and anatomical subjects, mild fictional violence, realistic portraits.
Stage 2 — output
Scores the finished file
Not configurable by any parameter, by us or by anyone. Images fail with IMAGE_SAFETY, video with rai_media_filtered_reasons.
A Google setting, not a bypass
Our API exposes an optional relaxed_filter parameter. It sets safetySettings to the permissive threshold and personGeneration to ALLOW_ADULT — two documented, supported Vertex AI parameters, offered by Google to their own API customers. It is a configuration choice within Google's product, not a circumvention technique.
What it does not do
- It does not disable stage 2 — the output-side classifier scores every generated file regardless.
- It does not unlock any category Google prohibits. Sexual content, minors and real public figures remain refused on every attempt, with or without the flag.
- For video it is close to a no-op: Veo exposes no safety settings at all, so the flag only pins a default that Veo 3.1 already applies.
Its practical purpose is reducing false refusals on legitimate commercial work — the fitness brand whose sportswear shoot gets refused, the medical illustration, the film-style scene with a plausible-looking human.
Acceptable use and enforcement
Our published Terms of Service (section 5, “Acceptable Use”) prohibit the following, and we terminate accounts without refund for any of it:
- Illegal, harmful or abusive content
- Content infringing third-party intellectual property
- Deepfakes or non-consensual intimate imagery of real people
- Any child sexual abuse material
- Content promoting violence, terrorism or hate speech
- Attempts to bypass content safety filters or model restrictions
NSFW and adult content are not a product category for us, are not marketed, and are refused upstream in any case.
Prepaid balance only
No stored cards, no recurring billing, no subscription to dispute. Customers top up and spend down a balance.
Full audit trail
Every generation is logged with account, model, prompt, cost and outcome; every API call is logged separately. Refusals are recorded with the upstream reason code.
Automatic refunds
A refused or failed generation costs the customer nothing, which removes any incentive to retry abusively.
Rate limits and spend caps
Per-account generation limits, plus optional per-key daily spend caps on API access.
Verified accounts
A verified email is required to sign in, so every generation is attributable to a confirmed account.
Enforcement
Administrators can block an account on any violation report. Abuse reports go to [email protected].
At a glance
| Adult / NSFW content | Prohibited by our Terms; refused upstream by Google |
| CSAM | Refused upstream in every configuration; zero tolerance, account termination |
| Deepfakes of real people | Prohibited by our Terms; real public figures refused upstream |
| Chargeback exposure | Very low — prepaid top-ups, no subscriptions, no stored cards |
| Physical goods / shipping | None — digital only |
| User-to-user transactions | None — no marketplace; the only payout is an affiliate commission on our own revenue |
| Regulated categories | None — no gambling, finance or pharma |
| Content retention | Generated files are deleted automatically after 30 days |
Questions and abuse reports
Compliance and abuse: [email protected]