What Makes an AI Image Generator Private?

A private AI image generator keeps inference on your device so prompts and photos are not sent to a generation server. You can check with airplane mode after a model download. Marketing “private” without on-device generation usually still uploads for processing.

Search for a private AI image generator and you will find dozens of apps using that word. In most of them, your prompt and your photos travel to a server and are processed on hardware you cannot inspect. “Private” as a marketing word and private as a fact about how the app is built are two different things, and you can check which one you are getting.

This article is a checklist for judging whether an AI image generator keeps your data on your device. None of the tests ask you to trust a company’s word, and most need nothing more technical than airplane mode.

Privacy is a question about architecture

The first question about any image generator is where the model runs. If it runs on a server, your prompt and any photo you edit have to be sent there, and no policy changes that. Everything else in the privacy story is about data the company has already received: how long it keeps that data (retention), whether it trains on it, how it is encrypted, and which staff can reach it.

Running the model on your device removes the need to upload prompts and photos for inference. It does not, by itself, prove that an app never sends other data or syncs content later. Check both where generation happens and what the privacy policy says about analytics, backups, sharing, and retention.

The checklist

• Does it work with networking disabled? Download a model, turn on airplane mode, and explicitly turn Wi-Fi off as well. Generate a new image. Success demonstrates an offline generation path. It does not prove what the app sends after networking is restored.
• Did you download a model? Local inference needs local weights. These may be bundled with the app, downloaded separately, or provided by the operating system. A substantial model download supports a local-processing claim, but download size alone cannot settle it.
• Is there an account? Local generation needs no sign-in. There is no usage quota to meter on a server, and no per-user history to sync. A mandatory account is not proof of cloud processing. It does mean your activity can be tied to an identity. Ask why the app needs one.
• Does performance depend on the device? Chip, memory, model choice, and temperature can affect local inference. Apps may choose different models by device, and timing alone does not reveal where a computation runs.
• Where do edited photos go? Test an edit with networking disabled, then review the policy for later uploads and export behavior. Completing an offline edit shows that the edit itself needed no server.

What a network inspector shows

A network trace can help check which services an app contacts and when. Inspecting contents is harder because requests are usually encrypted, so timing and transfer size are clues that need context.

Local apps may send analytics after a button tap; model downloads and Photos sync can also overlap a session. Compare the requests with the documented data flows before drawing a conclusion.

A large upload after selecting a photo deserves investigation. Establish which app or system service sent it and what action triggered it. A network trace without request contents does not prove that the AI model ran remotely.

“We don’t store your data” is not “we never receive it”

The most common privacy claim in this category sounds like “your images are not stored” or “deleted after processing.” Read literally, it concedes the important part. The data was received. Once your photo has crossed the network, its fate depends on settings you cannot see. Request logs, crash dumps, caches, backups, third-party GPU providers, and every future change to the retention policy all affect it. An app that never sends the photo avoids all of that.

Review privacy claims again after app updates. A local inference path can coexist with network services, and future versions can change those services. Keep copies and sharing settings in scope when deciding where an image may go.

Look for an explicit statement about where inference runs and which other data is collected. Encryption in transit is relevant whenever a service receives data, but says nothing by itself about whether image generation is local. An app can use encrypted analytics and still generate offline.

App privacy labels help, but read them as claims

App Store privacy labels summarize declared collection. Usage data or purchase history can be collected by an app whose image generation remains local. Read the data types and purposes rather than treating any collection as proof of a photo upload.

Developers write these labels themselves. They summarize a policy, not an audit of how the app is built. So a label cannot tell “we process your photo on a server but do not retain it” apart from “your photo never leaves the device.” Both can truthfully display a clean label. Treat the label as a first filter, then apply the airplane-mode and traffic tests above, which measure what the app does rather than what its developer declares.

How Onhand answers the checklist

Onhand runs its installed image models on compatible iPhones and iPads, through MLX or Core ML depending on the package. It has no account or advertising network. Prompts, source photos, and generated images are not uploaded for inference. Catalog delivery, Apple-hosted model downloads, RevenueCat purchases, and limited PostHog analytics use the network. Saving or sharing an image follows the destination’s settings, including iCloud Photos if enabled. See the privacy policy.

Onhand is coming soon to iPhone and iPad and is not on the App Store yet. Join the launch list on this site to hear first. For editing photos without upload, see private AI photo editor. For a 60-second offline proof test, see generate AI images offline on iPhone. Full data practices: privacy.

Apple’s Image Playground takes a different route: in iOS 27 it generates on Private Cloud Compute rather than on the device. Onhand vs Image Playground compares where each one generates and what that means offline.

FAQ

What makes an AI image generator private?

Inference runs on your device, so prompts and photos are not sent to a generation server. Retention and “we delete after processing” policies still assume the server received your data. Architecture first; policy second.

Does “we don’t store your data” mean private?

No. That phrasing usually means the company received your image, ran it on their hardware, and claims not to keep a copy. Receipt is the privacy event. An on-device generator never sends the photo for inference in the first place.

Can I check without trusting the company?

Yes. Download a model pack, turn on airplane mode (and turn Wi‑Fi off), then generate. If the image appears with the radios off, that run did not need a generation server. Restore networking afterward and read the privacy policy for analytics and purchases.

Does Onhand upload photos?

No for inference. Prompts, source photos, and generated images stay on the device for generation and editing. Catalog downloads, purchases, and limited analytics use the network separately; details are on the privacy page.

Is Onhand on the App Store?

Not yet. It is coming soon to iPhone and iPad. Join the launch list on this site to hear first.