Most AI image apps work the same way. You type a prompt, maybe pick a photo, and both travel to a company’s servers. A model runs there, and an image comes back. It is fast and convenient, and it means your prompt and your photo have left your phone before anything else happens.
Onhand, our image studio for iPhone and iPad, takes the other path. It downloads the image model to the device and runs generation there. Your prompts and the photos you edit are not uploaded to a server to make an image. Onhand is not released yet; it is coming soon, and onhand.arkaledge.com (opens in a new tab) has a launch list.
The problem with the round trip
A privacy policy can promise a lot about what happens to an upload: how long it is kept, who can see it, whether it trains a model. Those promises are worth reading. But they all describe what happens after the data has already been received. The prompt you typed at midnight and the family photo you wanted to retouch are on someone else’s computer either way.
For a lot of what people make with image tools, that is fine. For some of it, it is not. Personal photos, half-formed ideas, and work that isn’t public yet are exactly the things many of us would rather keep on our own device. On-device generation removes the question instead of asking you to trust the answer.
What on-device means in Onhand
Onhand is built with Swift and Apple’s MLX and Core ML frameworks, and the image model runs on the device’s own GPU. Once a model pack has downloaded, generation works in airplane mode. Today the app’s workflows are:
- Text to image: start from a description.
- Image to image: start from a photo you choose.
- Inpainting: paint over part of an image and replace it.
- Resize 2×: enlarge an image.
- Background removal: keep the subject and drop the rest.
What it costs
Running a diffusion model on a phone is not free, and we would rather say so plainly.
The models are large, so each one is a real download the first time you use it. Memory is the hard limit: iPhones do not swap to storage the way a laptop can, so a device’s memory decides which models and image sizes it can run reliably. Onhand targets compatible iPhones and iPads running iOS or iPadOS 26, and a device with more memory can run more of the model library.
Speed varies too. How long an image takes depends on the model, the quality setting, the device, whether the model is already loaded, and how warm the phone is. Early in development we quoted timings measured on our own devices. We have since taken them out of our general product guidance, because a number from one device and one model is not a promise for everyone. The fairest test is your own device.
Generation also works the chip hard, so a long session warms the phone and uses battery in a way a cloud app would not. That is the trade: the work happens on hardware you own, so you are the one supplying the power.
What still uses the network
“On-device” describes the image generation, not every byte the app ever sends, and it would be misleading to blur that. Onhand uses the network to download model packs and catalogue updates, to handle purchases, and for limited app analytics. The Onhand privacy policy (opens in a new tab) lists each service and what it receives.
Images you export or share follow the rules of wherever you send them. If iCloud Photos is turned on, an image saved to Photos syncs like any other picture.
Why build it this way
Apple’s recent chips are capable enough to run real image models locally, and frameworks like MLX and Core ML make that practical for a small team. When the device can do the work, sending your prompts and photos away to have it done somewhere else stops being a necessity and becomes a choice. We think more creative tools should make the private option the default.
Onhand is still in development, and the details here describe the app as it stands today. If you want to know when it arrives on the App Store, join the launch list at onhand.arkaledge.com (opens in a new tab), where the Onhand blog (opens in a new tab) also goes deeper into on-device image generation.