On-Device AI Image Generation, Explained
Most AI image apps send your words or photos to a server and send a picture back. An on-device app does the whole job on your own iPhone, iPad or Mac. Here is what that means in practice.
How it works
The app downloads an image model once, a file of a few gigabytes. After that the model runs on the chip in your device, on the Neural Engine and the graphics processor in Apple silicon, through Apple's Core ML and the MLX framework. Your prompt goes in, the picture comes out, and nothing travels over the network.
What you gain
- Privacy. Your prompts and photos stay on the device. There is no server to store, read or leak them.
- Offline use. Once a model is downloaded, it works in flight mode, on a train or in a cabin with no signal.
- No queue. You never wait behind other people's jobs, and the service cannot be busy or down.
The trade-offs
- Storage. Each model takes a few gigabytes. You can remove one and download it again later.
- Memory decides what runs. Bigger models need more memory, so the app should show only what your device can run well. In Visualmee, iPhones and iPads with 6 GB of memory run a small, fast model; devices with 8 GB or more also get the larger models and editing with words.
- Speed and heat. On a phone a picture takes from about half a minute to a couple of minutes; a Mac is usually faster. Long sessions warm the device, and a cooler device is faster.
How to tell if an app is really on-device
Download the model, turn on flight mode, and create a picture. If it works, the work happened on your device. Then read the privacy label on the App Store page: an on-device app has little or nothing to declare about your content.
Visualmee works this way on iPhone, iPad and Mac. See what it can make.