On-device processing: what it is and why it matters
Quick facts
- Definition
- Running a calculation or an AI model inside the phone itself, without sending the data to an external server.
- Contrast
- With server processing, the photo leaves the device, is computed on another machine and the result comes back.
- Advantages
- Less data in transit, immediate response and operation without relying on a connection for that step.
- Limit
- The device has less power; heavy tasks, such as generating an image, usually stay on the server.
On-device processing is the execution of a calculation inside the user's own phone or computer. The data goes in, the model runs right there and the result stays there, with no trip to a third-party server. It is the opposite of cloud processing, in which the image is sent, processed elsewhere and sent back.
What is the difference in practice?
In an app that deals with faces, the difference shows up in three points: privacy, speed and network dependence.
| Aspect | On the device | On the server |
|---|---|---|
| Where the photo goes | Stays on the phone | Leaves over the internet |
| Speed | Almost immediate | Depends on the connection and the queue |
| Power | Limited to the phone's hardware | Much greater |
| Typical use | Detecting the face, marking landmarks | Generating images with heavy AI |
The computer vision that locates eyes, nose, mouth and contour usually runs on the device. Generative AI, which creates a new image, usually requires a server, because the model is too large for a phone.
Why does this matter for facial data?
The face is facial biometric data and, when linked to an identifiable person, falls into the sensitive personal data category of the LGPD. The less this data circulates, the smaller the risk surface: there is no transit to intercept and no server copy to leak.
This does not zero out the professional's obligation. If the photo is sent to any step outside the device, the patient needs to know, with a clear purpose, and the acceptance must come before sending. Local processing is a good data-minimization practice, not a substitute for consent.
How does Harmy use this concept?
Harmy divides the work explicitly. Face detection and the calculation of the marking points happen on the device itself. The "after" image, however, is generated on a server, by a third-party AI, and that is why the professional gives an acceptance in the app before the photo is sent.
What to ask when choosing a tool?
These answers help build the informed consent form and respond to the patient who asks "where does my photo go".
Also called: on-device, local processing, edge computing (in the context of apps).
Frequently asked questions
Does processing on the device remove the need to comply with the LGPD?
No. It reduces the exposure of the data, but if the photo is later sent to another service, or stored, the LGPD (Brazil's data protection law) still applies, with legal basis, purpose and consent.