What happens to your face when you upload it to a headshot generator

An AI headshot service needs your face. Usually several photos of it, sometimes a dozen, because the model is being fitted to you specifically. That is a different kind of upload from sending a document, and it is worth being clear about why before deciding whether you mind.

This is general information rather than legal advice, and none of it is an accusation about any particular service. It is the set of questions that make the answer knowable.

A photo of your face is personal data, and sometimes more

Under the GDPR, a photograph of an identifiable person is personal data. That much is uncontroversial.

The sharper category is biometric data, which the regulation treats as a special category with stricter conditions. The distinction turns on processing rather than on the image: a photograph becomes biometric data when it is put through specific technical processing that allows or confirms unique identification of a person. A face template computed for recognition falls in that bucket. A photo sitting in a folder generally does not.

So “is this biometric” is not a question about your photo. It is a question about what was computed from it, and that is usually invisible from the outside. Which is why it is worth asking directly.

Four clocks start when you hit upload

It helps to separate these, because a single reassuring sentence usually only covers one.

Transit. The file leaves your machine and crosses networks to a server. HTTPS handles the obvious risk here, and this is the least interesting of the four.

Retention. The copy now exists somewhere. “We delete it after processing” is a real commitment about one stretch of time, beginning after the file has arrived. On its own it does not tell you whether backup snapshots exist, how long access logs live, or whether the filename and your account history outlive the image. None of that is sinister, it is just outside the sentence.

Training use. Whether your face contributes to a model that is not yours. This is the clock people most often assume is stopped and most often is not, because it is typically governed by the terms of service rather than the privacy page.

Derivation. Whether a face template, embedding or identity model was computed and whether that is deleted when the photo is. A model fine-tuned on your face is a derived artefact. Deleting the source photos does not automatically delete it, and a promise about photos does not necessarily cover it.

The questions worth sending

If you want a usable answer, ask for specifics and ask in writing.

How many days are the uploaded photos kept, as a number rather than “promptly”? Are backups included in that, and if so what is their separate retention period? Is any model, embedding or template derived from my photos, and is it deleted on the same schedule? Are my images used to train or evaluate any model beyond producing my own results, and can I opt out without losing the service? Which sub-processors receive the images, and in which countries? How do I exercise deletion, and what is the response time?

A service that has thought about this will answer quickly and in numbers. That in itself is informative.

The version where the question does not arise

There is a second route, and it changes the shape of the problem rather than the answer. If the image never becomes the body of a network request, there is no retention period to ask about, no transit to secure, no sub-processor list to read and nothing to chase in six months.

That is how this site works, and the mechanism is ordinary browser machinery. You choose a file, the browser hands the page a handle to bytes already on your disk, the page decodes it into a canvas and runs the models there. Two models, about 20 MB together, downloaded once and cached. The direction of travel is inward: something large arrives, nothing comparable leaves.

There is a second property worth separating from privacy, because it is a different claim. Nothing here is generated. The models only locate things, which pixels are background, hair, skin, clothes or accessories, and where 478 points on your face sit. Everything after that is ordinary image processing inside those masks: matting, relighting, masked retouching, cropping, resampling. No face is redrawn, so the person in the result is the person in the photo. A generative service, by construction, cannot make that claim.

The face mesh is computed, used to place masks and choose a crop, and discarded when you close the tab. It is never transmitted and never compared against anything.

Checking it yourself takes two minutes

Do not take this on trust either. Chrome and Edge both ship DevTools, and the Network panel logs what a page requests.

Open DevTools with Control+Shift+J, or Command+Option+J on a Mac, and click Network. DevTools only records while it is open, so reload the page once with the panel showing, and leave the filter on All so nothing is hidden. Then generate a headshot and watch.

You will see the model weights arrive on the first run, which is the large download. You will not see your photo go anywhere, because there is no request carrying it.

The stronger version of the test takes one more step. Generate once so the models are cached, then turn off your Wi-Fi and generate again. It still works, because at that point nothing on the page needs the network. A tool that uploads cannot pass that test, and it is difficult to fake.

If that is the trade you want, drop a photo in and try it. Our guide to shooting a headshot with the phone you already have covers getting a source photo worth processing, since the two things no tool can fix afterwards are focus and light.