What is AI photography?

The term covers two different things. One is AI inside the camera and the edit — autofocus, noise reduction, upscaling, background removal — where a real photo is taken and a model improves it. The other is AI-generated photography, where an image that looks like a photograph is produced without a photograph being taken. The second is the one that raises questions about disclosure and truth.

Is an AI photo the same as a filter?

No. A filter alters an existing photograph; the pixels underneath were captured by a camera. A generated photo is built from a model of your face and a described scene, with the lighting and shadows constructed as part of the image rather than laid over it. That is why generated results can place you somewhere you have never been, and a filter cannot.

How many photos do I need to upload?

Around ten, and variety matters far more than quality. Photos taken on different days, in different light, at slightly different angles teach the model what your face is. Five excellent photos from one afternoon teach it one lighting setup.

Can people tell a photo is AI-generated?

Sometimes. The usual tells are over-smoothed skin, lighting that could not have happened, hands, and backgrounds that do not quite add up. Restrained, naturally lit results are the hardest to identify, because they look like ordinary photographs. Dramatic ones are easier to spot precisely because they are dramatic.

Do I have to say a photo is AI-generated?

Where a platform has a rule, follow the rule. Beyond that, a workable principle is to disclose when the photograph makes a claim about the world rather than about how you look. A styled portrait sits in a context everyone reads as styled. A photo of you somewhere you have never been asserts that you were there, and that is the assertion people object to.

2026-08-08

The ultimate guide to AI photography

What AI photography actually means, how a generated photo of a specific person is made, what the technology is bad at, and where you should not use a generated photo at all.

"AI photography" is used for two genuinely different things, and most of the confusion around it comes from nobody separating them. This guide separates them, explains how the second one actually works, and is honest about what it cannot do — including the situations where you should not use a generated photograph at all.

The two things people mean by AI photography

The first is AI inside the camera and the edit. Your phone already does this. Face and eye detection driving autofocus, noise reduction on a dark frame, automatic upscaling, sky replacement, background removal, one-tap object erasure. A real photograph is taken and a model improves it. This has been mainstream for a decade and almost nobody calls it AI photography any more — they call it the camera working.

The second is AI-generated photography: an image that looks like a photograph of a real person, in a real place, which was never taken. No shutter opened. That is the sense this guide is about, and it is the sense that raises every interesting question.

The distinction matters practically. Editing a photo of yourself changes a record of something that happened. Generating one creates a record of something that did not. Those are different acts with different rules attached, even when the results look identical.

How a generated photo of a specific person works

Generic image models can produce a convincing stranger. Producing you takes an extra step.

A face model is trained on your own photos. Around ten ordinary selfies, taken once. The model is learning the structure of your face across different light and angles — which is why variety in those photos matters far more than their quality.
You choose a template. A template is a finished photographic setup: the lighting, the setting, the framing, the wardrobe and the composition are already decided. It is closer to choosing a photographer's shot than to writing a prompt.
The two are combined. Your face model is rendered into that setup at full resolution. In Artisse this takes roughly three minutes per photo.

The reason serious tools use templates rather than free-text prompts is repeatability. A prompt is a description, and descriptions produce a distribution of results — you write "professional headshot, soft light" and receive a lottery. A template is a fixed set of decisions, so the variable is your face and nothing else. That is what makes a set of photos look like it came from one session rather than from twenty different ones.

What it is genuinely good at

Volume and consistency. Thirty usable photos in an evening, all recognisably the same person, is not something a shoot produces.
Situations you cannot easily reach. Weather you would have to wait for, a place you are not in, a wardrobe you do not own.
Iteration without cost. If a result is wrong you generate another. There is no rebooking, and nobody is standing there while you decide.
The photo nobody has. Most people own plenty of pictures from holidays and none from an ordinary week, because nobody photographs an ordinary week. A current, well-lit, unremarkable photo of yourself is the hardest one to actually have.

What it is bad at — the honest list

This is the part usually left out, and it is the part worth reading.

Hands and text. Both have improved a great deal and neither is solved. Check fingers, and check any writing in the frame.
Groups. One trained face is reliable. Several people in one frame, all of whom must be right, is materially harder.
Specific objects. A generated photo will not faithfully reproduce your actual jacket, your actual dog or a particular product. It produces something of that kind, not that thing.
Extreme angles and expressions. Likeness is strongest near the range covered by your uploads and degrades outside it.
Anything that has to be true. A generated image is not evidence of anything, and no amount of realism changes that.

How it compares to a studio shoot

A shoot costs more and takes longer, and for a single flagship portrait it can still be the better answer. What a photographer supplies that a model does not is judgement in the room: reading how you are standing, noticing that you relax after twenty minutes, choosing to try something that was not planned, and taking responsibility for the result. That is a real service, not a legacy one.

What generation supplies instead is range and revision. It suits the case where you need many photos, in several contexts, updated more than once a year — and it is available at eleven at night when you realise the profile picture you have is four years old.

Where you should not use a generated photo

Some of this is law, some is platform policy, and some is simply what a reasonable person would expect.

Identity documents. Passports, driving licences, visas, right-to-work checks. Never.
Anything presented as evidence. Insurance claims, legal filings, journalism, "proof" of any kind.
Anywhere a platform forbids it. Rules differ and they change; the platform's own policy is the authority, not a general rule of thumb.
Claims about the physical world. A photograph of you in a city you have never visited asserts that you were there. That is the assertion people object to, not the pixels.
Other people's faces. Generating a recognisable person who did not agree to it is not a grey area.

Should you say a photo is AI?

A workable rule: disclose when the photograph makes a claim about the world, not when it makes a claim about how you look.

A styled portrait sits in a context everyone already reads as styled — nobody assumes a headshot was captured accidentally, and retouching has been normal for a century. A photo of you on a beach in another country is a different object: it says something happened. Between those, the closer you are to asserting a fact, the more disclosure stops being courtesy and starts being honesty. Where a platform has a rule, the rule wins regardless.

Getting a better result

Upload variety, not quality. Ten photos in different light, on different days, at slightly different angles. Poor phone snaps in varied conditions beat five good photos from one afternoon.
No sunglasses, hats over the eyes, or heavy filters in the uploads. You are teaching a face; do not hide it.
Generate more than you need. The one that works is usually not the one that looked best in the grid, because scale changes everything.
Match the template to the job. A dramatic, high-contrast photo is a worse profile picture than a plain one, however good it looks full-size.

Where this is heading

Two things are moving quickly: control and speed. Control means being able to specify a result precisely rather than sampling until one is right, which is why the useful products increasingly look like a catalogue of decisions rather than a text box. Speed keeps falling, and once generation is instant the interaction stops resembling a render queue and starts resembling a camera.

What is not moving is the underlying constraint. A generated photograph is a plausible image, not a record. Everything sensible built on this technology respects that line, and everything that gets into trouble crosses it.