An AI generator sells a beach family portrait in about four minutes, but it invents hands, jewellery and backgrounds instead of recording your real family. A real session is roughly ninety minutes at golden hour, delivered as a private 4K gallery in one to three business days. Consumer research puts demand for clear AI disclosure near 90 percent.
If you search for family portraits in Cancún today, a portion of what comes back is not photography. AI portrait generators now bid on the same words that photographers do, and they promise a beach family portrait in four minutes for the price of a coffee. That is a real offer and it deserves a real answer rather than a defensive one. So here it is, from a studio that has delivered more than 500 sessions since 2023: what those tools actually produce, how to recognise generated images inside a portfolio, what our own policy is on where a machine may and may not touch your photographs, and the one case where we think a generator is genuinely the right choice.
What an AI portrait generator actually produces
Most consumer portrait generators work the same way. You upload ten to thirty photographs of a person, the service fine-tunes a small adapter on top of a large image model so that it learns a rough approximation of that face, and then it generates new images from a text prompt. The output is not a photograph that was retouched. It is a new image, synthesised pixel by pixel, of a person who resembles you standing somewhere that does not exist.
For a single adult that works surprisingly well, and anyone who says otherwise has not looked recently. Headshots for a profile picture are a genuinely solved problem. Families are where it falls apart, and the reason is structural rather than temporary. Each additional person multiplies the number of identities the model has to hold consistent in one frame, and physical interaction between them is exactly where these models are weakest. Hands on shoulders, a child carried on a hip, four people at four different heights with correct relative scale and shared lighting. Those are the frames that make a family portrait a family portrait, and they are the frames generators get wrong.
There is a quieter problem too. A generator can only produce a version of what it has seen a great deal of, so its Cancún is a generic tropical composite: nowhere beach, nowhere palm, nowhere sunset. It cannot produce the resort where you actually stayed.
The tells: hands, jewellery, backgrounds, repeated faces
You do not need software to spot generated images. You need to know where to look, and to zoom in, because these images are built to survive a thumbnail and rarely survive full size.
Hands and contact points. Count fingers, then look at where two people touch. Generated images routinely produce a hand resting on a shoulder that has no visible arm connecting it, or fingers that merge into fabric. Contact between bodies is the single most reliable tell.
Jewellery, watches and glasses. Rings that change design between one hand and the other, earrings that do not match, a watch face with no coherent dial, glasses whose frames dissolve where they meet the ear. Small repeated objects need consistency across a frame and models do not track them.
Backgrounds and text. Zoom into the background people. In a generated beach scene the figures behind the family will be smeared, faceless or physically impossible. Any lettering, a hotel sign, a boat name, a t-shirt, will be almost-language rather than language.
Light that does not agree. Check where the shadows fall on each face, then check the sun's position in the sky. In generated images faces are often lit from the front while the sky behind indicates the sun is behind them. Real backlit golden hour has a specific rim on the hair and shoulders and a darker face, and it is very hard to fake convincingly.
Repetition across a portfolio. This is the tell that matters most when you are evaluating a supposed photographer rather than a single image. Scroll a whole gallery. If the same three faces recur in different clothes, if every sunset is the same colour, if every beach has the same three palms in the same arrangement, you are looking at outputs from one model rather than the work of one photographer.
What a real session captures that no model can invent
Set the technical argument aside for a moment, because the honest case for real photography is not that generators look bad. It is that they are answering a different question.
A generated portrait shows you what your family might look like arranged pleasantly on a beach. A photograph shows you what your family was actually like on a specific evening in a specific year. Those are not the same product. The value of the second one increases every year, and the value of the first one does not. Ten years from now nobody will care that a synthetic image had a nicer sky. They will care about how tall the children were, which cousin refused to look at the camera, and what your father's face did when the grandchildren piled onto him.
There are things a model has no access to and never will. It does not know that the youngest one needs twenty minutes of being ignored before she relaxes, or the specific way your family stands when nobody has told them to stand. That is data that only exists if somebody was there with a camera.
And practically: a session is ninety minutes in which your family is together, outdoors, at the best light of the day, with nobody holding a phone. No generator produces that.
Our AI policy: what a machine touches and what it never will
We think every studio should publish this, so here is ours in plain language.
Every image we deliver is a photograph. Real people, really present, in a real place, recorded by a camera on the date on your invoice. Nothing we deliver is generated.
What software may do. We use AI-assisted tools the way we use any other tool in a raw editor: subject and sky masking so an adjustment lands where we want it, noise reduction, lens corrections, sensor dust and blemish removal, and removing small distractions such as a stray plastic cup or a piece of litter on the sand. These are the modern versions of things darkroom printers did with their hands.
What we do not do. We do not generate faces, bodies, skies or backgrounds. We do not extend the edges of a frame with invented content. We do not composite a person into a photograph they were not in. We do not swap a head from one frame into another, and we do not reshape anybody's body. If a client specifically asks for a composite, for example to include a relative who could not be present, we will discuss it, we will tell you plainly that the result is a composite, and it will be delivered as a clearly separate file rather than mixed into the gallery.
Why this matters commercially. Consumer patience for undisclosed AI is dropping, not rising. Fractl's research found distrust of visibly AI-generated brand content roughly doubled year over year to about 40 percent of consumers, and that only around 7 percent trust a brand more when it uses visible AI content. Getty's global consumer research puts the share of people who want AI images clearly disclosed at close to 90 percent. A studio that quietly generates work is borrowing against its own reputation.
Ask, in writing: is every image you deliver a photograph of the people who were present, and do you use generative fill or generative expand on client work? A studio that does real work answers that in one sentence. Hesitation is your answer.
Content credentials and where camera provenance is heading
The industry is building an actual technical answer to this, and it is worth understanding because it will affect how you verify images within a few years.
The standard is C2PA, usually branded as Content Credentials, and it works by attaching a cryptographically signed record to an image describing where it came from and what was done to it. Camera manufacturers have started shipping it in hardware and firmware, beginning with Leica's M11-P and followed by professional bodies from the major Japanese manufacturers, and the main editing applications can now preserve and extend that record through the edit rather than stripping it.
It is not finished. Credentials are stripped whenever an image passes through a platform that does not preserve metadata, which today is most social networks, and absence of a credential does not prove an image is fake. But the direction is clear: within a few years, the meaningful question about a photograph will not be whether it looks real, it will be whether it carries a verifiable chain back to a camera. Studios that have always shot real work will simply turn the feature on. That is the whole reason we are comfortable telling you our policy now, in public, before anyone requires us to.
How to audit a portfolio before booking
Five checks, none of which take long, and any working photographer will pass all five without complaint.
Ask for one complete gallery. Not a curated portfolio, an entire delivered session from start to finish, with the client's permission. Generated portfolios do not have complete sessions, because there was never a session. Real ones have four hundred frames with a visible arc: arriving, warming up, the good stretch, the tired children at the end.
Verify the locations. Pick three images and ask exactly where they were taken. Then look at the place on a map or in the resort's own photographs. Real work is set in identifiable places with recognisable architecture, and a photographer who works in the region can tell you which beach access it was.
Look for named reviews. Reviews attached to real names on a platform that verifies them are much harder to fabricate than a website testimonial. Ours are on our Google Business Profile, 5.0 across 47 reviews, and the names on them are real people we photographed.
Get on a video call. Two minutes of conversation about your resort, its beach orientation and its vendor policy tells you whether somebody has stood on that sand.
Read the contract. A real studio's contract names a photographer, a date, a location and a delivery window. Ours promises a private 4K gallery within one to three business days and names the person who will be there.
When AI is genuinely useful in our workflow
We would be doing exactly what we criticise if we pretended none of this is useful. It is, and here is where.
Culling. Sorting four hundred frames by whether eyes are open and faces are sharp is mechanical work that software does faster than a person at eleven at night. It proposes, we decide. Noise reduction on frames shot in the last minutes of light, where the alternative used to be a frame we could not deliver at all. Masking, so a single adjustment lands on skin without touching the sky. Translation support in the bilingual side of the business, and drafting the first version of a timeline document that we then rewrite.
And one honest recommendation against ourselves: if what you want is a fun, obviously stylised image of yourselves as astronauts or in a Renaissance painting, a generator is the right tool and hiring us would be a waste of your money. Use it, enjoy it, and label it. What we are asking for is that the distinction stays visible, because the moment nobody can tell the difference, every real photograph loses the thing that made it worth keeping.
If you want the real version, the studio is based in Cancún and covers the Riviera Maya coast. Our approach to family photography is on the service page, more about how we work is on the about page, and the rest of the guides live in the Journal.
Questions people ask about AI and photography
For a single adult headshot, often yes. For a family, much less reliably. Each additional person multiplies the identities the model must keep consistent in one frame, and physical contact between people, hands on shoulders, a child on a hip, correct relative heights and shared lighting, is exactly where these models fail. Those interactions are what make a family portrait work.
Zoom in rather than judging thumbnails. Look at hands and where two people touch, at jewellery and glasses for inconsistency, at background figures and any lettering, and at whether shadow direction agrees with the sun in the sky. Then scroll the whole gallery: repeated faces, identical sunsets and the same palm arrangement across many images point to one model rather than one photographer.
Every image we deliver is a photograph of people who were actually present. We use AI-assisted tools for culling, noise reduction, lens correction, subject masking and removing small distractions such as litter on the sand. We do not generate faces, bodies, skies or backgrounds, we do not extend frames with invented content, and we do not composite people into photographs they were not in.
Content Credentials is the consumer name for C2PA, a standard that attaches a cryptographically signed record to an image describing its origin and the edits applied to it. Camera makers have begun signing files in hardware, starting with the Leica M11-P, and major editing software can carry the record through an edit. Platforms still strip metadata, so a missing credential does not prove an image is synthetic.
Yes, considerably, and that is a real trade rather than a trick. What you are buying is different. A generator produces a plausible image of people who resemble you in a place that does not exist. A session produces a record of your family on a specific evening in a specific place, which is the thing that gains value over the years rather than losing it.
When the goal is an obviously stylised image rather than a record. If you want your family as astronauts or in the style of an oil painting, a generator does that well and hiring a studio for it would waste your money. Our only request is that stylised images are labelled, so the line between a photograph and a generated image stays visible.