AI model likeness rights: four ways a real face gets in
Most brands think of AI model photography as one thing. For likeness purposes it is four different things, and the consent you need changes with each one.
| Scenario | Whose likeness is in the image | Consent you need | What goes wrong |
|---|---|---|---|
| Fully synthetic model, generated from a text description | Nobody's, if it is truly invented | None from a person; your tool's licence still applies | The face drifts toward a real or famous person |
| Digital twin of a model you hired | That model's | Separate written digital-replica consent: scope, use, pay, duration | Twin keeps being used after the contract ends |
| Face swap onto a real body photo | Two people: the face and the body | Consent from both, for this combined use | The body model never agreed to wear someone else's face |
| Garment put on a stock or library model | The person in the stock image | A licence that allows AI alteration and commercial use | Stock licence forbids manipulation or sensitive uses |
The first row is the one covered by most articles about AI-generated model photography for fashion, and it is the cleanest. The other three are where the new laws bite, because each one creates a picture of a real person doing something they were never photographed doing.
Digital twins of real models: what the new laws ask for
A digital twin is a set of reference photos of a real person that an AI tool uses to produce new images of them. You shoot a model once, in one outfit, and the tool gives you the same person from the side, the back, in the next season's collection or in a different setting. For a small brand it is the most practical use of AI on people, and it is also the use most clearly targeted by recent legislation.
New York. The Fashion Workers Act, in force since June 19, 2025, requires clear and conspicuous prior written consent for creating or using a model's digital replica, and that consent has to be separate from the representation agreement. It must spell out the scope, purpose, rate of pay and duration of the use. Powers of attorney that previously let agencies approve replica use on a model's behalf no longer count, and the state Department of Labor enforces it, with civil penalties reported at $3,000 for a first violation and $5,000 for each one after.
California. AB 2602, effective January 1, 2025, makes a contract term unenforceable for new performances if it lets a digital replica stand in for work the person would otherwise have done in person, unless the term carries a reasonably specific description of the intended uses and the person was represented by a lawyer or a union. A vague "all media, all uses" clause signed by an unrepresented freelance model is exactly the kind of term that fails.
Tennessee. The ELVIS Act, in force since July 1, 2024, extended the state's publicity law to cover voice as well as name, photograph and likeness, and it reaches those who make an unauthorised likeness available, not only those who create it.
Other states have similar bills, and a federal digital-replica bill has been proposed in Congress more than once. The direction is consistent: consent to a photo shoot is not consent to a digital replica, and broad consent is treated with suspicion.
If you could not show the model a list of every image you plan to generate of them, with what it is for and what you are paying, you probably do not have the consent the new laws describe.
The digital-replica clause: weak wording vs. wording that holds
You do not need a long contract. You need a short, separate document with five specific terms. Here is what the difference looks like in practice.
Weak (typical shoot release)
- "The model grants all rights to their image in all media worldwide in perpetuity."
- Replica use buried in the general release or agency agreement
- No mention of AI, new poses or new garments
- Paid once for the shoot day, nothing for generated images
- No end date and no way to withdraw
Holds up (separate replica consent)
- "The brand may generate new images of the model from these reference photos showing garments from the listed collections."
- A standalone document, signed separately and dated
- Names the channels: own site, named marketplaces, paid social
- States the fee: per collection, per image or a flat term fee
- Runs for a fixed term, with what happens to the twin at the end
Three terms deserve extra care.
- What the twin may show. Name the product categories. A model who agreed to knitwear did not agree to swimwear or lingerie, and a model who agreed to your brand did not agree to appear in a partner's ad. List any exclusions they ask for, such as alcohol, political content or competitor products.
- Pay that tracks the use. New York explicitly requires the rate of pay. A shoot-day fee that silently covers years of generated images is the clause most likely to be challenged. A per-collection or per-term fee is simple to administer and easy to defend.
- The end of the term. Say whether existing generated images may stay live after the term ends, for how long, and that no new images will be generated. Say that the reference photos and any trained model or saved identity will be deleted, and who confirms it.
For how copyright in the generated images works, which is a separate question from the model's consent, see Who Owns Your AI Product Photos?
Face swaps: two people, one image
Face swapping is common in fashion catalogs for a reason: you can shoot an entire collection on one body model and put your contracted face model on every shot, or keep the same face consistent across shoots done months apart. It also puts two real people into a single picture, and both have a likeness interest in it.
The face model is the obvious one. Their face is appearing on a body that is not theirs, in clothes they never wore, in a pose they never struck. That is a digital replica in every sense the New York law describes, and it needs the same separate written consent as a digital twin.
The body model is the one brands forget. Their figure, skin, tattoos, hands and posture are all still in the image, presented as someone else. A standard release for a body-double or fit-model shoot often says nothing about the face being replaced. If yours does not, add a line saying the body model agrees that a different person's face may be placed on the images, and that they will not be credited or identified.
- Get both consents before the swap, not after the images are live.
- Do not swap faces in photos with other identifiable people in them unless they are covered too. Tools that only change the chosen face and leave the rest of the frame untouched help, but they do not remove the consent question for the people left in the shot.
- Never swap in a face you found online, from a social profile, a stock site or a public figure. No licence you can buy for a downloaded photo covers putting that person's face into your product imagery.
Store the face model's replica consent and the body model's release in the same record as each swapped image. When a marketplace or a model's agent asks, you need to show both, for that specific picture.
When a "synthetic" model looks like someone real
A model generated from a text prompt has no one to sign a release, but that only stays true if the face really is no one. Image models learn from photographs of real people, and two habits push their output toward real faces.
- Naming a person in the prompt. "A model who looks like" a celebrity, an influencer or a known fashion model is asking for a likeness. Even phrased as a style reference, it puts that person's identity into your product imagery and exposes you to exactly the publicity claims the synthetic route was meant to avoid.
- Reference images of real people. Uploading a photo of an actual person "for the vibe" and then using the generated face is, in practice, a digital replica without consent.
Even without either habit, a generated face can land close to a recognisable person by chance. Before a synthetic model becomes the face of a collection, run three checks:
- Reverse image search the face on two engines and look for close matches to real, identifiable people.
- Show it to people outside the project and ask if it reminds them of anyone. Recognition is what publicity law turns on, and a fresh pair of eyes catches what the team stopped seeing.
- Record the prompt and settings that made the face, so you can show it was generated from a description and not from a person.
If a face fails any check, generate another. A new face costs a few minutes; a likeness claim on a face that is already on 400 product pages costs a re-shoot.
A likeness record for every image with a person in it
The practical failure is rarely a missing contract. It is a contract no one can find when the image goes to a new marketplace, a paid campaign or a wholesale partner three years later. A short record per person, linked to every image that uses them, solves it.
| Field | Example | Why it matters |
|---|---|---|
| Person and role | Face model / body model / synthetic | Tells you whose consent applies |
| Consent document | Replica consent v2, signed and dated | Proof of separate written consent |
| Permitted categories | Knitwear, outerwear; no swimwear | Stops scope creep across collections |
| Permitted channels | Own site, two named marketplaces, organic social | Paid ads and partners often need their own approval |
| Fee basis | Per collection | Shows pay tracks use, as New York requires |
| Term end and deletion | Twin deleted at term end; live images stay 6 months | The date your catalog has to act |
| Images using this person | SKU and image IDs | Lets you find and pull every image in an hour |
The last row is the one that saves you. When a model withdraws, a term ends, or an agent disputes a use, you need to find every image that person appears in, including the generated ones that were never part of a shoot. If you push images to your store from one place, keep the record there too, so the image and its consent travel together.
Where AI tools fit, and where the responsibility stays with you
AI tools now make every scenario above routine. In Retouchable, for example, you can build a model from photos of a person you already shoot with and generate new angles of them, or swap a contracted face onto your shoot photos while the rest of the frame stays as shot. Those are exactly the workflows that save a brand a season of re-shoots, and exactly the ones that need the consent covered above.
No tool can sign a consent on a model's behalf or know what your contract says. The responsibility for whose face is in your catalog stays with the brand. A sensible division of labour:
- Before generating: confirm a separate replica consent exists for every real person whose photos go in as reference, and that the garment category is inside its scope.
- Before publishing: check synthetic faces for resemblance and make sure each image is linked to its likeness record.
- When labelling: follow the disclosure rules for your markets. Regulators increasingly treat realistic synthetic people as content that has to be labelled; our guides on labelling AI-generated product images and the EU AI Act's labelling rules cover the details.
- At term end: delete reference photos and saved identities for that person, stop generating, and pull or keep live images exactly as the consent says.
Done this way, AI model photography is not a legal risk you tolerate for the savings. It is a better-documented process than most traditional shoots ever had, because every generated image of a real person has a consent, a scope and an end date attached to it.