AI Head Swap: How to Change the Model in a Product Photo

An AI head swap replaces a person’s whole head in a photo, hair and head shape included, while the body, garment, and background stay the same. For a brand, the technique is the easy part. The hard part is rights. Using a real person’s likeness in a commercial image needs their consent, and the model release from your original shoot probably does not grant it.

Most brand teams know this situation. The autumn shoot is done, the images are approved, and then someone asks whether the same set can run in a market where the cast reads wrong. Or the model’s contract term expires in March and the images are still on live PDPs. Or a retailer wants the range on a different body type and the next studio day is eight weeks out. The photograph exists. The person in it is the problem.

This covers what a head swap does, when a brand needs one, how to do it, and the rights and disclosure layer that decides whether the result can ship. It is written for fashion brands, ecommerce teams and the agencies that produce their catalogs, working with the consent of every person in the image. It is not for anyone making a meme, and it is not a guide to swapping faces of people who did not agree.

Key Takeaways

• Head swap replaces the entire head including hair and silhouette. Face swap replaces facial features only and keeps the original hair. Pick head swap when the hairstyle or head shape is part of what has to change.
• For commercial use, regenerating the shot on-model usually beats swapping a head in. A swap inherits the pose, lighting, and rights problems of the source photo. A regeneration does not.
• A standard model release rarely covers AI modification. New York’s Fashion Workers Act requires separate written consent for a model’s digital replica, and routine retouching is not a digital replica (dol.ny.gov, accessed September 2026).
• California Civil Code 3344 requires prior consent to use a person’s likeness in advertising. Damages are $750 or the actual loss, whichever is greater (leginfo.legislature.ca.gov, accessed September 2026).
• Article 50(4) of the EU AI Act has applied since 2 August 2026. It requires deployers to disclose deep fakes: realistic AI images, audio or video that could pass as real. A realistic head swap is likely one (digital-strategy.ec.europa.eu, accessed September 2026).

What is an AI head swap?

An AI head swap replaces one person’s entire head in a photograph with another, including hair, head shape, and often part of the neck. It keeps the body, clothing, pose, and background from the original image. The model matches lighting direction, skin tone, and viewing angle at the join so the edit reads as one photograph rather than two.

The distinction that matters commercially is whose head goes in. Swapping in a second real person carries that person’s likeness rights into your image. Generating a head that resembles nobody in particular does not, though it can still carry a disclosure duty, for example under New York’s synthetic performer law and the EU AI Act.

Head swap vs face swap: which one you need

Both edits replace part of a person. They differ in how much.

The practical test is the hairline. If the original hairstyle works for the new image, a face swap is the smaller and more reliable edit. If the hair is wrong, a face swap will leave a person wearing someone else’s hair, and it reads immediately as wrong in close crops.

Head swap is the larger edit and the harder one, because hair has thousands of soft edges against the background. That is also where the artifacts appear first.

When a brand needs this, and when it does not

Most brand requests that arrive phrased as “can we swap the head” are better solved another way. Three cases are worth separating.

Cases where a head swap is the right tool:

• The photograph is irreplaceable. A campaign hero shot, an on-location image, a set you cannot rebuild.
• The garment fit and drape in the original are the asset, and regenerating risks losing them.
• You need one specific change to an approved image late in a cycle.

Cases where regenerating on-model is better:

• You have flat lays or packshots and no on-model imagery yet. Nothing to swap into.
• You need the same garment across several model appearances. Swapping heads one at a time compounds artifacts; generating from the garment does not.
• The original model’s usage term is expiring. A swap built on their pose and body may still be a derivative of their performance.

Cases where the answer is a reshoot or a license:

• The person whose head you want to swap in has not consented.
• The image is going into paid media in a market with strict likeness law and you have no written grant.

The middle case, regenerating on-model, is the one teams use least. If you have the garment, generating a fresh on-model image removes the likeness license, because no real person’s likeness is in the result. Disclosure can still apply: New York’s synthetic performer law and the EU AI Act both cover realistic AI people in some ads, as the disclosure section below explains. DesignerBox’s Clothes on a model app and virtual try-on both work from the garment, and the wider tool set is compared in our guide to AI fashion model generators. For bags, jewelry and other worn products, AI on-model photography tools by product shows which tools cover each type. For free routes, free AI virtual try-on tools sellers can use checks which ones allow commercial use of the results.

How to swap the model in a product photo, step by step

If a swap is the right call, the sequence below produces the fewest retries. Steps 1 and 2 are the ones people skip, and they are the ones that decide the result.

1. Confirm the rights before you generate. Get written consent from the person whose likeness goes in. Check that your source photo’s release permits modification. If either is missing, stop here.
2. Match the source images on angle and light. A front-facing head will not sit on a three-quarter body. Match the direction the light falls from, and the height of the camera relative to the face.
3. Work from the highest resolution you have. Edits compound. Starting from a web-sized JPEG means the join is soft before you begin.
4. Write the edit as a preservation instruction, not a replacement one. Google’s own prompting guidance for this pattern is to “change only the [specific element] to [new element/description]. Keep everything else in the image exactly the same, preserving the original style, lighting, and composition” (ai.google.dev, accessed September 2026). Naming what must not change does more work than naming what must.
5. Check the hairline, the neck join, and the ears at full resolution. These fail first. Ears are the reliable tell, because they sit at the boundary and are easy to get subtly wrong.
6. Check the hands and any reflective surface. Skin tone that matches at the face but not at the hands is the second most common miss.
7. Log what you generated and from what. If disclosure is required later, you need a record of which images were modified.

Step 5 needs extra care. Most head swap output looks correct at thumbnail size and wrong at PDP zoom, which is the size that matters. The same resolution and accuracy discipline that applies to any generated product image applies harder here.

Holding a face across moving frames is the harder version of this same problem. If the shot is video rather than a still, see how to keep a character consistent in AI video. To keep the person and change the garment in a clip instead, see how an AI video clothes changer works.

Which models hold a face, and where they break

Identity preservation across an edit is a documented capability. Each provider states what its model holds, and those statements are worth reading.

Google’s Gemini API documents Nano Banana Pro (Gemini 3 Pro Image) as “the premium choice for the most complex visual tasks”. It mixes up to 14 reference images, including up to 5 character images to keep a character consistent, and outputs 1K, 2K or 4K (ai.google.dev, accessed September 2026).

Black Forest Labs documents FLUX.1 Kontext as built to “preserve unique elements of an image, such as a reference character or object in a picture, across multiple scenes and environments”. It also documents Kontext as built to “make targeted modifications of specific elements in an image without affecting the rest” (bfl.ai, accessed September 2026). Their docs add that Kontext “excels at character consistency, even after multiple edits”, and that “for new projects, we recommend FLUX.2 with multi-reference support and up to 4MP output” (docs.bfl.ai, accessed September 2026).

OpenAI documents GPT Image edits with mask support, and is candid about the limit: “Masking with GPT Image is entirely prompt-based. The model uses the mask as guidance, but may not follow its exact shape with complete precision” (developers.openai.com, accessed September 2026). That guide documents GPT Image 2.5 as the current version.

The honest counterweight comes from Google, on its own product page: advanced features “like masked editing, major lighting changes (like day to night), or blending multiple images may sometimes produce unnatural results, visual artifacts, or disjointed scenes” (deepmind.google, accessed September 2026).

Read those together and the picture is consistent. Region-locked editing is approximate across every major provider, by their own documentation. Budget for several tries rather than expecting a first-pass result, and check every result at full size.

In DesignerBox, the model is one step in a workflow. The workflow picks the image or video model for each step, Nano Banana Pro and Kontext Multi among them, and the model list names the rest.

Rights and consent behind a head swap

The rights decide whether the image ships. This is general information, not legal advice. Check the law in every market where the image runs.

A likeness in an advert needs consent. In the US, the right of publicity comes mainly from state law, with no single federal statute (law.cornell.edu, accessed September 2026). California Civil Code 3344 covers “any person who knowingly uses another’s name, voice, signature, photograph, or likeness, in any manner, on or in products, merchandise, or goods, or for purposes of advertising or selling” without prior consent. Damages are $750 or the actual loss, whichever is greater, plus profits and legal fees (leginfo.legislature.ca.gov, accessed September 2026). New York Civil Rights Law 50 and Tennessee’s ELVIS Act set similar rules. The ELVIS Act took effect on 1 July 2024 and protects a person’s voice as well as their likeness (Tennessee Public Chapter 588, accessed October 2026). This is ordinary right-of-publicity law, and a generated image is still a use.

California has since legislated digital replicas directly. Since 1 January 2025, AB 2602 can void a contract term that allows a digital replica of a performer, unless the term describes the uses clearly and the performer had a lawyer or a union (leginfo.legislature.ca.gov, accessed September 2026). AB 1836, effective 1 January 2025, makes people liable for digital replicas of a dead person’s voice or likeness without consent, with damages of $10,000 or the actual loss, whichever is greater (leginfo.legislature.ca.gov, accessed September 2026). Civil Code 3344 itself names digital replicas from 1 January 2027, the default start date for a California law without an urgency clause. California signed SB 1111 on 30 September 2026, and it says a person’s voice or likeness includes a digital replica (California SB 1111, accessed October 2026).

Your existing model release probably does not cover this, and one state now says so in law. Since 19 June 2025, New York’s Fashion Workers Act has required “separate and explicit written consent” for the use of a model’s digital replica. The model must approve in writing “the scope, purpose, rate of pay, and length of time the replica will be used” (dol.ny.gov, accessed September 2026). The New York Department of Labor draws the line explicitly: “A digital replica does not include routine photographic edits like color correction, minor retouching, or other standard post-production modifications” (dol.ny.gov, accessed September 2026). Routine retouching is outside that rule. Replacing someone’s head is likely inside it.

The risk grows when the same face appears again and again. A one-off swap needs consent for one image; a reusable AI brand character needs a release that names synthetic generation, a term, and a territory.

Consent requirements also differ by state. New York Civil Rights Law 50 requires written consent before you use a living person’s name, picture, likeness or voice in advertising, and using it without consent is a misdemeanor (nysenate.gov, accessed September 2026). California’s 3344 says “prior consent” and does not specify writing. Get it in writing anyway, because you will have to prove it later.

Licensed stock does not solve it either. Getty’s content license agreement, in its section on AI and machine learning use, states that “Getty Images does not represent or warrant that consent has been obtained for such uses with respect to model-released content” (gettyimages.com/eula, accessed September 2026). A model release attached to a stock photo is not a grant to synthesize that person’s likeness.

The model providers require consent too, independently of the law. Google’s Generative AI Prohibited Use Policy lists content that “violates the rights of others, including privacy and intellectual property rights”, with “using personal data or biometrics without legally-required consent” as an example (policies.google.com, accessed September 2026). Black Forest Labs’ terms, last updated 1 August 2026, say that where you generate output depicting “real, identifiable individuals”, you “represent and warrant that you have obtained that individual’s consent” (bfl.ai, accessed September 2026).

One nuance: in the policies we checked, neither Google nor Black Forest Labs bans face or head swapping as a technique. They regulate the outcome: impersonation, privacy violations and use of a likeness without consent. Swapping in a model who agreed in writing is a different act from swapping in someone who did not, and the policies are written to reach the second.

What the platforms require when the ad ships

Disclosure is a separate obligation from consent. Clearing one does not clear the other.

EU AI Act Article 50 has applied since 2 August 2026. Deployers of an AI system that generates or manipulates image content constituting a deep fake “shall disclose that the content has been artificially generated or manipulated” (AI Act Service Desk, accessed September 2026). A realistic head swap that could pass as a real photo is likely one. Deep fakes in clearly artistic, creative or fictional works still need a label, in a lighter form that does not spoil the work. The Commission’s guidelines say ads qualify for this only in specific cases. A plainly commercial product image usually does not, so brands should not plan around it (Commission guidelines, C(2026) 5054, published 20 July 2026, not binding, accessed September 2026).

New York’s synthetic performer law has applied since 9 June 2026. If you make an ad and you know it contains a synthetic performer, the ad must say so in a way people will notice. A synthetic performer is a digital asset that looks like a human performing, where the person is not recognizable as any real performer. An edited photo of a real, recognizable person is outside that definition, but a generated head that resembles nobody can fall inside it. The civil penalty is $1,000 for a first violation and $5,000 for each later one (nysenate.gov, accessed September 2026).

The two big ad platforms have different rules.

TikTok allows ads with AI-generated or heavily AI-edited media, and asks you to add the AIGC label or your own clear disclaimer, caption, watermark or sticker. Showing someone doing something they did not do counts as a significant AI edit. If TikTok finds undisclosed AI content, it rejects or restricts the ad (ads.tiktok.com, accessed September 2026). TikTok also does not allow AI content that falsely shows a public figure making or receiving an endorsement, even with a label (tiktok.com, accessed September 2026).

Meta adds an “AI info” label to ads made or changed a lot with its own AI tools, and not to minor edits such as resizing or color correction. It also checks ads for signs of third-party AI, such as C2PA metadata, and labels them when it finds signals. Its advertiser self-disclosure duty covers ads about social issues, elections or politics (facebook.com and transparency.meta.com, accessed September 2026).

Check platform policy on the platform’s own help pages before each campaign, including the rules described here. TikTok, Meta and Google changed their AI labeling rules several times in 2025 and 2026.

Google adds SynthID, an invisible watermark, to images from Nano Banana (ai.google.dev, accessed September 2026). Google says it is designed to survive cropping and filters, and may not be found after many edits. Plan disclosure as a normal production step rather than a risk to manage.

What this does not solve

Here are the limits, because the technique is often oversold.

A head swap does not fix a bad source photograph. Pose, lighting, and garment fit come from the original, and a new head on an awkward pose is an awkward pose. It does not reliably survive a PDP zoom on the first pass.

It does not give you rights you did not have. It does not produce a consistent model across a range either, because each swap is an independent edit. Holding one identity across forty SKUs is a brand consistency problem. Reference conditioning and saved workflows solve it, and repeating a swap does not. Holding one look across a whole drop takes four separate locks: model identity, lighting, framing, and garment fidelity.

Against a reshoot, the cost comparison is in our breakdown of product photoshoot cost. Whichever route produces the image, it still has to earn its place in the PDP gallery.

On-model runs in DesignerBox

The value sits in the second garment. Once a single on-model result is approved, save the setup as a workflow and run it again on the next garment, and the one after that. You set the brand rules, the light and the framing once. The workflow reads them on every run, so a saved workflow runs the same way on the next product. Every run shows its cost before you start it.

The approved shot, the next garments and your brand rules stay in one account, so you do not rebuild the set in a second tool. To use a generated model instead of a real person, start from an avatar. An avatar run returns nine fixed poses of one person. Uploading your own photos and the commercial license you need to run results as paid media start on the Pro plan. Virtual try-on starts on the Premium plan. What each surface does is on the product page.

For an agency or brand team that repeats this on every new garment, start from a model pose set template. Use a model who agreed to the use in writing, or a generated model who is no real person, and add your garment.

FAQ

What is an AI head swap?

An AI head swap replaces a person’s entire head in a photograph, including hair, head shape, and often part of the neck, while keeping the body, garment, pose, and background from the original image. The model matches lighting, skin tone, and viewing angle at the join. It differs from a face swap, which replaces only facial features and keeps the original hair.

What is the difference between a head swap and a face swap?

A face swap replaces the eyes, nose, mouth, and skin while keeping the original hair and head shape. A head swap replaces the whole head including hair and silhouette. Use the hairline as the test: if the original hairstyle suits the new image, a face swap is the smaller and more reliable edit. If the hair has to change, only a head swap will look right in close crops.

Can I use an AI head swap in a commercial advert?

Only with rights to the likeness in the result. California Civil Code 3344 requires prior consent to use a person’s likeness in advertising, with damages of $750 or the actual loss, whichever is greater, and New York Civil Rights Law 50 requires written consent. Model provider policies require consent as well. Generating a model who is not a real, identifiable person avoids the likeness question, though disclosure duties can still apply, for example New York’s synthetic performer law.

Does my model release cover AI modification of the photos?

Usually not. Since 19 June 2025, New York’s Fashion Workers Act has required separate and explicit written consent for a model’s digital replica, covering scope, purpose, pay and duration. Routine edits like color correction and minor retouching are not digital replicas (dol.ny.gov, September 2026). Getty’s license, in its section on AI use, disclaims any warranty that consent was obtained for such uses on model-released content (gettyimages.com/eula, September 2026). Treat AI modification as a separately negotiated term.

Do I have to disclose that a product image was AI-edited?

In the EU, the deciding question is whether the image is a deep fake. Article 50(4) of the AI Act has applied since 2 August 2026 and requires deployers to disclose deep fakes: realistic AI images that could pass as real. A realistic head swap is likely in scope. Ordinary retouching, such as color correction or lighting changes, usually is not. The lighter rule for artistic works covers ads only in specific cases. In New York, an ad with a synthetic performer must say so. Platform rules are separate and change often, so check current advertising policy before a campaign ships.

Should I swap the head or regenerate the shot on-model?

Regenerate, in most cases. A swap inherits the pose, lighting, and likeness rights of the source photograph, while generating an on-model image from your garment carries none of those. Swapping is the better answer when the original photograph is irreplaceable, such as a campaign hero or an on-location set, or when garment fit and drape in the original are the asset.

Which plan do I need for on-model work in DesignerBox?

AI video and virtual try-on both start on the Premium plan. The cost of an image run depends on the model, and the cost is shown before the run. Uploading your own photos and the commercial license needed to run results as paid media or on a live PDP start on the Pro plan. Plans and credits are on the pricing page. An agency that needs team features needs the Ultra plan, because every plan below Ultra is one seat.

Sources

Accessed September 2026, unless a line gives its own date.

• California Civil Code 3344 prior-consent requirement and damages: leginfo.legislature.ca.gov
• California AB 2602 on digital replica contract terms: leginfo.legislature.ca.gov
• California AB 1836 on digital replicas of deceased personalities, effective 1 January 2025: leginfo.legislature.ca.gov
• Right of publicity as mainly state law: law.cornell.edu
• Tennessee’s ELVIS Act, in effect since 1 July 2024: Tennessee Public Chapter 588, accessed October 2026
• California SB 1111, signed 30 September 2026, which adds digital replicas to Civil Code 3344: leginfo.legislature.ca.gov, accessed October 2026
• New York’s Fashion Workers Act, separate written consent for a digital replica: dol.ny.gov FAQ
• The routine-retouching exclusion from the digital replica definition: dol.ny.gov definitions
• New York Civil Rights Law 50 requiring written consent: nysenate.gov
• New York’s synthetic performer disclosure law, General Business Law 396-b: nysenate.gov
• EU AI Act Article 50 deployer disclosure duty: AI Act Service Desk and Commission FAQ
• The treatment of artistic works and ads, and the non-binding status of the guidelines: Commission guidelines, C(2026) 5054, published 20 July 2026
• Getty’s content license agreement disclaiming any warranty that consent was obtained for AI uses of model-released content: gettyimages.com/eula
• TikTok’s AI labeling rule for ads: ads.tiktok.com
• TikTok’s rule on AI content that falsely shows public figures: tiktok.com
• Meta’s AI info label and third-party AI detection: facebook.com/business/help
• Meta’s self-disclosure duty for social issue, election and political ads: transparency.meta.com
• Nano Banana Pro reference limits, resolutions, SynthID, and the prompt pattern for preservation instructions: ai.google.dev
• Google’s caveat that masked editing, major lighting changes and multi-image blending may produce unnatural results: deepmind.google
• Google’s Generative AI Prohibited Use Policy: policies.google.com
• FLUX.1 Kontext consistency across edits: bfl.ai and docs.bfl.ai
• Black Forest Labs’ terms on consent for real, identifiable individuals, last updated 1 August 2026: bfl.ai
• GPT Image mask behavior: developers.openai.com
• DesignerBox feature gating: DesignerBox pricing page (designerbox.ai/pricing), September 2026

Legal, policy and model capability claims verified from leginfo.legislature.ca.gov, nysenate.gov, dol.ny.gov, the Tennessee Secretary of State, law.cornell.edu, the European Commission, gettyimages.com, TikTok, Meta, Google, Black Forest Labs and OpenAI as of September 2026. California SB 1111 and the text of Tennessee’s ELVIS Act were re-checked on 2 October 2026. This is general information, not legal advice. Likeness and disclosure law varies by jurisdiction and changes quickly, so check current requirements for your markets before publishing. Individual results vary.