Weight-loss then-vs-now needs consented photos, not generated bodies

First-last only on real consented stills. Generating a before body is a likeness problem. The clinic already has the pair, or it does not have a then-vs-now. Inventing the earlier body — slimmer, heavier, "what they used to look like" — is a replica of a private person, not a production shortcut.

A deepfake of a private person is not a style. A patient is a private figure. TikTok's rule on that is permission, not labelling. Create a first-and-last-frame transition is allowed only when both photographs already exist and both were granted. The interpolate is the middle. The bodies are not available to invent.

Then-vs-now is two files you already own

The format is a transformation. Frame A is the consented before. Frame B is the consented after. Same person, same standing mark if you can get it, same lighting family. The model travels between those plates. That is first-last frame: a bounded problem with a destination, not a prompt that "shows the journey."

If you have only the after, you do not have a then-vs-now. You have a portrait. Asking image-to-video to "age them back" or "add the weight they lost" is generating a body the person never stood for in that state, from a likeness they did not grant for that act. A standard model release authorises using photographs that were made. It does not authorise making new ones. The clause list is a likeness release that covers AI generation. This is general information, not legal advice.

Do not "fill in" a missing before from a staff guess, a chart, or a reference of someone else "in the same category." If a reasonable person who knows them would say "that is them," you have a replica. Category is not consent.

A generated before is not anonymised

Blurring the face and keeping the body does not get you out. Likeness is not only a face. A body that is still their body, in a clinic ad, is still them. "We changed 15% of the torso" is the same non-defence the private-figure post already killed for faces.

Minors are a hard stop. No synthetic likeness of anyone under 18, and no "age them up" from a school photo to manufacture an adult before. Do not.

A fully invented person — no reference image of a living patient — is a different object. That person still cannot deliver a first-person result ("I lost the weight, here is my stomach") as if they were a customer. That is a fabricated testimonial under the reviews rule, which AI testimonials and the fake review rule already maps. Dramatising a real, granted quote is production. Inventing the experience is the violation.

What the interpolate is allowed to do

When both stills are real and granted:

1. Attach them. The agent passes first_frame_url and last_frame_url and picks a first-last-capable model.
2. Prompt the camera only — locked-off morph, no new room, no extra people, no wardrobe the stills do not show.
3. Inspect the midpoint. Ends are held by construction. The middle is where a generator grows a third body. If the middle is wrong, reroll the motion, not the stills.
4. Keep the clip short. Then-vs-now is a state change. It is not a 30-second identity pass of one body while the model guesses the other.

What you can generate around the pair without touching the person: a diagram of a process, type, a clinic exterior you photographed, an attorney-style explainer with a real clinician on camera. What you cannot generate: the missing still, a lookalike "UGC creator," a synthetic patient narrating results.

The clinic test

Did this person stand in both photographs, and did they grant generation of a transition between those photographs?

Yes → first-last. Publish the interpolate as motion on evidence you already hold.

No, we only have the after → do not manufacture the before. Film a clinician. Use the after as a still, not as a replica seed.

No, we have photos but the release is a 2015 model release with no replica clause → read it. If it authorises use of the shoot, it probably does not authorise new performances. Get the clause or do not interpolate.

The expensive mistake is treating the missing before as a prompting problem. It is a consent problem. Two granted stills beat a better paragraph.

FAQ

Can I generate a before body if the patient consents in Slack?

Get a replica grant that names face and body, the two stills, the transition, the platforms, and a take-down. A Slack thumbs-up is not that document. The release that covers generation is the clause list.

If I label the video AI-generated, can I invent the earlier body?

No. Labelling tells viewers the media is synthetic. It does not create permission from the person depicted, and it does not turn a fabricated before into a real one. TikTok's private-figure line is permission, not a caption.

Can first-last run on one photo plus "then they were heavier" in the prompt?

No. The first-last agent job needs two reference photos. Describe-the-end is a different job, and it will not land on a body you never attached.

Is a synthetic presenter talking about "typical results" safer?

As a spokesperson who is not pretending to be a patient, maybe — still disclose, still stay off personal medical advice from a fake expert. As a witness ("I used this clinic and here is my stomach"), no. That is a testimonial with no person behind it.