In Hollywood’s Digital Doubles, the Face Was the Easy Part

Hollywood has spent the past decade putting faces on bodies that don’t belong to them. An actor who died in 1994 returned as a villain. Stars in their late seventies played men decades younger. The headlines were about the face. The crews kept running into three other problems: how a body moves, whose voice comes out of it, and who signed off on any of it.

AI video models have made the face the cheap part. A generated avatar can be built from a few reference photos and a short voice recording, and it can look right in a single frame. A brand planning an AI spokesperson now faces the same three problems the studios did, usually with far less reference material and far less time to fix them.

Tarkin: a borrowed face and a paper trail

Peter Cushing played Grand Moff Tarkin in the first Star Wars film in 1977 and died in 1994. In Rogue One, released in December 2016, the British actor Guy Henry played the scenes, providing the voice and a motion-capture performance. Industrial Light & Magic then placed a digital likeness of Cushing over his performance.

Lucasfilm had permission from Cushing’s estate. Variety reported that Joyce Broughton, who oversees it, was moved by the result. But in 1993, Cushing’s company had signed a letter agreement with Tyburn Film Productions for a TV film that was never made, with a clause blocking the use of his likeness without Tyburn’s express prior written consent. Tyburn sued. England’s Court of Appeal struck out the claim in December 2025, nine years after the film’s release. The estate’s yes was valid. It still wasn’t the only document that mattered.

Bringing Tarkin back took a lookalike actor, a capture stage and a major effects house. A generated brand avatar can now start from photos pulled off a company website and a voice sample lifted from an old webinar. That lower barrier makes the consent chain matter more. When a likeness is this easy to build, the paperwork is what separates a licensed avatar from an impersonation.

Lawmakers have started to write that down. In March 2024, Tennessee added voice to its law protecting a person’s likeness, citing generative AI cloning. In June 2026, a US Senate committee unanimously advanced the NO FAKES Act, which would give every individual, celebrity or private citizen, a federal right over digital replicas of their voice and likeness. It is not yet law.

The Irishman: young faces, old bodies

For The Irishman (2019), Martin Scorsese wanted Robert De Niro, Al Pacino and Joe Pesci to play across decades without helmets or tracking markers on their faces. De Niro and Pesci were 76, Pacino 79. ILM built a rig with two infrared witness cameras flanking the main camera, plus software that rebuilt younger faces from the light and geometry it captured. The film holds 1,750 visual effects shots.

Hands show age, so they were altered too, and bodies were adjusted in compositing for some early flashbacks. A posture coach worked with the actors on set. Scorsese summed it up: “It’s about posture, it’s about movement, it’s about clarity of the eyes, everything.” At SlashFilm, Hoai-Tran Bui wrote that de-aging could smooth the skin but could not change the way a 76-year-old actor carries himself.

Generated avatars show the same mismatch. Give a model a reference photo of a 30-year-old and it returns a 30-year-old face, but the body language comes from the motion driving it or from generic patterns. Map a young face onto footage of an older presenter through video-to-video, and the source’s posture and gait come along. Leave a talking-head shot unguided, and a model tends to fill it with generic gestures: hands that move without purpose, nods that don’t land on the words. The fix is Scorsese’s: direct the body. Cast a reference performer who moves like the character, and reject takes where the hands say something the face doesn’t.

Dial of Destiny: hundreds of hours of reference

Indiana Jones and the Dial of Destiny (2023) opens with a prologue of about 25 minutes, set in 1944, in which Harrison Ford looks roughly 35 years younger. Ford began production at 78. ILM used what it calls FaceSwap, an umbrella for several techniques: a 3D computer-generated head, machine learning, paint, keyframe animation and lighting work, chosen shot by shot.

The team had hundreds of hours of footage from the first three Indiana Jones films, showing Ford under varied lighting, and compared its CG head side by side with that archive. And Ford played the scenes himself. “Harrison is super fit and energetic,” visual effects supervisor Andrew Whitehurst told VFX Voice. Face, body and voice all came from the same person.

ILM’s machine learning worked from a face documented across decades of film. A brand avatar usually starts from a handful of photos and a few minutes of video. That gap is a big reason generated characters drift. Each shot is a separate generation, and where the reference is thin, the model fills in what it hasn’t seen: the jawline, the hairline or the eye color shifts from one clip to the next, sometimes within a single shot as the head turns. More reference helps: photos from several angles and lighting setups, a short turnaround video, a locked first frame for each shot. So does review, with every shot compared against the reference sheet, the way ILM checked its head against the archive.

Digital doubles, from the screen to the contract

Milestones cited in this article

Digital-double milestones

Year	Event
1977	Peter Cushing plays Tarkin in the first Star Wars film
1994	Cushing dies
2016	Rogue One: Guy Henry performs, Cushing’s likeness added with his estate’s permission
2019	The Irishman: de-aging without face markers, 1,750 VFX shots
2023	Dial of Destiny: a 1944 prologue with a de-aged Harrison Ford
2023	SAG-AFTRA contract sets consent rules for digital replicas
2024	Tennessee adds voice to its likeness law, citing AI cloning
2025	Court of Appeal strikes out a claim over Cushing’s likeness
2026	US Senate committee advances the NO FAKES Act

What the actors negotiated, and what it means for a brand

By 2023, the question had reached the bargaining table. SAG-AFTRA members struck from July 14 to November 9, 2023. The new TV/Theatrical contract, ratified on December 5, 2023 with 78.33% of the vote, set rules for digital replicas.

Consent before creation. At least 48 hours’ notice before a scan, and consent before a replica is made.

Consent that can’t be buried. Clear and conspicuous, signed or initialed separately, with a reasonably specific description of the intended use.

Consent for each use. With limited exceptions, every use needs consent, and covering several projects means naming each one.

Payment, and voice counts. Performers are paid for the use, and the same rules apply to a replica of a voice as to an image.

Generated brand avatars rarely fall under a union contract, but the logic transfers to an avatar of an employee. The contract separates an employment-based replica, made with the performer’s participation for a specific project, from an independently created one, built from existing material for work the performer never did. An avatar made at a capture session with your head of sales is the first kind. One assembled later from her old webinar footage is the second, and deserves at least as much scrutiny.

Consent per use means an avatar approved for onboarding videos is not automatically approved for a TV spot. And the union treats a synthetic character that resembles a real performer as a replica, a useful test for an “invented” brand face that looks a lot like someone in the office.

What to ask before building a brand avatar

A convincing face is no longer the expensive part. Movement, voice and permission still are. Disclosure and what an AI spokesperson may claim are covered here. Before the first generation, ask for these in writing.

Consent for face and voice. Signed by the person, naming both, for this use.

A defined scope and term. Which videos, channels and languages, for how long, and what happens if the person leaves the company.

A check of earlier contracts. Agency deals, past endorsements, old releases. Cushing’s estate said yes, and a 1993 letter still went to court.

A reference package and a matching performance. Enough photos, angles and voice material to hold the character across shots, a reference performer whose movement fits the avatar’s age, and a clear answer on how drift between shots will be caught before delivery.