Reference Acting to AI Generation

All Missions

W-20

Film Making · AI Pipeline

One Take, Two Faces

Reference Acting to AI Generation

One actor's take, one voice recording, and a different face on screen — the same performance re-cast after the shoot rather than before it.

Client

Confidential

Year

2026

Sector

Film Making

Outcome

One Take, Two Faces

Mission Brief

The objective, and how we met it.

// The Challenge

Casting is decided before a frame is shot, and it is the most expensive decision to reverse. The question was whether a performance could be captured once and cast afterwards — with the timing, the delivery and the voice all preserved well enough to survive a side-by-side comparison rather than a flattering edit.

// What We Deployed

The character is built before any performance touches it. A reference portrait fixes the face, the wardrobe and the light; a turnaround sheet then locks that same character across front, three-quarter, profile and back, so every later frame has something to be consistent with. Only then does the performance arrive: a single take of an actor delivering to camera, whose timing, beats, head turns and closing gesture are transferred onto the built character and lip-synced to the original voice recording. Nothing is re-shot, nothing is re-read, and no motion-capture rig is involved. The output holds frame-for-frame against the reference across the full take, which is the part that actually decides whether a shot is usable: drift of even a fifth of a second reads as wrong long before a viewer can say why.

// BUILD

Before the performance

The character is made first.

A generated face drifts the moment the head turns unless there is something for it to be consistent with. These two plates are that something.

01 · Character portrait

01 · Character portrait

Generated first. Fixes the face, the wardrobe and the key light — everything the performance will later be hung on.

02 · Character sheet

02 · Character sheet

The same character turned through front, three-quarter, profile and back, with head detail. This is what a generated face is checked against when it moves.

// THE FILM

Watch it end to end

AI-driven acting — Reference matched expressions

0:33 · 1920×1080 · 30 fps

Play the case film

The reference clip is licensed and its watermark is left in deliberately.

// RESULTS

What It Proves

1

Take Required

28s

Held In Sync

0

Re-shoots

24fps

Generated Take

1080p

Delivered Master

// SURVEILLANCE

Visual Record

Captured from the field.

01 PLATES · DECLASSIFIED

// SCOPE

Deliverables

What we shipped.

Disciplines

  • AI Film Pipeline
  • Performance Transfer
  • Post Production

Tags

  • Film
  • AI
  • Performance

01

Single reference take transferred to a different character

02

Lip-sync locked to the original voice recording, not re-read

03

Frame-accurate across the full 28-second take

04

No motion capture, no re-shoot, no second performer

// LOG

Engagement Log

How the mission ran.

Every PAR2 engagement runs the same four-phase doctrine — adapted here to the specifics of this operation.

01

Recon

Mapped the terrain for Reference Acting to AI Generation — film making · ai pipeline constraints, users, and the real objective behind the brief.

02

Strategy

Set the operating plan and success criteria, aligning ai film pipeline & performance transfer & post production into a single line of attack.

03

Build

Engineered and iterated in the field — single reference take transferred to a different character, validated continuously.

04

Deploy

Shipped, monitored, and tuned. One Take, Two Faces stands as the headline outcome of the deployment.

// CREDITS

With thanks

Reference performance

Actors Intro to Camera Video

Headshot Company

The take that drives the timing, the beats and the closing gesture. Used with the owner's licence; the on-screen watermark is left in.

Character generation

Reference image and character sheet

Higgsfield

Both plates above were generated on Higgsfield before any performance was applied.

This case study documents a technique. Both performances are shown to explain how the work was made — not to suggest that either performer endorses, is associated with, or has any connection to PAR2 LABS or anything we sell. Neither likeness is licensed, sold, or offered for use by anyone else, and the film is not an advertisement for any person or product. All rights in the original performance and in the depicted likeness remain with their owners. If you hold either and want this taken down, write to ceo@par2labs.com and we will remove it.

Have a mission like this?