Published: January 13, 2026
Deepfakes (a portmanteau made from deep learning, a type of machine learning, and faking things) and face replacement in film are the practice of changing who a face appears to be in a finished shot, using a mix of VFX work and machine learning so the face fits the scene’s camera, lighting, and performance. The goal is shot-level realism that holds up in the edit, including close-ups.
This article covers face identity work in filmed scenes and the real production choices that come with it (planning, approvals, post workflow, and delivery checks). It does not cover voice cloning, full-body character generation, or general “AI video” topics unless they directly change how you replace a face.
Face work matters in film because your brain treats faces as proof. When a face reads as real, you follow the character and the scene without friction. When a face reads as wrong, you notice the technique, and the moment loses focus. There is also a second layer that matters just as much: consent, rights, and trust. Face replacement can solve real production problems, and it can also create serious harm when it is used carelessly. If you want FilmDaft’s baseline for how AI works in creative tools, read What Is AI? A Plain-English Guide for Creators.
Practical note: Rules depend on your country, contracts, and distribution plan, so when in doubt, always contact an entertainment lawyer. Treat this page as production guidance that helps you ask the right questions early, then confirm details with your legal and union contacts when a real person’s likeness is involved. For FilmDaft’s longer consent and permissions breakdown, see Consent and Digital Replicas: What Creators Should Know.
Key terms you will hear on set and in post
People use the word deepfake to mean a lot of different things, so it helps to name the parts. Once you label the work correctly, you can budget, schedule, and review shots with clear expectations. If you want a quick language guide for AI terms that show up in video tools, Common AI Terms in Video Tools can help.
- Deepfake: A general label for synthetic media made by a trained model. In film talk, it often means a model-driven face swap that maps one face onto another across angles and expressions. For the difference between machine learning, deep learning, and generative AI, see Machine Learning, Deep Learning, and Generative AI.
- Face replacement: The broader craft category. A face can be replaced with 2D compositing, a 3D digital head, model output, or a mix of all three.
- Plate (original footage): The original live-action shot you are integrating into. Most face problems show up when the replacement stops matching the plate’s light, blur, grain, or motion. If you want the FilmDaft definition in plain words, see What Is a VFX Plate?.
- Digital double: A CG version of a performer (head or full body) that can carry motion and lighting in a shot when practical footage cannot do the job.
- De-aging: Face work that targets age cues while keeping identity consistent, so the shot reads as the same person at a different time.
- Digital replica: A contract and labor term tied to permission, control, and compensation when a person’s likeness is captured or recreated for use in a project or later reuse. FilmDaft covers the practical consent side in Consent and Digital Replicas.
What makes a face replacement believable
A good face replacement is rarely “about AI.” It is about matching the reality the camera recorded. If you keep that mindset, you judge shots with practical checks instead of gut feeling, and you can explain what is wrong in a way your post team can act on.
Motion match
Faces feel real when the face moves like it belongs on the head in that shot. You look for stable head tracking, correct jaw motion, and blinks that land on natural beats. Small timing issues matter, especially when the camera is close.
Light match
Lighting mismatch is one of the fastest ways to break the illusion. Skin has soft shadows, sharp specular highlights, and subtle color shifts that react to the key and fill. If the replacement keeps the “same face” under changing light, you will spot it.
Texture match
Real footage has pores, fine wrinkles, lens softness, sensor noise, and film grain. A replacement face can fail because it looks too smooth, or because it carries a different noise pattern than the plate. Texture also changes with focus, so depth of field has to match.
Cut-context match
You never watch a face in isolation. You watch it after the previous shot and before the next one. A face that seems fine on a still frame can feel wrong when it cuts next to a practical close-up with real skin response and real lens behavior.
Sound and lip sync
Dialogue makes face work stricter because you track mouth shapes without thinking. If the edit changes timing, the face has to change with it. A clean comp can still feel wrong when mouth motion drifts away from the final sound edit.
How face replacement works in practice
Film-quality face replacement follows a pipeline because the shot has many moving parts. A model can help with the face layer, yet the shot still needs tracking, compositing, and approval passes that match the rest of your post workflow. If you want FilmDaft’s role-level view of who does what in the final stage, see What Does a Compositor Do in Film?.
- Define the intent at the shot level. Write down what must stay true: identity, performance, and camera reality (lens, distance, movement, exposure). This gives you clear pass or fail checks when you review the shot at full resolution.
- Collect structured reference. You need target reference (the face you want on screen) and a source performance (the person acting in the shot). Good reference is planned: neutral angles, a set of expressions, and lighting samples that match the scene.
- Track the head and face motion. The team tracks head movement and facial motion so the replacement follows the same motion as the plate. You will hear terms like matchmove (matching camera and head motion) and face solve (tracking facial motion). If the track jitters, the face layer jitters too.
- Create a first-pass replacement. Some shots start from model output, and some start from a 3D digital head or traditional comp work. Either way, the first pass is a draft that proves alignment, expression range, and eye line.
- Composite for plate realism. Compositing matches color, contrast, highlights, shadow edges, depth of field, motion blur, grain, and noise. Hairlines, ears, glasses, hands, smoke, and fast props often need rotoscoping and paint cleanup so edges behave like the original footage. If you want FilmDaft’s clean definition of “plate,” keep What Is a VFX Plate? nearby.
- Review, revise, and approve in context. You review at real speed and at full resolution, then you check frame-by-frame on the hard parts. You also compare against nearby shots for continuity because a face that “works” alone can fail inside the sequence.
Planning the shoot so post has fewer surprises
Face replacement gets easier when you plan it the same way you plan any VFX-heavy moment. You keep the performance readable, you reduce variables that break tracking, and you capture reference that matches the scene instead of collecting random clips.
Camera choices that keep tracking stable
Fast motion blur, heavy shake, and extreme profile turns raise the difficulty. If the scene demands chaos, you can still make the shot more workable with choices like a faster shutter, cleaner key direction, and enough exposure to keep facial detail in the plate.
Lighting reference that matches the scene
Stylized lighting can look great, yet it also raises the bar for face work. If you have neon, strobes, firelight, or strong colored sources, capture reference under those same sources. Simple tools help here, including gray and chrome spheres that show light direction and intensity.
Performance coverage that protects the face
Face replacement depends on what the actor does in frame. Long dialogue takes with heavy head turns can be harder than shorter beats with clear eye lines. If you can, plan coverage that gives you at least one angle where the face stays readable and consistent.
When a double is involved
Face replacement often connects a body performance to a specific identity, especially in stunts or demanding physical action. If you want the on-set baseline for what a double does (before post work enters the picture), see What Is a Body Double in Film?.
On-set documentation that saves post time
Small notes become big later. Lens info, camera settings, and a clean record of which takes are “VFX intended” can prevent confusion during turnovers. When you have time, capture a quick neutral reference of the source performer under similar lighting; that reference can rescue a tricky shot.
Asset security as part of production, not an afterthought
Face reference, scans, and high-resolution plates are sensitive because they can be reused outside the project. Limit access, watermark internal review exports, and keep datasets and model files behind the same permissions you use for unreleased cuts. If you want a broader checklist you can apply to professional deliveries, see Risk Checklist for Using AI in Client Work.
Real use cases and what they look like on screen
Face replacement shows up in different types of projects for different reasons. The examples below show different goals, which help you think about what level of realism and what level of consent you need.
Finishing scenes after an actor’s death
In Furious 7 (2015, Original Film), Paul Walker’s remaining material was completed with stand-ins and large-scale face replacement work.
Weta Digital completed Paul Walker’s work across roughly 260 shots, depending on how shots are counted, and many shots relied on full digital head replacement plus compositing work that matched the plate’s lighting and texture. The craft goal was continuity inside the cut; the production goal was finishing the film without breaking the character’s presence.
Replacing a face on a body performance for a digital character
Rogue One: A Star Wars Story (2016, Lucasfilm) is widely discussed for Grand Moff Tarkin. The shot work combined an on-set performance with a CG facial/likeness pipeline designed to match Peter Cushing’s face in the plate and in the cut.
This kind of work becomes strict in close-ups and extreme close-up shots because you see eye moisture, tiny mouth shapes, and skin highlights that must match the plate and match the performance.
De-aging for a long flashback
Indiana Jones and the Dial of Destiny (2023, Lucasfilm) used de-aging for extended sequences. Machine learning helped ILM search and match archival reference on a frame/shot basis, then compositing and CG work finished the shot for the plate’s light, blur, and texture. and the work relied on large amounts of archival reference to keep younger facial features consistent.
De-aging still has to obey the same rules as face replacement; the face must match the shot’s lighting, blur, and grain, or you notice the technique.
Protecting identities in documentary
Welcome to Chechnya (2020, HBO Documentary Films) used machine-learning face replacement (often described as deepfake-like) as a practical witness-protection method for faces on screen.
The idea is simple to describe and hard to execute: you keep the real performance beats, then you overlay a different identity so the person stays safer. In this context, consent and harm reduction sit at the center of the workflow, and the effect exists to protect real people rather than to create spectacle.
Risks you need to plan for
Face replacement carries extra risk because it touches identity, and identity is tied to rights and reputation. If you treat the effect as a planned decision with written permissions, approval gates, and security controls, you reduce the odds of expensive surprises later. FilmDaft collects adjacent guidance in the Ethics, Law, and Provenance in AI Filmmaking section.
Consent, contracts, and “digital replica” terms
If you put a real person’s likeness on screen, you need written permission that matches the actual use. Scope matters (which shots and what context), and distribution matters (where and for how long). Union and contract language may treat a recreated likeness as a digital replica, and you may see requirements around informed consent, compensation, and usage reporting. For release basics that help you write clearer paperwork, see What Is a Talent Release Form? and What Is a Model Release Form?.
False performance and meaning drift
A face swap can create a performance that the person never gave. Even with permission, a new edit can shift meaning, especially in dialogue scenes where expression changes tone. That is why approval paths often focus on moments where face work changes intent, implies endorsement, or changes what a person seems to say.
Disclosure expectations and trust
Trust drops fast when face manipulation appears inside a format that signals authenticity, such as documentary testimony, news-adjacent work, or client statements. Regulation also moves in this direction. For FilmDaft’s practical breakdown of the transparency rule, see Explaining the EU AI Act Deepfake Disclosure. If you also want a workflow for keeping a paper trail and machine-readable provenance, Content Credentials and Provenance (C2PA): A Creator Workflow connects well to this topic.
Security and leakage
Training clips, scans, and clean face reference can leak. Once that material exists outside your control, it can be reused in ways you never approved. Tight access control, watermarking, and clear data retention rules help, and they also help you explain your process to partners who care about confidentiality. If you want a practical set of checks for professional deliveries, Risk Checklist for Using AI in Client Work is designed for that.
Craft risk and schedule risk
Most face replacement failure looks like small mismatch. Eyes slide, teeth flicker, skin looks too smooth, or the face stays stable while the plate has real micro-shake and sensor noise. These problems take time to fix because they usually require tracking updates, paint cleanup, and careful compositing passes. If you want a broader mental model for predictable AI mistakes across formats, see Limits and Failure Modes in AI Output.
Credit and labor boundaries
Face replacement blends a source performance, a target likeness, and post work that builds the final facial motion you see. That can create credit tension if you never defined who approves the final face, who owns the creative decision, and how the work is documented. A simple habit helps here: name the approvals and credit plan before the first heavy VFX turnover.
A real-world reminder about rights disputes
Rights arguments can show up years after release. In December 2025, the UK Court of Appeal struck out Tyburn Film Productions Limited’s unjust enrichment claim in (1) Lunak Heavy Industries (UK) Limited (2) Lucasfilm Ltd LLC v Tyburn Film Productions Limited, a dispute tied to the digital recreation of Peter Cushing’s likeness for Rogue One: A Star Wars Story (2016, Lucasfilm). It is a reminder that likeness questions can stay alive long after delivery, so you want clear written permission, scope limits, and approval records from the start.
A decision workflow you can run before you commit
Before you greenlight face replacement, you want a workflow that forces clarity. These questions keep you focused on the shot problem, the consent problem, and the realism problem, and they keep you from spending money on work you could have solved with coverage or editorial. For a wider overview of where AI fits across production phases, you can also skim Artificial Intelligence in Filmmaking: A Practical Guide and Overview.
- Name the shot problem. Write it as a sentence you can hand to post, such as “We need a safe stunt double face,” or “We need to protect a subject’s identity while keeping expression readable.”
- Check for simpler options. Alternate takes, reframes, cutaways, and ADR can reduce or avoid the need for face replacement in many sequences.
- Confirm written permission. Make sure consent matches scope, platforms, duration, and whether reuse is allowed. If the project touches union terms, confirm what counts as a digital replica in your context. FilmDaft’s longer permission guidance is in Consent and Digital Replicas.
- Test source performance compatibility. Review head angles, lighting stability, motion blur, and occlusions from hair, hands, glasses, smoke, or props.
- Set a realism target that fits the shot type. Wide action coverage tolerates more than tight dialogue close-ups. Set your target before post starts so reviews stay consistent.
- Define approvals. Put names on who approves the final face, who can request changes, and what counts as “final” for delivery.
- Decide disclosure and documentation. If context calls for labeling or client disclosure, decide the wording and placement early, then carry it into delivery notes and archives. FilmDaft’s practical disclosure guide is EU AI Act Deepfake Disclosure.
Quality control before delivery
Face issues hide in still frames and show up in motion, especially after color and compression. A steady QC routine helps you catch problems while fixes still fit the schedule. If you are working with generated footage in general and want extra context on motion instability and visual drift, AI Video Generators Explained: Limits and Uses pairs well with this section.
Review in cut context first
Watch the shot in the full scene at real speed. Then scrub frame-by-frame on the hard moments, such as blinks, teeth, and fast head turns. Always compare against the surrounding shots because continuity problems often appear at the cut.
Check eyes, mouth, and blink beats
Eyes are the first place you notice drift. Look for eye jitter, dead blinks, and a gaze that does not match the actor’s intent in the plate. For dialogue, check mouth shapes against the final sound edit, and do this after the sound timing is locked.
Match grain, noise, and compression behavior
Confirm that the face carries the same grain and noise as the plate, including any texture changes from denoise, sharpening, or compression. A face can look fine in a clean VFX export and then break once it hits a streaming encode, so test real deliverables.
Stress-test edges and occlusions
Hairlines, ears, glasses, hands, smoke, and fast props expose edge chatter. Check those areas at 100% scale, and check again after color because contrast changes can reveal seams.
Documentation that keeps you safe later
Good documentation keeps your project stable when edits change, partners ask questions, or a distributor needs clarity. It also protects you when you need to prove what was approved and what assets were used. If you want a structured version of this thinking for professional work, Risk Checklist for Using AI in Client Work is designed as a repeatable checklist.
Consent package
Store signed releases, approval emails, usage limits, and any restrictions in one place. Tie the consent to a shot list or timecode list that matches the current cut so scope stays consistent across versions. If you need refreshers on release language, FilmDaft’s starting points are Talent Release Forms and Model Release Forms.
VFX package
Archive lens data, color pipeline notes, reference captures, and versioned outputs. If you trained or fine-tuned any model for the project, record what data was allowed, where it was stored, and who had access. If you only ran inference on an external tool, record which assets you uploaded and what your retention settings were.
Provenance and disclosure notes
Write down which shots include face replacement or de-aging and why the effect was used. If disclosure is required by context, client agreement, or territory rules, record the exact wording and where it appears (credits, end cards, platform descriptions, or client documentation). If you want a deeper workflow for keeping evidence and metadata attached to files, see Content Credentials and Provenance (C2PA).
Summing Up
Deepfakes and face replacement work best when you treat them as a planned pipeline. The core craft job is matching the plate with correct motion, lighting, and texture, then reviewing the result inside the edit where problems show up. The core responsibility job is consent, scope control, and trust, especially when a real person’s likeness appears in a context that implies authenticity. If you decide early, document clearly, and run consistent QC, you can use face replacement for real production needs without turning the project into a rights and reputational mess.
Read Next: Curious how AI fits into the editing room?
Explore our full AI in Post-Production section to see how AI tools can support editing, audio cleanup, transcription, and visual effects—without replacing your creative judgment.
This section builds on key ideas from our Practical Guide to AI in Filmmaking, which covers where automation helps, where it falls short, and how to stay in control of the final cut.
Also, check out our full guide on AI Tools for Filmmaking to compare models, task types, and how different tools handle writing, editing, color, audio, and animation.
Or step back and explore the broader AI Filmmaking section for insights across pre-production, VFX, animation, and delivery.
