AI by department · VFX and finishing
The most mature adoption in the industry, and the most exposed workforce.
Denoise, matchmove, depth estimation and roto assist were normal practice here before anybody outside VFX used the word. Netflix’s Q2 2026 shareholder letter, July 2026, reports generative AI workflows on roughly 300 titles this year, the largest concentration of that work in post-production. It is the only hard sector number published.
Last checked September 2026.
What is in use
The workflows on shows now.
Task-trained models inside the comp
Hand-roto a few frames, train a small model on them, run it across the shot
Standard
Roto, paint and clean-up at volume
Marker and rig removal, tracker paint-out, garbage mattes. Assistive, not autonomous
Standard
Matchmove, camera solve and depth
Including depth from a single monocular plate, which is now a button
Standard
Denoise, degrain, upres, restoration
Check the licence tier: several plans restrict commercial use above a revenue threshold
Standard
De-ageing and face replacement
The technical question is settled; the consent question belongs to the performer
Consent required
The evidence
The workflow worth learning first.
Train it on your own frames
Nuke’s CopyCat is the opposite of a prompt: you supply original frames and the same frames with your fix applied, and it writes a model you run over the sequence. No third-party training data, no provenance question, nothing unreleased leaving the building. Foundry’s list prices in September 2026: Nuke $3,839 a year, NukeX $5,219, Nuke Studio $6,379. Cattery, its library of pre-trained models, states no licensing terms for individual models.
What it does badly
What it does badly.
- List item text
- List item text
- List item text
- List item text
- List item text
These tools compress volume work and barely touch hero work, which is why the employment effect lands on juniors. In the UK the contraction came first: Bectu’s Big Survey of August 2025 found around half the television workforce out of work and did not identify AI as a cause at all.
Before you agree to anything
The four questions, and the skills that survive this.
Which model produced any supplied asset, not which interface; whether it is indemnified, and on which plan; whether de-ageing or face replacement has documented performer consent; and whether uploading client plates to a cloud service breaches the production’s content security requirements. Nothing on this page is legal advice.
