AI VFX covers two very different things: machine-learning tools that speed up established visual effects tasks such as rotoscoping, cleanup, retiming and upscaling, and generative models that create new imagery. The first is already standard on most shows. The second changes the creative and legal risk of a delivery, depending on who controls the result.
Most debates about AI in VFX happen at the wrong altitude, as policy arguments about whether AI belongs in the industry. The person who actually has to decide is usually a supervisor or post lead asked whether a specific technique can be used on a specific shot this week. That decision is answerable. This guide shows how to make it: what is already normal, where the risk changes, which questions to put to a vendor, and what to write down.
Key takeaways
- Much of what could be called AI VFX, including masking, roto, retiming, upscaling, object removal and voice isolation, is built into standard software and carries no unusual risk.
- The risk profile changes when a generative model produces final footage rather than assisting an existing process.
- The U.S. Copyright Office says a film with AI-generated effects is still copyrightable as a whole, even if the AI elements are not protected on their own.
- Artist-controlled generative work, with human reference, structural inputs and correction at each step, is a different proposition from prompt-and-accept.
- The moment a tool moves from exploration into final footage is when credit and copyright questions attach, so record it.
- A vendor should be able to say which category its work falls into. If it cannot, that is the finding.
Table of contents
- What is AI VFX?
- Which AI tools are already standard in VFX?
- Where the AI VFX risk profile changes
- The three-category test for any shot
- What artist-controlled generative VFX looks like
- Questions to ask a VFX vendor
- Documenting the transition point and the bid
- Cost and schedule reality
- Keeping artists in control of every shot
- Frequently asked questions
What is AI VFX?
AI VFX is any use of machine learning in visual effects and finishing work. It splits into two families that behave very differently on a production.
The first family is assistive. Models help artists do established tasks faster: separating a subject from a background, tracking an object, removing a rig, retiming a shot, enlarging a lower-resolution plate. The artist still decides what the shot should look like; the tool handles a tedious, technical part of getting there.
The second family is generative. Models create new pixels that no camera captured and no artist drew or modeled in the traditional sense: a set extension, a crowd, a creature, a destroyed building, a stylized look. Here the question of who authored the image becomes live, and so does the question of who owns it.
Keeping these families apart is the first step to any sensible policy. Treating them as one thing leads either to banning tools that every show already uses, or to waving through generated shots nobody can account for.
Which AI tools are already standard in VFX?
Before deciding what to permit, be precise about what is already happening on essentially every show. Much of it lives inside the editing, color and compositing tools teams use daily.
Blackmagic Design’s DaVinci Resolve Studio page, for example, lists Neural Engine features including Magic Mask for object isolation and tracking, Super Scale for creating 4K and 8K images from HD sources, speed warp retiming, object removal, a 3D depth map, and voice isolation and dialogue separation. Compositing and editing packages from other vendors ship comparable machine-learning roto and masking tools.
The U.S. Copyright Office’s January 2025 report on copyrightability records the same reality from the industry side. It summarizes Motion Picture Association comments that studios use AI tools in visual effects and post for color correction, detail sharpening and de-blurring, and that AI is frequently used in rotoscoping.
The common thread is that these tasks answer questions with a defined correct answer. There is a right place for a matte edge. Nobody makes an expressive choice when a wire is painted out of a plate. No reasonable policy should prohibit this category, and none of it needs special paperwork beyond normal project records.
Where the AI VFX risk profile changes
The meaningful shift happens when a generative model produces final footage rather than assisting an existing process. At that point, two things change at once.
Creatively, an element now exists that nobody photographed, painted or animated in the usual sense, so the question of who shaped it matters. Commercially, copyright attaches differently depending on how much a human controlled the specific expressive result.
Generative final footage is already on screen. In July 2025, Netflix co-CEO Ted Sarandos said visual effects artists had used AI to create a building collapse in Buenos Aires for the series El Eternauta, and, according to Entrepreneur’s report of the earnings call, that the sequence “was completed 10 times faster” than with standard tools and workflows. He framed it as “real people doing real work with better tools.” That is exactly the kind of claim a buyer will want to see backed by records of who did what.
This is where AI VFX stops being a tooling question and becomes a delivery question. A chain-of-title review does not care which plugin was used. It cares whether the licensed material is owned.
Key insight: The line that matters is not “AI or no AI.” It is “assisting an artist’s shot” versus “generating the shot,” and within generation, whether a person or the model decided what the frame looks like.
The three-category test for any shot
A workable framework separates three forms of AI in VFX rather than treating AI as a single switch. For any element, ask one question: did a person determine the specific expressive result, or did the model?
| Category | VFX examples | Copyright position | Handling on a show |
|---|---|---|---|
| Utility and embedded AI | Roto and masks, tracking, paint-out, retiming, upscaling, depth maps, denoise | Assistive use; does not affect protection | Normal; no special approval |
| Artist-controlled generative (Human Generative Workflows) | Set extensions or stylized passes from custom models trained on the show’s art, conditioned on plates, layouts or rough animation, corrected by hand | Human expression, arrangement and modifications protectable | Approve per sequence; archive inputs and correction rounds |
| Machine generative | Prompted shots, digital performances or environments accepted as generated | Generated material not protected on its own | Escalate to producers and counsel; check consent and rights |
There is reassuring news for most shows in the Copyright Office’s Part 2 report: it states that a film including AI-generated special effects or background artwork is copyrightable, even if the AI effects and artwork are not protected separately. The whole film stays yours. What may not be exclusively yours is an unprotected generated element, which matters if that element is a creature, a design or a sequence you expect to license or reuse.
The same report concludes that prompts alone do not give sufficient control, while human expression visible in the output, and creative selection, arrangement and modification, can be protected. For a supervisor, the practical version is simple: if the honest answer to “who decided this frame?” is the model, treat it as category three.
What artist-controlled generative VFX looks like
Artist-controlled generative work brings specialized models inside professional tools under an artist’s direction. It is closer to a new rendering method than to prompting.
In practice, that means a small set of recognizable habits:
- Reference first. Artists paint or build reference that fixes the look before any model runs.
- Custom training. Small models such as LoRAs are fine-tuned on the show’s own artwork rather than relying only on general models.
- Structural conditioning. ControlNet-style inputs such as depth from the plate, pose, line art or rough animation fix the layout, so the model renders inside it.
- Correction and retraining. Wrong results are painted over or inpainted, and the model is retrained on the corrected frames.
- Traditional finish. Generated elements go through the same compositing, grade and review as any other element.
Google DeepMind’s animated short Dear Upstairs Neighbors shows the approach at small-film scale. Its production write-up describes fine-tuning custom Veo and Imagen models on the production designer’s concept art, animators creating rough animation that the models turned into stylized video, and tools for refining specific regions of a shot with adjustable control. The stated goal was to give artists generative tools without handing their control to the model’s unpredictability.
Key insight: The practices that make generated VFX match the show’s look, custom training, structural inputs and hand correction, are the same practices that create evidence of human authorship. Craft and compliance point the same way.
Questions to ask a VFX vendor
Any vendor offering AI VFX should be able to answer these plainly. Vague answers are themselves an answer.
- Which category does this work fall into? Utility, artist-controlled generative or machine generative, shot by shot.
- What was the model trained on? Your material, the show’s material, licensed data or a general model. Provenance questions come down the chain, so ask first.
- What human inputs constrained the generation? Reference art, plates, depth or pose passes, rough animation, blocked layouts. There should be specifics.
- What does the iteration loop look like? Paint-over and retrain cycles indicate artist control. Accepting the first output does not.
- What will you hand over? Reference, correction passes, training records and workflow files are the evidence that makes a chain-of-title answer possible.
Documenting the transition point and the bid
The most consequential moment on an AI VFX show is the one nobody writes down: when a tool used for look development, previs or a quick test turns into final pixels.
That transition is where consent, credit and copyright begin to apply. Record it: which tool, which shot, what date, which human inputs constrained it, and who approved it. It takes minutes at the time and is close to impossible to reconstruct in a delivery audit months later.
The same thinking belongs in the bid. Useful bid language states which categories of technique the vendor expects to use, whether any generative components will contribute to final footage, what training data those components rely on, and what documentation will be delivered with the work. Agreeing this up front avoids the far worse conversation where a distributor raises it after picture lock.
Cost and schedule reality
Artist-led generative VFX is not the automatic saving it is sometimes assumed to be, and budgeting it that way causes trouble.
A controlled pipeline needs dataset curation, model training, conditioning setup and several correction and retraining cycles. That is skilled labor. What it can change is what is achievable at a given budget, such as shot count, iteration volume or look complexity, rather than removing the people.
The clearest savings sit in the utility category: roto, cleanup, retiming and other repetitive work that was always mechanical. Schedule behaves differently from cost. Once an artist-led pipeline is built, iteration can be much faster, but building it is front-loaded work that does not compress. A show adopting AI VFX for the first time should budget setup honestly. The second show on the same infrastructure is a very different proposition.
Keeping artists in control of every shot
AI VFX is not one decision. It is a series of specific decisions about specific elements, and each one is answerable with the three-category test. Utility tools are already standard. Artist-controlled generative work is valuable and defensible when the control is real and documented. Machine-generated substitution for human creative work is a different category with different consequences.
Two practical supports make that control easier to exercise. When passes pile up, PeakView puts two versions side by side, synced to the frame, so VFX artists, colorists and directors can catch what changed and leave notes at the exact timecode before anything reaches delivery. For teams creating new material, Big Idea offers prompted VFX on your own shot footage, with consistent characters and environments across shots, while creators keep full control. To see how these steps fit into a complete production, read our guide to making a film with AI from script to final render.
