Will AI replace roto artists? Where AI belongs in post-production

AI is already changing roto and cleanup, and anyone who says otherwise isn’t looking. The technical limits are real: AI mattes still break on hair, motion blur and occlusion, so skilled review stays part of the job. But the change is felt first in entry-level work, because that’s where roto and paint have traditionally lived. This post covers what AI does well and badly in roto today, what union agreements and studies say, the entry-level problem, where we think AI belongs, and the skills that keep compounding for artists. It’s our perspective, and we build AI tools for this work, so read it with that in mind.

What AI already does well in roto and cleanup

AI is good at propagating a shape or a selection through frames where the subject is clear, and at repetitive work on large numbers of similar shots. Those are exactly the tasks that made roto slow. The Roto++ paper (SIGGRAPH 2016) cites an experienced artist rotoscoping about 15 frames a day on average, depending on the shot (Li et al.).

Three examples from production and testing:

• Trained on the show’s own data. On Dune: Part Two, Foundry says its CopyCat machine-learning tool was used on the film’s eye shots, and “forty per cent of the 1,000 eye shots completed for the film required no additional touch-ups, saving artists from thousands of hours of roto work” (Foundry). Note the other side of that number: the remaining 60% still needed an artist.
• Assisting the artist’s own shapes. Foundry’s SmartRoto for Nuke (released July 2026) propagates splines the artist draws on key frames, rather than guessing the subject from scratch. In fxguide’s July 2026 coverage, testers worked about twice as fast, with some reaching 4x on complex shots (fxguide). Those are vendor-reported test results.
• One-click masks in editing apps. After Effects Object Matte, Premiere’s Object Mask and Resolve’s Magic Mask now handle simple isolations inside the NLE. For a clear subject against a clean background, that’s often enough. We compare them in AI masking in After Effects, Resolve and Premiere.

Where AI still falls short

The hard parts of roto are still hard for AI, and they’re the parts that decide whether a shot holds up. In Larry Jordan’s April 2026 comparison of AI masking in Final Cut Pro 12.2, Premiere 26.2 and DaVinci Resolve 21, all three struggled with hair: Final Cut left “a ‘halo’ around her hair that makes the mask look fake”, and Premiere showed “obvious edging around her hair”. Each needed manual adjustment afterward (Larry Jordan).

Foundry’s own explanation of SmartRoto makes the same point from the tool-builder’s side: optical flow “breaks down at the edge”, which is exactly where precision matters, and splines stay “subpixel-accurate” (fxguide, July 2026).

Task	What AI tools tend to handle	What still needs an artist
Clear subject, clean background	Propagating the selection through the shot	Checking edges and the frames where it drifts
Hair, fur, motion blur	A soft starting matte	Judging and fixing semi-transparent edges, avoiding halos
Occlusion, crossing subjects	Tracking while the subject stays visible	Keeping identity through the cross; fixing swaps and dropouts
Temporal stability	Frame-to-frame propagation	Spotting chatter and flicker that only shows at speed
Cleanup and paint	Plausible fills for small areas	Grain, perspective, and fills that hold over the whole shot
Delivery	A matte or a mask	Choke, feather, premultiplication, and whether the comp actually works

We cover these limits in more detail in is AI rotoscoping good enough for client work?. The short version: AI removes a lot of the drawing. It doesn’t remove the judgment.

What industry agreements and studies say

The formal picture is uneven, and none of these bodies has endorsed any tool, ours included.

• The Animation Guild and a commissioned study (January 2024). A study by CVL Economics, co-commissioned by The Animation Guild (IATSE Local 839) and others, surveyed 300 entertainment industry leaders in late 2023. It estimated that about 21.4% of film, television and animation jobs, roughly 118,500, would be consolidated, replaced or eliminated by 2026. Of the leaders surveyed, 75% said generative AI had already supported job elimination, reduction or consolidation in their division, and 25% predicted compositors would be affected (Cartoon Brew; TAG). These are forecasts from executives, not measured job losses.
• IATSE’s 2024 Basic Agreement. As reported in July 2024, the agreement keeps employees who operate AI systems covered by the contract, bars producers from using AI prompts in a way that “results in the displacement of any covered employee”, and gives employees consultation on AI use at their request, where production allows. Some VFX and concept design members argued the protections fell well short of what writers and actors won in 2023 (TheWrap).
• SAG-AFTRA’s 2026 agreement. Ratified in June 2026, it requires producers to show “significant additional value” before using synthetic performers in place of live actors (RedShark News, June 2026). It concerns performers, not VFX crews, but it shapes what clients will and won’t allow on screen.

Many VFX artists, especially freelancers and people outside the US, aren’t covered by any of these agreements. For them, the terms are whatever the studio and the client decide.

The entry-level problem, stated plainly

The concern people raise most often isn’t that senior compositors will disappear. It’s that the tasks junior artists used to be hired for, roto and cleanup among them, are the ones AI speeds up first. Fewer of those tasks can mean fewer first jobs, and fewer first jobs means fewer people who ever get the hours to become senior.

Ruoqi Jin’s June 2026 analysis puts it in two halves. The technical argument, that AI can’t yet do a compositor’s job at film resolution and bit depth, “broadly holds”. The employment argument has “clearly broken down” at the entry level (Ruoqi Jin). We think that’s a fair reading.

We don’t have a clean answer to this, and we won’t pretend a tool can solve it. What we can do is be clear about who we build for and what we think the time should be used for.

Where we think AI belongs

Our manifesto asks a different question from “how much can AI do in filmmaking?” It asks which work AI should do to genuinely help the people who make films. Our answer:

• AI does the work behind the shot; people decide. Separating a subject from its background or painting out a rig “directly affects the final quality of a shot, and it demands real expertise”, as the manifesto puts it. AI can take on the repetition. The artist decides whether the edge is right, which version goes forward and what the shot should look like.
• Results are proposals, not finals. The tools we make are designed around review: results come back as layers and mattes an artist can refine, compare and reject, not a baked video to accept as is. Only the pixels you mark are meant to change.
• The time should go to the people doing the work. We sell to freelancers and small studios who do their own roto and cleanup, so the time saved stays with them: more attention on the hero shots, fewer 2 a.m. roto sessions, and shots they’d otherwise have to turn down. We don’t sell headcount reduction, and we won’t market it that way.
• Some lines we don’t cross. We don’t generate performances from nothing, clone voices or create synthetic performers, and we only change a performer’s face or lips (an eyeline, a lip-sync, a stunt double) when they agreed to that use in writing.

Other tool makers state similar principles. Foundry, for example, says it develops AI tools “to augment and accelerate each artist’s work, not to replace their creative judgment” (Foundry). And the open-source keyer CorridorKey was described by Hackaday in March 2026 as taking “one of the most annoying parts of the VFX pipeline” and automating it (Hackaday). That’s the kind of AI we think belongs in post: it takes the tedious part and leaves the decisions with the artist.

What artists can do now

Invest in the skills that make AI output usable, because those keep their value as the tools change.

1. An eye for edges. Knowing what a good hair edge, a correct motion-blur falloff or a clean holdout looks like, and seeing chatter at speed, is what separates a usable matte from a rejected one. If you’re starting out, learn what rotoscoping actually involves by hand first. It trains the eye you’ll use to judge every AI matte.
2. Comp judgment. A matte is only right if the comp works: light wrap, grain, spill, alpha handling. The closer you work to the final image, the harder you are to automate.
3. Supervising AI output. Choosing between versions, knowing when to fix a result and when to redo it by hand, and explaining the call to a supervisor or client. That’s a real skill, and teams are going to need people who have it.
4. Pipeline knowledge. Color spaces, bit depth, formats, conform, and where a model’s 8-bit or 1080p stage breaks a 4K log plate. Many of the AI failures that matter are pipeline failures.
5. Cleanup craft. Clean plates and wire and marker removal still depend on grain, perspective and consistency across the shot. That’s taste, not only technique.

If you run a small team, the same logic applies. Use the time AI frees on the shots that make the work better and on training the people who’ll supervise the next generation of tools.

AI does the work behind the shot. People decide what’s on screen. That’s the future we want to build, and you can read the full reasoning in our manifesto.

Questions

Will AI replace roto artists?

Not in the sense of making the skill unnecessary: AI mattes still fail on hair, motion blur and fast occlusion, and someone has to judge and fix them. But AI is already reducing the amount of manual roto on some productions, and the effect is felt most in entry-level roles, where roto and cleanup have traditionally been the starting point.

Is rotoscoping still worth learning in 2026?

Yes, as a foundation rather than a whole career. Understanding edges, motion blur, holdouts and how a matte behaves in a comp is what lets you judge and fix AI output. Pair it with compositing, supervision of AI-assisted work and pipeline knowledge, which get more valuable as the tools change.

What do union agreements say about AI in VFX?

IATSE's 2024 Basic Agreement, as reported in July 2024, keeps employees who operate AI systems covered by the agreement, says producers can't use AI prompts in a way that displaces covered employees, and provides consultation on request. Some VFX and design members said the protections were too weak. Coverage varies a lot, and many VFX artists work outside union agreements.

Sources

1. Roto++: Accelerating Professional Rotoscoping using Shape Manifolds, by Li et al. (SIGGRAPH 2016) · accessed 2026-09-24
2. The untapped potential of machine learning in VFX, Foundry (CopyCat on Dune: Part Two) · accessed 2026-09-24
3. Foundry's SmartRoto: AI-assisted roto that aims to work the way you do, fxguide (July 16, 2026) · accessed 2026-09-24
4. Compare AI-assisted masking in Final Cut, Premiere and Resolve, by Larry Jordan (April 29, 2026) · accessed 2026-09-24
5. AI principles, Foundry · accessed 2026-09-24
6. Union study says generative AI will disrupt 204,000 jobs in three years, Cartoon Brew (January 31, 2024) · accessed 2026-09-24
7. AI and animation, The Animation Guild · accessed 2026-09-24
8. IATSE members oppose contract over AI protections, TheWrap (July 15, 2024) · accessed 2026-09-24
9. SAG-AFTRA's AI deal and synthetic actors, RedShark News (June 8, 2026) · accessed 2026-09-24
10. Will compositors lose their jobs? (Ruoqi Jin, June 7, 2026) · accessed 2026-09-24
11. CorridorKey is what you get when artists make AI tools, Hackaday (March 18, 2026) · accessed 2026-09-24

Written by the nolanlabs team. nolanlabs does AI shot work for post-production; product names mentioned belong to their owners. Tool details were accurate on the date shown above. Check the vendor’s documentation before relying on them.