The Last Frame: Manual Rotoscoping vs. AI in 2026
Mocha Pro, Silhouette and the new wave of AI tools like Beeble, measured where it hurts: hours per shot.
VFX · Rotoscoping · 9 min

Picture a two-minute interview. One person, one chair, a background the director now hates. In a traditional studio pipeline, isolating that person could take close to 40 hours of roto work. Today, an AI model can hand back a first-pass matte of the same clip before your coffee cools.
So is rotoscoping, one of the most patient crafts in visual effects, about to disappear? Or is it just changing hands?
We put the industry standards, Boris FX Mocha Pro and Boris FX Silhouette, next to the new AI-first tools led by Beeble, and compared them on the thing every producer asks about first: time.
The real price of a hand-drawn matte
Rotoscoping means drawing a shape around a subject and adjusting it on every frame it moves. At 24 frames per second, one second of footage is 24 separate drawings. A 30-second clip at 25fps is 750 frames.

How fast can a professional go? According to FXbuddy’s 2026 guide, a skilled roto artist on a complex full-figure shot typically finishes one to three seconds of footage per hour, even with the best shape-propagation tools. On the hardest shots, with fine hair, see-through fabric or heavy motion blur, Morphic’s glossary puts it at a full day or more per second of footage.

That is why roto has always lived at the bottom of the VFX budget and the top of the schedule. It is precise, skilled, and slow.
The professionals: the tools behind the blockbusters
Mocha Pro 2026: the tracking specialist
Mocha Pro is the planar tracker most compositors learn first. Instead of tracking single points, it tracks flat surfaces, so a roto shape can "ride" a moving object with far fewer keyframes. That alone cuts manual work sharply.
The 2026 release, launched in December 2025, leaned into machine learning. Boris FX added Matte Refine ML, which cleans up the edges of existing mattes on the hard stuff: fine hair, soft blur and fast motion. It joins the earlier AI masking tools Object Brush ML, Matte Assist ML and Face ML. A 2026.5 update in June added AI-based point tracking for shots that used to be untrackable.

Best for: shots with surfaces and structure, screen replacements, cleanup, and artists who want tracking data they can inspect and correct.
Price: subscriptions start at $48/month or $325/year, and Mocha Pro runs as a plugin inside After Effects, Premiere Pro, Nuke, Fusion, Flame, Avid and VEGAS.
Silhouette 2026: the roto artist’s home
If Mocha is the tracker, Silhouette is the paint-and-roto workshop. It has been around for about 20 years and is widely used in film and broadcast VFX, with detailed spline tools, non-destructive paint and deep pipeline support.

The May 2026 release brought AI into the core: tools that detect, segment, propagate and refine mattes, plus AI head, object and point tracking. It even ships an MCP server, so external AI assistants can drive Silhouette’s scripting. A roto supervisor at Untold Studio told Boris FX that the new Point Track ML gives artists steadier tracks with less worry about occlusion.
Best for: feature film and episodic roto, beauty work, paint cleanup, and studios that need every edge to hold up on a cinema screen.
Price: the standalone is $2,195 perpetual or $165/month; the plugin is $1,195 perpetual or $103/month.
The built-in options
Most editors also have roto tools already. Adobe After Effects has Roto Brush, DaVinci Resolve has Magic Mask on its Color page (Resolve 21.1, released at IBC this month, added a Magic Mask 2.0 with better edges), and Nuke offers its own roto and machine-learning toolsets. These are great for fast selections, but pros still reach for Mocha and Silhouette when the shot gets hard.
The AI wave: no splines, no install
Beeble: roto as a cloud service
Beeble pitches itself as an AI VFX platform for filmmakers, and rotoscoping is only one tool in its kit. It also offers relighting (SwitchX), reframing, SDR-to-HDR conversion, VFX pass generation and upscaling, all in the browser. Its client list includes names like Framestore, Skydance, Crafty Apes and Corridor.

For roto, the work happens in Beeble’s Background Remover, and the flow has four steps:
- Upload a video, image sequence or single frame (MP4, ProRes, PNG, JPG).
- Review a keyframe with a rough alpha, and fix mistakes in the Edit Alpha brush before spending GPU time.
- Track every frame on cloud GPUs, which carry the matte through hair, motion blur and partial occlusion.
- Export a grayscale matte, 10-bit RGBA MOV or PNG sequence for Nuke, After Effects or Resolve.
The default model, MatAnyone, pulls a matte from ordinary footage with no green screen. Shot on green? A second model, CorridorKey, handles keying with spill removal. Beeble’s own claim is blunt: a shot that costs a roto artist hours comes back in seconds. It also admits that no pro trusts 100% automation, which is exactly why Edit Alpha exists.

Limits worth knowing: input goes up to 2K and 60fps, with clips up to 3,600 frames on the Professional plan. Pricing starts at $16/month (Creator, billed yearly) and $60/month (Professional), and MatAnyone costs one credit per 30 video frames.
Other AI tools in the race
Beeble is not alone. DaVinci Resolve’s Magic Mask lets you click on a subject and track it, with add/subtract clicks and paint tools for touch-ups. After Effects Roto Brush uses AI propagation inside Adobe’s timeline. And the professional tools are absorbing AI themselves: Silhouette’s Matte Assist ML builds animated mattes from a few inputs, and Mocha’s Object Brush ML does the same inside the tracker.
The line between "pro software" and "AI tool" is blurring fast.
Head to head
Where AI still hands the pencil back
AI roto is fast, but "fast" and "final" are not the same word. Here is where artists still step in:
- Resolution ceilings. A 2K cap is fine for social, music videos and many commercials. A 4K or 6K feature plate needs a pipeline that works at full resolution.
- Temporal consistency. A matte that looks perfect on one frame can shimmer across 100. Edge chatter on hair is still the first thing a supervisor spots.
- Creative intent. AI finds "the person." It does not know that the director wants the sword included but the shadow excluded, or that a strand of hair should be cut on purpose.
- Data and security. Cloud processing means your footage leaves the building. Beeble says it never trains on user data, but big studios often have strict rules about unreleased material.
- Review and fixes. Even Beeble’s own site says no pro trusts full automation. The time saved on drawing partly moves into checking.

The smartest studios are not choosing one side. They let AI pull the first pass, then send the hard frames to Mocha or Silhouette for a human finish.
The clock doesn’t lie: time per stage
Let’s make it concrete. Take one 10-second shot (240 frames) of a person walking, with loose hair and some motion blur. Here is roughly where the hours go in each workflow.
*These are illustrative estimates, not lab results. Real numbers swing widely with footage, resolution and the artist’s experience.*
Look at where the time moves. In manual work, tracking the shape frame by frame eats half the day. With AI, that stage nearly vanishes. What’s left is the part only a human can do: looking at the result and deciding if it’s good enough. In the AI workflow, review and fixes become the biggest slice of the job.
So the real shift isn’t from human to machine. It’s from *drawing* to *judging*.

So… who draws the last frame?
If a matte that took a full working day now takes a coffee break, what happens to the hours we just got back? Do studios deliver the same film faster and cheaper, or do they spend the savings on shots nobody could afford before?
And the harder question: when the machine draws 95% of the frames, is the roto artist disappearing, or becoming the most important quality-control eye in the building?
Your turn. Have you swapped Mocha or Silhouette for an AI tool yet, or are you keeping your splines close? Tell us in the comments which frame you still wouldn’t trust to a machine.