Silhouette 2027 puts editable splines in motion

Body ML tracks limbs, Video Inpaint ML takes the wires out of a coffee shot, and a solved camera carries a painted repair across a castle wall.
Worker in hazmat suit spraying near large yellow industrial tank behind metal security fence outdoors.

For those who don’t know the tool, Silhouette 2027 is Boris FX’s application for rotoscoping, digital paint and compositing. Artists use it to draw shapes around moving subjects, build mattes and paint things out of shots. If the work goes well, the audience has absolutely no idea somebody spent Tuesday putting a wall back together.

My favourite detail in Silhouette 2027 is a foot that gets away. The new Body ML tracker follows a person by a window, breaks the body into articulated parts and gives a roto artist a spline for each one. A second pass tightens the outlines. Down by the shoe, the fit eventually gives out. Here comes the good bit. Delete the offending shape keys and the foot still has its underlying track. Stabilise the viewer on it, redraw the contour and carry on. Somebody has given quite serious thought to what happens five minutes after the impressive demo.

That sets the tone for a substantial release. We get editable roto, a four-stage matte pipeline, local video inpainting and enough 3D to remove a security camera from a Scottish castle. Yes, there is MCP news. For once it can wait until we have dealt with the bloody shoe.

A body of work, in actual splines

Feed the plate into a Roto node, connect the matte to its matte input, then pick Body ML in the Object Tracker and track the frame range. The local model builds a layer and shape for each major body part it detects. An arm can articulate independently of the torso. The artist gets a hierarchy, motion and deformation in one pass. Look at the timeline, though, and that first result can have a key on every frame. Anyone who has tried to adjust an entire stride one ankle key at a time will spot the catch.

Three people with digital motion tracking overlays, showing facial and body pose detection in various activities.

Roto ML takes those shapes and fits them to the matte. It also works with a loose spline you draw yourself, on a human or any other object. Give it a few keys that describe the poses you care about, let it fit the intervening outlines, then check the places where the silhouette changes character. The point of drawing the initial shape yourself is that you get to decide where one ends and the next begins. A machinelearned model can find the outside of a coat quite convincingly and still know remarkably little about which department needs the sleeve.

Two separate simplification operations tackle the dense result. One reduces the number of spline points, the other reduces animation keys. Undo either if a simplification costs you the toe of the shoe. Convert curves to Bezier, B-Splines or X-Splines, clear automatically generated keys, and use per-point keyframing to correct one control point without keying every other point on the outline. There are two kinds of jitter to diagnose, too. Smooth the spline when the contour wriggles. Use the tracker’s Smooth control when the matrix transform from Body ML or Roto ML makes the entire shape twitch.

A giraffe outlined with splines beside its individual point keyframes.

Shape Tracker gives selected shapes another route through a shot. It follows movement and deformation inside their boundaries and can snap their outlines to a propagated or connected matte. For a single intervention, go to the frame that needs attention and press K. Auto Roto Keyframe fits a new key to the edge there. A keyboard shortcut that deals with the frame currently ruining your afternoon qualifies as a modest progress for civilisation.

A fine matte of detail

A spline is a lovely deliverable right up to the moment somebody grows curly hair. The new Matte ML compound arrives with the image and data connections already wired between Mask ML, Matte Assist ML, Matte Refine ML and an optional Matte Deflicker ML. Each stage remains accessible. Inside the compound, separate controls choose which result the viewer shows, which parameters you edit and which stages actually run. You can watch the refined result while you change an earlier selection, or bypass a stage to see exactly what it contributes.

Start in Mask ML. A configurable hover overlay previews the segmentation before you click. The Object tool replaces the rectangle tool and can select a single object or extract its constituent parts at sparse or dense settings. Matte Assist ML then propagates the selection through time. Its model and inference controls let you trade processing time and GPU memory against detail. A perfectly serviceable garbage matte has little to gain from an expensive afternoon at the hairdresser. And vice versa – what does a hairdresser need a garbage matte for. Somewhere in there is a joke about long hair, but it’s stuck.

Young woman with curly hair using laptop outdoors, surrounded by green trees, with partial silhouette overlay.

One particularly good control hides behind the unassuming name Auto-DOD. If the subject occupies a small patch of a large plate, sending the whole image to inference can squander fine detail as the model scales it down and back up. Auto-DOD analyses the shot and creates an upstream domain-of-definition node to restrict the image area sent to the model. The subject gets more of the available inference resolution. Choose a static domain when a downstream node objects to an animated one.

Matte Refine ML adds the Video Matte 2.1 model and Standard and Enhanced quality modes. This is where to spend the extra compute if the binary edge from propagation cuts straight through hair. Play the result at speed before reaching for the fourth stage. Matte Deflicker ML, also available on its own, targets fine frame-to-frame variations that make a stationary boundary appear to boil or jiggle. It starts disabled. Keep its temporal range short on delicate detail, because a longer range can introduce visible transitions of its own. There is an unexpected satisfaction in deciding to leave a button alone.

The disappearing act runs locally

Video Inpaint ML takes an alpha channel marking the area to remove and generates replacement pixels across time. A person in a drone plate, tracking markers on a screen, suspension wires holding coffee beans in mid-air: three quite different clean-ups, one intelligible input. Here is AI worth cheering for. Draw around the thing, run the model on your own machine, inspect the result. No tokens to buy for this operation, no reason to transfer the development budget to OpenAI or Anthropic, and the coffee beans can finally support themselves.

Coffee beans suspended above ceramic cup with green leafy background, left side showing strings, right side edited.

Start by denoising the plate and giving the node a tracked alpha. Generate the current frame first. That buys you a cheap visual test and, on a suitable shot, a clean patch you can track back into place. Change the random seed for a different fill. Increasing Quality gives the model longer to work, although it takes only one lovingly computed wrong answer to appreciate the value of trying a different seed. The second-level joke in this example? Coffee beans are, botanically speaking, seeds, while coffee itself is, chemically speaking, a solution.

So, after an espresso, we go for moving detail or changing illumination, and generate the work range. Processing runs in the background. Use Latest loads the new result, and the Asset Browser lets you switch between versions. Silhouette saves the generated frames in the project folder, so a chosen pass can become a source in the tree (No joke there, coffee fruit grows on shrubs or treelets). Check the fill in motion, finish any stubborn edges with paint and put the grain back. A wire can vanish beautifully while its unusually well-behaved little strip of noise gives the game away.

Paint the castle once

Let’s get into the example from the video: Consider a security camera bolted to the wall at Eilean Donan Castle (I assume you have seen either Highlander, Entrapment and/or Braveheart – you know the castle). It sits on a conveniently solid piece of masonry. The camera filming it moves. To remove the fixture, choose the frame that gives the clearest view of the wall, paint a patch there and project that repair onto a surface in the solved scene. As the shot camera moves, the wall carries the patch through perspective. This is an excellent use of 3D for people who had merely hoped to spend the morning painting out a camera.

In Silhouette 2027, the SynthEyes camera tracker has its own node. Use the solved point cloud to position a card or primitive at the repair, and use Scene Transform to move, rotate or scale the scene or nominate a point as the world origin. The 3D Composite node lets you add geometry, cards and cameras, hide items and reparent them without changing the upstream solve. For roto, a card can stabilise an unprojected viewer. Carry the SynthEyes data port into Roto and import the track into a layer to take its shapes along for the ride. A plane will still only approximate a rounded object. Refine the solve or give the difficult surface its own geometry.

Watch the alpha on the way back to 2D. Passing the full plate alpha through Paint projects an entire opaque face of the proxy, including its imperfectly aligned corners. Switch the footage output to RGB and let Paint supply the alpha for the repair alone. For a sign that straddles wall and ground, make two projection branches. Each projection targets one geometry, so assign one branch to the building and the other to the ground, then composite the two patches together. A small geometry lesson with a fairly large saving in swear words.

Green tracking points and a perspective grid over Eilean Donan Castle.

Projection and texturing also solve different problems. To replace the modern roof, unproject a patch of stone wall into an orthogonal view, paint out the windows and feed the result to 3D Texture. Select the roof primitive and the numbered face that should receive the stone. Mask the chimney and battlements where necessary. The expanded Unwrap node also takes supported primitives and imported meshes into a flat view for painting. Match the operation to the surface and the camera move does much of the remaining work.

Heads, depth and the occasional nostril

Head Track ML graduates to a dedicated node. Its Solid Mesh option fills openings around eyes, nostrils, mouth and ears, giving Unwrap a more continuous surface to flatten. Export the tracked, anchor-adjusted mesh as Alembic when another application needs the head motion. Closing a nostril turns out to be an entirely reasonable pipeline request, not just a wish for the colleague at the next desk.

Depth Map ML adds V3 Faster and Better models. Boris FX credits both with steadier depth over time and lower GPU memory use. The older V2 model also gets a claimed 50 per cent reduction in memory use. That figure refers specifically to V2 memory consumption. A compositor choosing between the models still has a more revealing test than a single attractive depth frame: play the shot and watch whether two overlapping surfaces keep their order.

Woman in floral dress standing indoors, digital code overlay, color filters and effects applied to image.

Write a node before the kettle boils

SeExpr, Open Shading Language (OSL) and WebGPU Shading Language (WGSL) get live image nodes. Write a procedural matte, a gradient, a QC view or a custom image operation directly in the tree. Define parameters in code and they appear as ordinary animatable controls. Adjust the code and evaluate the change without compiling a plug-in or restarting the application. A facility can give artists a control labelled for the job it does, while the person who wrote the maths gets to stop explaining it at every desk.

The expanded Python API supplies more building blocks for integrations, including MCP-driven tools. Package custom nodes, menu actions, key bindings and session templates into an installable extension for the team. Yes, an LLM can help write the shader. The more interesting hand-off happens afterwards, when the next artist installs the extension and gets the same working controls.

There is also a genuinely separate route for anyone who wants a hosted model. The Runway extension brings text-to-image, text-to-video, image-to-video, video-to-video, Act-Two and image and video upscaling into the graph. It requires a Runway developer account and API key. Keep that distinction in mind when budgeting a shot. Choosing a Runway node is a service decision. Running the local body tracker or video inpainting a wire is a workstation decision.

The Background Tasks window shows a render in progress.

Get the render out of the way

The Tasks dock tracks supported jobs including renders, Roto ML, Runway processing, and Paint Rebuild. Completion actions let you bring in a render. The Output node now has an output port of its own. In Passthrough mode it forwards the upstream result. Switch to Rendered Sequence and downstream nodes read the frames on disk. That makes a useful boundary after an expensive branch. If the upstream work changes, render it again before judging the composite against yesterday’s pixels.

Output, matte repair and compositing branches in the Silhouette node graph.

Smaller conveniences add up nicely on a big tree. Shift plus F1 through F8 assigns a node to a viewer shortcut. Press the corresponding function key to return to it. A Compare node helps inspect results, custom playback ranges isolate the troublesome stretch, and individual enable switches let you bypass steps inside compounds. You can add, remove, reorder and publish custom parameters, or change the interface theme. Of these, being able to jump from the final comp straight back to the matte that made the fringe is the one I would want at half past six.

Hosts, licences and the price of admission

Silhouette comes as a standalone application and as a plug-in, with macOS, Windows and Linux builds. Supported hosts include After Effects, Premiere, Nuke, DaVinci Resolve, Fusion and Flame. Subscriptions start at US$545 a year or US$103 a month. Active Silhouette and Boris FX Suite subscribers and customers on an active Silhouette Upgrade & Support plan get the update. Permanent licences are also available. The unrestricted trial runs for 14 days.

Boris FX also lists a 25 percent Black Friday discount from 23 November through 4 December 2026, applied automatically. For a first trial shot, I would dig out the one with the troublesome shoe. Anybody can make a matte look good on the frame they choose for the press release. Being able to fix the frame you choose is a much better reason to upgrade.

For another take on segmentation inside a host, see our Continuum 2027 coverage.

DeveloperBoris FX
ProductSilhouette 2027
PlatformsmacOS, Windows, Linux
Plug-in hostsAfter Effects, Premiere, Nuke, DaVinci Resolve, Fusion, Flame
Local MLBody tracking and video inpainting
External serviceRunway nodes require a Runway developer account and API key
SubscriptionsFrom US$545/year or US$103/month
Trial14 days, unrestricted