Nexus is designed to work without an analysis pass, a tracking session or a single keyframe. Read this page and you can use it. Everything after it is refinement and reference.

1Apply Nx Nexus. Drop it onto your clip. It lives under Nx Color in every host.

2Choose a Detection Mode. Start with Smart Detect for people, vehicles and objects. If the thing you want isn’t a recognisable object, see the mode table overleaf.

3Let it play. Nexus reads the shot as it goes, finds your subjects and gives each one a stable identity. No pre-analysis, no waiting for a solve.

4Choose what to show. Boxes and labels are on from the start. Pick a frame style, and switch on centre markers and connections when the look calls for them.

5Pick a subject. Click one in the viewer, or type the number from its label into Subject 1. Region FX, Subject Glow and matte exports then apply to that subject alone.

For most shots, you can stop here

Two settings are worth knowing beyond that. Revive Frames (in Tracking) decides whether an identity survives a subject going briefly out of sight — it starts at 0, and around 15 is a much better default for real footage. Blur Strength (in Rack Focus) has to be above zero for a focus pull to do anything at all.

How to read this manual

Parts I and II explain what Nexus is and what it can do. Part III covers install and licensing. Part IV is the full control reference — every control, its range, its default and what it depends on.

Passages marked “How it works”

Some sections carry a HOW IT WORKS marker. Those explain the machinery underneath, for readers who want it. Skip every one of them and nothing about using Nexus gets harder.

Part I

The System

What Nx Nexus is, and what changes the moment you drop it on a clip.

The Idea

What Nexus Actually Is

Every tracking shot you’ve admired — the reticle that never slips, the label that follows a running subject through a crowd, the interface that seems to understand the frame — was traditionally built by hand. Track a point. Fix the drift. Parent a box. Redo it all when the edit changes.

Nx Nexus removes the hand. It is a subject-tracking engine: something that finds what matters in a frame, gives it an identity, holds that identity through the shot, works out where it sits in space, and then lets you do something with it. Overlays, mattes, focus, effects, exported channels — all of it downstream of one idea.

Find anything. Track anything. Make it do something.

The part people notice first is the AI. A neural detector reads every frame and recognises what’s in it — people, vehicles, animals, eighty kinds of everyday object, each one named. But the detector is an input, not the product. Nexus doesn’t assume that everything worth tracking is a person or an object.

Sometimes the thing you want is a face. Sometimes it’s a car. Sometimes it’s a light, a colour, a bright reflection, a logo, a particle, or an abstract shape with no name at all. So Nexus has several ways of finding a subject — neural object detection, dedicated face detection, and six classical computer-vision detectors that lock onto brightness, edges, colour, a picked hue, or local bright and dark spots. Once found, every one of them enters the same tracking system, gets the same stable ID, and reaches the same overlays, the same depth, the same mattes and the same effects.

One architecture: find a subject however the shot allows, then everything downstream is the same.

Downstream, it doesn’t stop at overlays. Nexus estimates a depth map of the scene from a single camera, so it knows which subject is in front of which. It can cut an AI matte around a subject — hair, fingers, the gap under an arm — and use it to blur a face, glow an outline, or hand a clean alpha channel to the rest of your comp. It can pull focus onto a moving actor with a smoothness no hand-keyframed blur will match.

Local AI

No cloud. No uploads. No subscription inference. Every model ships inside the plug-in and every frame is analysed on your own machine. There is no service to sign into, nothing to download on first run, and no footage, frames or metadata ever leave your computer. On client work under NDA, that is not a convenience — it is the requirement.

What It Feels Like

You drop the effect on a clip and the frame comes alive. Corner brackets snap onto every person in the shot, each with an ID and a name — PERSON · ID 3 — and the IDs hold. Someone walks behind a pillar and emerges still ID 3. The camera pans and the boxes stay glued to their subjects instead of smearing across the frame.

Then you start playing. Connection lines web the subjects together and route themselves like circuit traces. Labels pick up speed and depth readouts. A cyan rim traces the exact silhouette of one actor. Everything behind them hazes into atmospheric fog. A rack focus follows your lead through the shot. It feels less like applying an effect and more like switching on an instrument that was already watching.

The Nexus HUD running on footage — Ten subjects in a street crowd. The labels for the group behind the lead have stacked themselves apart, and that arrangement holds from frame to frame.

Part II

The Engine

How the frame is read, how identity survives, and everything the plug-in can do with what it finds.

The Pipeline

Detect. Track. Render.

Every frame travels a three-stage pipeline. The analysis runs once per frame — not once per pixel — so the heavy thinking happens exactly once, and the GPU’s only job is painting the interface. That split is why styling changes feel instant even on 4K.

Detect

Stage one

The frame is read once and handed to the detector you’ve chosen. Smart Detect runs a neural object detector that recognises eighty classes of real-world object and names each one. Face Detect runs a dedicated, very fast face model. Six classical modes analyse luminance, edges, saturation, a chosen colour, or locally bright and dark spots — with hysteresis thresholding, so regions that sit right on the threshold don’t flicker in and out. The output is the same either way: a list of scored, classified boxes.

Track

Stage two

Raw detections become subjects. A ByteTrack-style multi-object tracker runs a motion-prediction filter per subject — on position and box size — then matches predictions to new detections optimally rather than greedily. Identity follows the box, not the name the detector gives it, so a boat the detector briefly calls a truck is still the same subject. Global camera motion is estimated and subtracted first, so a pan doesn’t read as sixty objects sprinting sideways. Up to 64 subjects hold IDs at once.

Render

Stage three

The tracked subjects are handed to a single-pass GPU renderer that draws everything you see: seven frame styles, ten marker shapes, six connection topologies with six line routings and five arrowheads, labels on rounded legibility plates that automatically step out of each other’s way in crowds — every layer anti-aliased and composited in order, with a thin dark outline that keeps light overlays readable on bright footage and an optional glow — plus Region FX, Subject Glow and depth fog composited underneath. Because the drawing is cheap, styling changes are instant.

Fast by design

Detection and depth run concurrently, so switching depth on costs far less than it sounds like it should. Everything the AI needs ships inside the plug-in — there is nothing to download, sign in to, or configure. Curious where the time goes? Turn on Show Performance Stats for a live readout of detection, depth and matte time, the AI backend in use, and the live tracker count.

Detection

Eight Ways of Seeing

Nexus doesn’t assume that everything worth tracking is a person or an object. Sometimes the thing you want is a face. Sometimes it’s a car. Sometimes it’s a light, a colour, a bright reflection, a logo, a particle, or an abstract shape.

So there are several ways of finding a subject — from neural detection to classical computer vision. Once found, they all enter the same tracking system. One control, Detection Mode, decides which one is used; everything downstream behaves identically whichever you pick.

Which mode should I use?

I want to track…	Use
People, cars, animals, everyday objects	Smart Detect (AI)
Faces	Face Detect (AI)
Lights, highlights, stars, particles	Bright Spots
Dark objects against a bright background	Dark Spots
A specific colour — a marker, a costume, a brand	Select Color
A textured or highly detailed object	Edges
Bright or dark regions generally	Luminance
Vividly coloured regions	Saturation

The eight modes in full

Mode	Locks onto	Notes
Smart Detect (AI)	80 classes of real object, each named and scored	The flagship mode. Anything with people, vehicles, animals or props. Detections carry a class name, which becomes the subject’s label and follows its whole history rather than the latest frame.
Face Detect (AI)	Faces	A tiny, very fast dedicated model — ready almost the moment you select it. Interviews, crowds, privacy work. Faces carry no class name, so the Object Name label prints nothing for them.
Luminance	Bright regions — or dark, with Invert	Practicals, headlights, hot spots, silhouettes. Effectively free.
Edges	High-detail, high-contrast regions	Textured or graphic subjects, motion-graphics plates. Measured on the analysis grid, so the response doesn’t change with source resolution.
Saturation	Vividly coloured regions	Costumes, signage, sports kit. Deliberately gated by brightness so shadow noise never registers as colour.
Select Color	Regions matching a colour you pick	Green screens, coloured markers, brand colours. In this mode Threshold becomes the colour tolerance — and in Resolve the control relabels itself to say so.
Bright Spots	Features brighter than their surroundings	Star fields, LEDs, specular hits, particles. Finds local highlights regardless of the overall exposure of the frame.
Dark Spots	Features darker than their surroundings	Dark objects on bright ground, dust, sensor spots, birds against sky.

Smart Detect naming a person, a dog, a skateboard and two cars

Face Detect on three people, with Visualize Segment Masks tinting each face

Narrowing what Smart Detect looks for

With Smart Detect running, Class Filter restricts the search to one category. Filtering happens before tracking, so excluded objects never become subjects and never consume an ID or a tracker slot.

People

People. Plus People + Vehicles and People + Animals as ready-made unions.

Vehicles

Bicycle, car, motorcycle, aeroplane, bus, train, truck, boat.

Animals

Bird, cat, dog, horse, sheep, cow, elephant, bear, zebra, giraffe.

Sports Equipment

Ball, racket, bat, glove, skis, snowboard, skateboard, surfboard, frisbee, kite.

Food & Drink

Bottles, glasses, tableware and prepared food.

Electronics

TV, laptop, mouse, remote, keyboard, phone.

Furniture

Chair, couch, potted plant, bed, dining table, toilet.

Bags & Accessories

Backpack, umbrella, handbag, tie, suitcase.

Street Objects

Traffic light, fire hydrant, stop sign, parking meter, bench.

Some household objects the detector knows — clocks, books, vases, kitchen appliances — don’t belong to any filter group. Leave Class Filter on All Objects to reach them.

First run on a new machine

The first time an AI mode runs on a given computer, Nexus prepares the neural model for your GPU in the background. This happens once per machine — one shared, prepared copy serves every host you have installed — and is kept until you delete it (see Disk space under System requirements). Nothing freezes: the plug-in renders without overlays until the model is ready, usually within seconds to a minute, and the viewer refreshes on its own the moment it is. Face Detect uses a tiny model and is effectively instant. Once the detector is up, Nexus quietly prepares the depth, face and matte models one after another while you work, so they are normally ready before you reach for them.

The panel keeps up

In DaVinci Resolve and After Effects, controls that don’t apply to the current mode grey themselves out, so the panel always shows you what matters right now. Premiere Pro and Final Cut show every control as active — their plug-in interfaces have no way to grey things — so use the group structure and the label hints as your guide there.

Tracking

IDs That Hold

Detection tells you what’s in this frame. Tracking tells you it’s the same one as last frame — and that second claim is the hard one.

In practice there are only two things to set. Temporal Stability is the master switch and is on by default. Revive Frames decides how long an identity survives while its subject can’t be seen, and it starts at 0 — raise it. Everything else has a sensible default and a job it does quietly.

How it works · the five mechanisms

Five things run under that single switch. You never configure them, but knowing they exist explains why the IDs behave the way they do.

Prediction

Motion filter

Each subject carries a motion model that predicts where it will be next frame — position and box size. Detections are matched against predictions, not against last frame’s pixels, so fast movers and growers (someone walking toward camera) stay locked. Confident detections correct the model strongly; shaky ones barely nudge it. The size filter is deliberately asymmetric: it grows fast and shrinks slowly, so a box never crops a limb as a subject approaches.

Association

Staged, optimal

High-confidence detections are matched first; leftover subjects then get a second chance against faint ones — which is how someone half-hidden behind another person keeps their ID instead of being dropped. Faint detections can extend an existing subject but can never spawn a new one, so partial occlusion doesn’t breed ghosts. The matching itself is solved optimally, not greedily, and it goes by the box rather than the class name: where boxes alone can’t tell two subjects apart, the colours inside them break the tie.

Camera compensation

Automatic

The global camera translation between frames is estimated and subtracted before any matching happens. Pans, tilts and handheld shake stop reading as object motion, and IDs survive camera moves that would churn a naive tracker completely. There is nothing to configure; on flat or ambiguous frames where no reliable estimate exists, the plug-in reports no motion rather than guessing.

Occlusion coasting

Revive + Ghost

When a subject vanishes behind something, its track can coast on its motion prediction for as many frames as Revive Frames allows, then reacquire on the far side with the same ID. Turn on Show Occluded (Ghost) to keep drawing the predicted box at reduced opacity while it coasts — far better on screen than a box that blinks out.

A sticky top list

Max Trackers

When a scene holds more objects than Max Trackers allows, the biggest subjects win the slots — but a subject that was on screen last frame gets a substantial head start. A newcomer must be decisively bigger to displace it, and a subject that is briefly hidden keeps its place until it returns. Without that stickiness, labels flicker on and off at the boundary in any crowded shot.

Read this before you file a bug about IDs

Revive Frames starts at 0. At zero, a subject that goes undetected for a single frame is dropped outright and comes back with a new ID. Occlusion tolerance is opt-in, and it is the single most useful control in this group: set it to 10–20 for handheld and crowd work, higher for shots with long, hard occlusions. From 2 upward it also carries a box, at full strength, through a frame or two that the detector simply misses. If you’ve ever thought “my IDs keep changing,” this is almost always why.

Two dials for calm

Temporal Smoothing trades responsiveness for steadiness — raise it and boxes glide, lower it and they snap. Confirm Frames debounces: a subject must persist that many frames before it is drawn at all, which silences one-frame false positives entirely.

One dial for reach

Max Move Distance caps how far a subject may travel between frames and still keep its ID. Leave it at 0 and Nexus sizes the gate automatically from each subject’s own dimensions — the right answer almost always. Raise it only for genuinely fast, small subjects.

Stable IDs through an occlusion — Nine subjects, nine identities — people, two cars, a handbag, two traffic lights — each one held from frame to frame while the crowd crosses behind the lead. The number on the label is the one every other control refers to.

How it works · Determinism

The Same Every Time

A tracker is, by nature, a thing that remembers. Feed it frames out of order and a naive one gives you different answers every pass — different IDs on scrub, different IDs on export, different IDs on the machine next to yours. Editing hosts do exactly that, constantly: they scrub, they play backwards, they re-render a paused frame after every slider nudge, they split a render across threads.

Nexus solves it by keeping a rolling per-frame record instead of a single “current” state. Every frame it renders, it stores the tracker state, the camera-motion reference and the published result for that exact frame number. When you ask for a new frame, it seeds from the frame before it — not from whatever ran most recently. When you ask for a frame it has already produced, it hands back the stored answer verbatim.

Scrubbing back is cheap

Dragging back over frames you’ve already played re-uses the stored tracking result: the detector doesn’t run again, and the IDs can’t change.

Order doesn’t matter

Reverse playback, out-of-order export and multi-threaded rendering reproduce the IDs a straight playthrough would give.

Focus pulls and mattes too

Depth levels, the automatic focus pull and a matte’s frame-to-frame checks ride the same record, so focus moves and cutouts reproduce too.

Tweaking is safe

Changing a styling control never re-runs analysis. Changing a detection control re-detects, but identities carry across rather than resetting.

If you jump before you play

A frame you jump straight to has nothing before it in the record. In After Effects, Nexus fetches the frames leading up to it and tracks them first, so the frame you land on is the frame a playthrough gives; a long jump takes a few seconds the first time. Resolve does the same over a run-up of about a dozen frames. A longer jump there, and any jump in Premiere Pro, Final Cut Pro or Motion, where Nexus works only from the frames it is handed, starts a fresh chain that may briefly show different ID numbers than a full playthrough will.

A fresh chain repairs itself: play through the shot and each frame recomputes once from the correct chain as the playhead passes over it. So outside After Effects — play the shot through once before you judge the IDs. The rolling record holds roughly 600 frames (about 20 seconds at 30 fps); beyond that window the oldest frames retire.

The Third Dimension

Depth, Fog & Rack Focus

Nexus estimates a full depth map of the frame from a single camera — no stereo rig, no LiDAR — then reads a robust depth value for every tracked subject. It samples the inside of each box densely, weighted toward the middle, and takes the dominant surface it finds there, so the backdrop between spread arms, the floor at someone’s feet or a neighbour at the edge of the box can’t drag the reading off. The flat list of subjects becomes a scene with front and back, and six things become possible.

Colour by depth

Boxes and markers tint by distance — nearest warm, farthest cool — stretched across the tracked group, so the spread reads clearly even in a shallow scene.

Occlude by depth

A depth-aware 3D HUD: wherever a nearer object covers a farther one, the farther subject’s box, label and connectors are hidden behind it. Overlays stop floating on the image and start living in it.

3D connections

With 3D Tilt, links taper thin and dim toward the farther subject and bow into the scene. 3D Connection Depth makes the proximity, nearest-neighbour and spanning-tree graphs link true 3D neighbours instead of screen-space ones.

Depth fog

Edge-aware atmospheric haze graded by distance. Set where it starts and where it saturates, pick a colour, and the background recedes — cheap separation that would otherwise need a rotoscope.

Rack focus

A real depth-of-field with bokeh discs and depth-aware sampling, so a sharp foreground doesn’t halo into a blurred background and a blurred near foreground spills softly over what’s behind it. Follow a tracker ID, the nearest subject, the farthest, or set the distance by hand.

Depth as a matte

Inside each box, keep only the pixels sitting at that subject’s distance. Hair, smoke and held objects survive; the background behind them doesn’t. Fast, and it works while the outline AI is still loading.

Enable Depth is not the master switch you think it is

Enable Depth governs the features in the Depth group itself — fog, occlusion, 3D tilt, the Z label, and colour-by-depth. Four other things are their own statement of intent and switch the depth engine on by themselves: Visualize Depth Map, an Output Mode set to Depth, a Matte Source of Depth Silhouette while something is actually consuming the matte, and a Focus mode with Blur Strength above zero. You never need to hunt for a checkbox to make those work.

The corollary matters too: picking a Focus mode and leaving Blur Strength at 0 does nothing at all. A focus mode with no blur has nothing to express.

Depth map of two people walking out of a colonnade — Visualize Depth Map on two shots: near is white, far is black. Left, two figures step out from a colonnade; right, the ground falls away toward a treeline. Every subject takes its distance from this map, and fog and rack focus read straight from it.

Depth map of a hiker in a field, the ground falling away to the treeline — Visualize Depth Map on two shots: near is white, far is black. Left, two figures step out from a colonnade; right, the ground falls away toward a treeline. Every subject takes its distance from this map, and fog and rack focus read straight from it.

What this depth is, honestly

It is relative depth inferred from a single image — not a measurement. Values are normalised to the scene with 1 meaning nearest, and the range is carried from frame to frame: it follows a real change in the scene only gradually — about a tenth of the range per second — so levels don’t pump when a subject enters or leaves frame, and it starts afresh at a cut. It is excellent for separation, focus, fog and mattes. It is not a survey instrument, and an exported depth pass should be treated as a relative depth channel, not metric distance.

Subject Mattes

Outlines, Not Rectangles

A tracking box tells you where a subject is. A matte tells you which pixels are the subject. Nexus can cut one automatically — the actual silhouette, with hair, fingers and the gap under an arm — using the tracker’s own boxes as prompts to a segmentation model. No roto, no keyframes, and it follows the subject for the whole clip.

Matte Source, in the Subject Mask group, decides how tightly the subject region hugs each subject. It shapes both Region FX and the matte you export from the Output group; Subject Glow always follows the AI outline.

If a matte source isn’t ready — the AI still loading, or no depth in a shot — Nexus quietly steps down the chain rather than rendering nothing.

Visualize Segment Masks: every subject in a street scene tinted in its own colour — Diagnostics › Visualize Segment Masks, each subject’s outline tinted in its own colour. Hair, a cyclist astride a bicycle, a basket held in two hands — these are the silhouettes Region FX, Subject Glow and the Output group all cut with.

Visualize Segment Masks on a woman with a bicycle basket and the crowd behind her — Diagnostics › Visualize Segment Masks, each subject’s outline tinted in its own colour. Hair, a cyclist astride a bicycle, a basket held in two hands — these are the silhouettes Region FX, Subject Glow and the Output group all cut with.

Matte Quality

Segmentation runs in one of two modes. High (Per-Subject), the default, zooms the AI into each subject’s neighbourhood and analyses it there — so edge detail scales with how much of the frame your subject occupies, and a quarter-frame actor gets roughly four times the edge resolution of a whole-frame pass. Standard analyses the whole frame once. It is meaningfully faster and perfectly good for censoring, spotlighting and glow.

Switching High → Standard is the single biggest speed lever once segmentation is on.

Shaping the edge

The AI matte is solid inside the outline and soft only along it. Feather and Grow / Shrink work on that edge.

Matte Grow / Shrink

−20 to +20 px

A true expand and choke, not a blur. Push it positive to swallow a stubborn edge fringe; pull it negative to pull the matte inside the subject and avoid a halo.

Matte Feather

0 to 1

Widens the soft ramp at the edge. The place to reach when hair or motion blur is shimmering frame to frame — a little feather buys a lot of stability.

Matte Range

Depth Silhouette only

How wide a band of distance around the subject counts as “the subject.” Tighter cuts closer; wider keeps outstretched arms and held objects.

Being straight about it

Each frame’s matte is cut from that frame’s own pixels and then checked against the sixteen frames before it. A part of the subject the model suddenly wavers on — a fan held in a hand, a strand of hair — is held where it was for up to sixteen frames, and a short run of frames the model gets plainly wrong is replaced by the last good one. Nothing is averaged across frames, so a moving subject leaves no trail behind it. On difficult material you may still see a little edge shimmer, and a small part can drop out for a single frame. Matte Feather is the fix for the shimmer. This is an outstanding automatic matte; it is not a substitute for hand roto on a hero VFX shot.

Subject Glow

A neon rim and halo that hugs the segmented outline and screen-blends over the image, so it reads as emitted light rather than a drawn stroke. Set intensity, width and colour; point it at specific subjects with the Subject Mask selection; let it pulse with Animate Overlays. Glow always uses the segmentation outline, whatever Matte Source is set to — it’s the one feature that needs the outline to look right.

The first time you use a matte

The segmentation model is the largest of the AI models and takes the longest to prepare — the one-time preparation on a new machine is the longest wait in Nexus, and each later launch of a host needs a few seconds to bring it back. Nexus starts that work in the background soon after the effect first renders, so it is usually done before you ever reach for a matte control. If you get there first, you’ll see MATTE AI PREPARING... in the top-left corner, the matte falls back to depth or box in the meantime, and the picture updates by itself the moment the model is ready. Show Performance Stats keeps a running clock on it if you want one.

Region FX

Effects That Follow

Because Nexus always knows where its subjects are — and, with segmentation on, exactly what shape they are — it can treat the pixels inside them, or everything outside them, with no masks, no roto and no second tracking pass. One menu, twenty effects, applied underneath the overlays so boxes and labels stay crisp on top.

Tracker Zoom

Magnifies the content inside the box, anchored on its centre. Always rectangular by design.

Spotlight (Dim BG)

Dims everything outside the subjects. Instant attention.

Blur

A softer privacy option, or a way to push a background back.

Pixelate (Mosaic)

Censor blocks that scale with FX Amount. The standard privacy tool.

Solid Fill

Flat fill in the Tracker colour. Redaction bars on boxes; a clean rotoscoped silhouette with the Segmentation matte. Amount = opacity.

Sharpen

Unsharp mask on the subject only — the “make them pop” control. Amount = strength.

Black & White

Desaturates by FX Amount — colour-pop the subject, or drain it.

Invert

Negative image, blended by FX Amount.

Colorama (Hue Shift)

Rotates hue while preserving luminance.

Posterize (Cel Shade)

Quantises brightness into flat tones while keeping the colours. Amount = fewer levels.

Thermal

Heat-camera palette: black through violet and orange to pale yellow. Amount = mix.

Night Vision

Green image intensifier with scanlines and sensor grain.

Halftone

Newspaper dot screen at 45°: bigger dots in the shadows, on warm paper. Amount = cell size.

Ink Sketch

A pencil drawing on paper: contour lines, and hatching that deepens to cross-hatching in the shadows. Amount = line weight and shading.

Dither (Retro)

Ordered 4-level dither in a hand-held-console green palette. Amount = pixel size.

ASCII Mosaic

Terminal look: cells drawn as glyphs chosen by brightness, in the cell’s own colour. Amount = cell size.

Hologram

The picture as a projection made of light: blue-to-cyan tone map with lifted blacks and bloom, thin scanlines, a slow rolling band, flicker and rare row tears. Reads best in motion.

Chromatic Ghost

RGB split with a faint double-exposure offset.

Glitch

Row-wise datamosh with an RGB channel offset.

Heat Ripple

Heat-shimmer displacement over the subject; animates with Animate / Speed. Amount = amplitude.

Every effect works on the subjects selected in the Subject Mask group. Apply To flips the whole thing inside out: Inside Subjects treats them, Outside Subjects treats the background. And because the region is whatever Matte Source says it is, switching to Segmentation makes every one of these effects follow the subject’s actual outline instead of its rectangle.

Two effects deliberately ignore that: Tracker Zoom keeps rectangle semantics (a silhouette-shaped magnifier looks wrong), and Spotlight always dims the outside.

The animated effects — Night Vision grain, Glitch, Hologram and Heat Ripple — are driven by the frame number rather than a random seed, so they render identically every pass, and identically whether the frame was drawn by the GPU or on the CPU. No boiling on re-export.

Thermal’s sensor noise and the Ink Sketch hatching and paper grain don’t move at all. Like the Halftone screen they are fixed to the frame, so a still subject stays still.

Effects stay clean right up to the frame border: a subject touching the edge of the picture shows no dark fringe from Zoom, Blur, Glitch or Chromatic Ghost.

The one-minute face censor

Detection Mode Face Detect → Region Effect Pixelate → Apply To Inside Subjects → switch off Boxes and Labels. Every face in the shot is found, tracked and censored for the whole clip, hands-free. Add Revive Frames 15 so a head-turn doesn’t drop the censor for a frame.

Pixelate following the segmentation outline of a woman by a lake

Dither (Retro) applied to two subjects, the room behind them left alone

The Look

Anatomy of the Overlay

Four independent layers stack into the HUD. Each has its own group in the panel, its own colour and its own switch — use one layer or all four. Boxes and labels are on when you apply the effect; centre markers and connections wait for their switch. Overlay sizes given in pixels — line widths, marker size, dash length, arrowheads, glow width — are measured at HD and scale automatically with the frame, so a look you build on a 1080p proxy survives the move to 4K.

Tracker Frames

7 styles

Corner brackets, rounded rectangles, targeting circles, hexagons, diamonds, reticles and target rings, with control over line width, bracket length, circle gap and opacity. Every stroke is anti-aliased and the layers are composited in order, each with its own coverage: a one-pixel line is a clean one-pixel line, boxes that overlap blend where they cross instead of one recolouring the other, and opacity is a real per-layer setting. Box Color Mode switches between one fixed colour, a distinct colour per object class, or a warm-to-cool tint by distance. With Animation on, the circular styles spin.

Center Markers

10 styles

A precise mark at each subject’s centre — dots, plus, cross, crosshair, ring, target, dot-and-ring, diamond, square, triangle. Markers pulse gently when Animation is on. The Velocity Arrow lives here too: a real motion vector drawn from the centre of anything that’s moving.

Connections

6 topologies × 6 routings

Lines that web the subjects together: sequential chains, star, full mesh, proximity webs, nearest-neighbour graphs, or a minimum spanning tree. Route them straight, curved, as a wave, as a circuit-board trace with a 45° dogleg, or as right-angled runs with sharp or filleted corners. Solid, dashed or dotted; five arrowhead styles that ride the actual route. Dashes march when Animation is on.

Labels

Anti-overlap

Object name, ID, coordinates, speed, depth and confidence — any mix, on a dark, rounded legibility plate that keeps text readable over any footage; the text sits cleanly on the plate, with no halo at any plate or text opacity. When labels collide in a crowd, a layout pass stacks them apart automatically and holds that arrangement steady between frames so they don’t jitter between rows. A label sits just above its box, clear of the brackets. Where there is no room it goes below, and on a box that fills the height of the frame it sits inside, along the bottom edge; having moved, it stays put until there has been room above for a dozen frames, so it doesn’t hop while a subject bobs at the frame edge. The whole plate always stays inside the picture. There is nothing to tune.

Outline and glow

Two finishes apply to all four layers at once, from the Overlay Style group. Overlay Outline draws a thin dark edge under every stroke and every glyph and around the label plate, so a white bracket stays readable against a white wall. It is on by default, at 0.5: higher is wider and darker, and 0 switches it off. Overlay Glow adds light around boxes, markers, arrows and connections, screen-blended onto the picture underneath the overlays. Each overlay glows in its own colour unless you choose one with Overlay Glow Color, and Overlay Glow Width sets how far the light reaches.

Labels deliberately take no glow: the text stays crisp, and the plate hides whatever glow falls behind it. The top-left status messages and the performance readout take neither finish.

Reading a label

Whatever you switch on appears in a fixed order, on one line:

The ID is the number you type into Subject 1–4 or Focus Tracker ID. Z runs 0–100 with 100 nearest.

Rounded Rectangle frames on children sitting on a tractor — Two of the seven frame styles and two of the six line routings. Every combination is a menu choice, and none of it costs a re-analysis.

Corner Bracket frames and Curved connections on a herd of elephants — Two of the seven frame styles and two of the six line routings. Every combination is a menu choice, and none of it costs a re-analysis.

Targeting

Choosing Your Subject

By default, Region FX, Subject Glow and matte export apply to every subject Nexus is tracking. Very often you want one person, not all of them. That’s what the Subject Mask group is for — and the fastest way to use it is to click the subject in the viewer.

Click-to-select

Click a subject → it becomes the only selected subject.

Click another → it’s added — up to four subjects at once.

Click a selected subject again → it’s removed. Remove the last one and nothing is selected.

Click empty background → the selection clears.

No modifier keys, no drag. If subjects overlap, the smallest box containing your click wins — so you can pick a face inside a body. Selected subjects get a soft white halo and everyone else’s overlays dim, so you can see the selection at a glance. The numbers it sets are exactly the ones the labels show.

You can always set it by hand. Subject 1 at -1 means all subjects, 0 means none, and a number selects that one subject. Subject 2, 3 and 4 add more, and 0 means unused. Read the numbers off the labels; they stay put as long as Temporal Stability is on. Select All Subjects and Clear Selection do exactly what they say. With nothing selected, the subject effects simply wait until you pick someone.

Clicking works in DaVinci Resolve, After Effects, Final Cut Pro and Motion. In Resolve, the OpenFX overlay must be visible in the viewer. In After Effects, select Nx Nexus in the Effect Controls panel, then click in the Composition panel with the Selection tool. In Final Cut Pro, select the clip and click Nx Nexus in the Video inspector, then click in the viewer; in Motion, select the layer and the filter. Premiere Pro doesn’t pass viewer clicks to effects, so there you type the numbers into the Subject controls and use the two buttons.

Delivery

Handing Off — Mattes & Depth

Nexus doesn’t have to be the last word. The Output group turns it into a source of channels for the rest of your node tree or timeline.

Output Mode	What comes out
Normal	The rendered image with your overlays. The default.
Matte Only	The subject matte as a white-on-black image, replacing the picture. Feed it straight into a key or a mask input.
Matte in Alpha (Cutout)	The normal render is kept, and the matte is written into the alpha channel — a ready-made cutout to composite over anything.
Depth Map	The depth pass as a greyscale image. Use Depth Black Level and Depth White Level to pull the range tight around your scene before it leaves the plug-in.
Depth in Alpha	The normal render, with depth in the alpha channel.
Overlays Only	Just the overlays — boxes, markers, arrows, connections and labels, with their outline and glow — on a transparent frame, premultiplied. Put a glow, a blur or a blend mode on the HUD alone, then lay it back over the plate. Region FX, Subject Glow, fog and rack focus are left out: they treat the picture, not the overlays.

Honours your selection

Matte exports follow the Subject Mask selection, so you can export a clean alpha for one actor in a crowded frame — automatically, for the whole clip.

Invert Output

Flips the exported matte or depth in one click, for garbage-matte and hold-out workflows.

If a depth mode is selected before the depth engine has a map, Nexus renders the normal image rather than a black frame.

In Overlays Only the status messages and the performance readout count as overlays, so switch Show Performance Stats off before a final export. In demo mode this output stays opaque — overlays on black — so the watermark remains visible.

Pixels Nexus does not draw on pass through exactly as they arrived. On a linear, ACES or HDR timeline, super-whites and values below black survive the node untouched — only the alpha channel is kept within range.

Honestly

Know the Limits

Everything in this part works, and works well. None of it works everywhere. These are the places where the technology has real edges — worth knowing before a deadline finds them for you.

AI detection misses things

Smart & Face Detect

Heavily obscured subjects, very small ones, extreme angles, unusual costumes and heavy motion blur can all fall below the detector’s confidence. Lower AI Confidence to catch more (at the cost of false positives), and raise Revive Frames so a subject that briefly disappears keeps its identity. If the AI simply doesn’t recognise your subject, that is what the classical modes are for.

Classical often beats AI

And that’s the point

For unusual visual targets — a moving light, a laser dot, a coloured marker, a particle, a reflection, an abstract shape — a classical detector is not a fallback, it is the better tool. It is also far cheaper. Reach for the AI when you need the object to be named; reach for classical vision when you just need it found.

Depth is relative, not metric

Monocular estimation

Depth is inferred from a single image. It is normalised to the scene, held steady from frame to frame and excellent for separation, focus, fog and mattes — but it is not a measurement. An exported depth pass is a relative depth channel, not calibrated distance, and its absolute values shift as the scene’s depth range changes.

Segmentation depends on the subject

AI mattes

Matte quality scales with how large the subject is in frame, how much it contrasts with its background, and how complex its edges are. A close-up figure against a clean background cuts beautifully; a distant figure in matching tones does not. Each frame is checked against the sixteen before it, which holds a part the model wavers on and replaces a frame it gets plainly wrong, but the edge itself is not averaged over time: difficult material can still shimmer there, and a small part can drop out for a single frame. Matte Feather is the mitigation. Excellent automatic matte; not a replacement for hand roto on a hero shot.

Long occlusions break identity

Tracking

Revive Frames buys real tolerance, but it is finite. A subject hidden for longer than that, or one that leaves frame and returns much later, comes back as a new identity. Two similar subjects crossing with heavy overlap can also swap — the tracker weighs overlap, motion and the colours inside each box, which prevents most of it, not all of it.

Camera compensation is translation only

Automatic

Nexus estimates and subtracts a global frame-to-frame shift, which covers pans, tilts and handheld drift. A shift only counts when most of the textured picture shares it, so a passer-by filling a third of the frame can’t pass for a camera move. It does not model rotation, zoom or perspective, and on flat or ambiguous frames it deliberately reports no motion rather than guessing. This is identity maintenance, not camera tracking.

Reproducibility has a window

Determinism

Identities reproduce across scrubbing, reverse play and out-of-order rendering within a session. The rolling record holds roughly 600 frames; a render that starts cold, far into a long clip, begins its own chain rather than inheriting one. In Premiere Pro and Final Cut Pro, Nexus works only from the frames it is handed, so there a jump always begins a new chain until the shot has been played in order.

Very small frames pass through

By design

Frames under 200 pixels on either side render untouched. Below that size there isn’t enough information for detection to be meaningful, so Nexus declines rather than producing noise.

One diagnostic caveat

Visualize Detection Field shows what the detector sees. In the classical modes it draws the field they threshold (red = above Threshold), which is exactly what you want while dragging Threshold. In Smart Detect and Face Detect it paints each detection’s box in its class colour, brighter and more opaque the more confident the detector is, over the dimmed frame — a quick read of what was found and how surely.

Part III

Getting Started

Install, license, lock on. Everything you need to get a tracked HUD on screen in minutes.

Requirements & Hosts

Where It Runs

Nx Nexus runs on macOS and Windows, GPU-accelerated through Metal, CUDA or OpenCL depending on your platform. The AI models ship inside the plug-in — the installer delivers everything, and nothing needs to be downloaded, signed into or configured afterwards.

DaVinci Resolve & other OpenFX hostsMAC · WIN · OPENFX

Adobe Premiere ProMAC · WIN

Adobe After EffectsMAC · WIN

Final Cut Pro & MotionMAC · FXPLUG

In every host the effect is named Nx Nexus and lives under the Nx Color group. The engine — detector, tracker, camera compensation, depth, mattes, label layout — is identical everywhere; only the surrounding panel chrome changes.

System requirements

	macOS	Windows
Operating system	macOS 14 Sonoma or later	Windows 10 / 11, 64-bit (x64)
Processor	Apple Silicon or Intel (universal binary)	64-bit x86
Graphics	Any Metal-capable GPU	DirectX 12 GPU for AI acceleration. NVIDIA/CUDA for GPU rendering in Premiere Pro; other GPUs and Software Only mode render on the CPU.
Memory	16 GB minimum, 32 GB recommended for 4K work with depth and segmentation enabled.
Disk space	~2 GB with all three host components	~700 MB per host component; ~1.3 GB for both
Internet	Required once, to activate. Afterwards Nexus runs offline and checks in quietly on its own whenever you happen to be connected.

Each host component carries its own private copy of the AI models, which is why the disk figures look large — installing fewer components saves proportionally.

On macOS, allow a further ~1.6 GB for the prepared AI models in ~/Library/Caches/NxColor/nexus/coreml-cache. It is built once, shared by Resolve, the Adobe apps and Final Cut Pro, and safe to delete at any time — Nexus rebuilds it on the next launch (the first frames then take a little longer while the models are prepared again).

No GPU? Still works.

Every host has a full CPU path that runs the identical engine, so the picture is the same — just slower. The AI models fall back to CPU inference too. Smart Detect in particular is much happier with a GPU; if you’re CPU-bound, the classical detection modes are nearly free and give you the same tracker, the same overlays and the same identity behaviour.

Installing

Adding the Effect

Quit your host applications first — plug-ins are only loaded at launch. Then run the installer for your platform and pick the components you want.

macOS

Open Nx-Nexus-1.0.0-macOS.pkg and follow the prompts. Choose any combination of OpenFX, Adobe and Final Cut components. The Motion template that Final Cut needs is installed for you; nothing else to register.

Windows

Run Nx-Nexus-1.0.0-Windows.exe as administrator. The installer detects which hosts you have and pre-selects the matching components.

Finding it in your host

DaVinci Resolve

In the Effects library, open OpenFX › Nx Color and drag Nx Nexus onto a clip, or onto a node on the Color page.

Premiere Pro

In the Effects panel, open Video Effects › Nx Color › Nx Nexus and drag it onto a clip.

After Effects

With a layer selected, apply Effect › Nx Color › Nx Nexus.

Final Cut Pro

In the Effects browser, find Nx Nexus under Nx Color and drag it onto a clip.

Motion

In the Library, choose Filters › Nx Color › Nx Nexus and apply it to a layer.

Before activation

Until the machine is activated, Nexus runs in demo mode — every feature unlocked, no restrictions, so you can evaluate it properly on your own footage — with a watermark over the output. Nothing else is held back. See the next section to remove it.

Uninstalling

On macOS, double-click nx-uninstall-nxnexus.command in /Library/Application Support/NxColor/. It releases your license seat before removing the files, so the seat is free for another machine. On Windows, use Add or Remove Programs.

Licensing & Activation

Your License

At a glance

One purchase = one license key, emailed right after checkout. It looks like NXNS-XXXX-XXXX-XXXX.

A license covers 2 machines (2 seats). Activate on each machine once.

One activation licenses every supported app on that machine — Resolve, Premiere Pro, After Effects, and on macOS Final Cut Pro and Motion.

Works offline. After a one-time online activation, Nx Nexus keeps running without an internet connection and checks in quietly on its own whenever you’re connected.

Activating

1Buy Nx Nexus — your license key arrives by email within a few minutes. (Check spam if you don’t see it.)

2Install the plug-in, then open any supported host and add the Nx Nexus effect to a clip.

3Open the License group at the bottom of the effect panel and click License…. Paste your key and click Activate. If the key is already on your clipboard, the field fills itself.

The Activate dialog with a license key entered — The Activate dialog. Paste the key from your confirmation email and click Activate.

4The watermark disappears and the License window shows Status: Activated. You’re done — and licensed in every supported app on this machine.

The License window showing Status Activated, license type, masked key and Machine ID — The License window after activation: status and version, license type, your key with the middle masked, and this machine’s ID — the four things support may ask you for.

Moving to a new machine

Your license works on two machines, and it’s easy to move between them.

Switching machines? Deactivate first. On the machine you’re leaving, open the License window, click Deactivate (bottom-left in the window above), and confirm — that seat is freed. (Uninstalling deactivates it too.)

Old machine gone for good? Just activate on the new one. Activating beyond your two machines clears the earlier activations — then re-activate any machine you still use.

Refresh License asks the server for your current status on the spot — useful after a seat change or a long stretch offline.

Good to know

Ordinary upgrades — RAM, GPU or drive — don’t burn a seat; Nx Nexus still recognises your machine. If you stay offline for a long stretch, the License window warns you a week before it needs a check-in; connect once and it resets. Lost your key? It’s in your order email, or sign in at nxcolor.com under My Account. Anything off — email [email protected] and we’ll sort it out.

In Final Cut Pro and Motion the License… button opens a separate Nx Nexus window rather than a dialog inside the host — that’s a Final Cut requirement, and the flow is otherwise identical.

Quick Start

Five Minutes to a Tracked HUD

The five steps at the front of this manual get you a working shot. This is the same path with the refinements filled in — what to reach for, and in what order, once the boxes are on screen.

1Drop it on a clip and play. Boxes and labels appear on everything the detector recognises. On the very first use, give it a few seconds to prepare the model.

2Tidy up the detections. Too many subjects? Raise AI Confidence toward 0.5 and lower Max Trackers. Only want people? Set Class Filter to People.

3Make the IDs stick. In Tracking, set Revive Frames to about 15. This is the difference between IDs that hold through an occlusion and IDs that renumber constantly.

4Choose your look. In Boxes, pick a Tracker Style and colour. In Labels, switch on only the content you want. Center Markers and Connections start switched off; turn each on in its own group when you want it.

5Add one idea. A Region FX on the subject. A Subject Glow rim. Depth fog. A rack focus following your lead. One is usually enough.

If the shot is abstract

No people, no cars, nothing the AI recognises? Switch Detection Mode to Bright Spots (lights, particles, stars) or Select Color (pick the colour you want to follow) and tune Threshold until the right things are found. Everything downstream — tracking, labels, Region FX, glow — behaves exactly the same.

The Nx Nexus panel in DaVinci Resolve with the Detection group open — The panel in DaVinci Resolve with Detection open — sixteen groups, ordered the way you work through a shot. In Smart Detect the classical-mode controls have greyed themselves out.

Part IV

Reference

Every control, every default, every dependency — plus diagnostics, recipes and the honest list of host differences.

Control Reference

Sixteen Groups

The panel is ordered the way you work through a shot: find the subjects, hold them, then draw on them, then treat them, then deliver. Drawing sizes given in pixels are measured at HD and scale with frame size; the two analysis distances — Max Move Distance and Max Connection Distance — are not scaled, so re-check them if you re-tune a look at a different resolution.

Reading the tables

The Default column is what a fresh instance gives you. The line beneath each description tells you what has to be true for the control to do anything — in Resolve and After Effects, controls that fail that test grey out on their own.

Typing past the slider

A few controls accept larger typed values than the slider shows: Text Size up to 0.10, and the Subject and Focus Tracker ID fields up to 999 — tracker IDs keep counting up over a long clip, so the wider range matters. In Final Cut Pro and Motion the sliders already span the full range.

01DetectionWhat to look for

Control	Range	Default	What it does
Detection Mode	8 options	Smart Detect	Chooses what the tracker locks onto — an AI object detector, an AI face detector, or one of six classical image-feature detectors.
AI Confidence	0.05 – 0.95	0.30	Minimum detection score for a subject to appear. Raise it to show only strong, certain detections; lower it to catch distant or partly hidden ones. In Smart Detect a subject that is already on screen is kept down to about 70% of this value, so a brief dip in confidence doesn’t drop its box.Smart Detect and Face Detect only
Class Filter	12 options	All Objects	Limits Smart Detect to one category. Filtered-out objects never become subjects, so they never consume a tracker slot or an ID.Smart Detect only
Target Color	colour	red	The colour that Select Color mode hunts for.Select Color mode only
Threshold	0.0 – 1.0	0.25	How strong a region must be to register. In Select Color mode this is the colour tolerance instead, and Resolve relabels the control to say so.Classical modes only
Blob Size Scale	0.5 – 10.0	1.5	Multiplies the size of every box a classical mode finds. Raise it if boxes sit inside your subjects.Classical modes only
Blob Sensitivity	0.01 – 0.5	0.10	How much accumulated signal is needed before a region counts as a subject. Lower finds more, smaller, weaker things.Classical modes only
Max Trackers	1 – 64	20	Hard cap on how many subjects are tracked and drawn at once. Beyond the cap the largest subjects win, with a bias toward whoever was already on screen.
Detection Detail	8 – 32	16	Analysis resolution for the classical modes. Higher separates subjects that are close together and tightens their boxes; lower merges them and is faster.Classical modes only
Invert	off / on	off	Track what’s below the threshold instead of above it — dark subjects on a light background. Labelled Invert (track below threshold) in After Effects and Premiere Pro.Luminance and Saturation modes only
Pre-Blur	0.0 – 8.0	0.0	Softens the analysis before thresholding, so grain and noise stop generating phantom subjects. 0 = off.Classical modes only
Mask Cleanup	0 – 5	0	Removes speckles and fills small holes in the detected regions. Each step up removes larger specks and closes wider gaps, measured in cells of the analysis grid. Raise it when a subject keeps fragmenting into pieces.Classical modes only
Max Blob Size %	5 – 100	40	Rejects — and clamps — any box wider or taller than this percentage of the frame’s longer side. The cure for one blob swallowing the shot. 100 = no limit.Classical modes only

Detection Mode Smart Detect (AI) · Face Detect (AI) · Luminance · Edges · Saturation · Select Color · Bright Spots · Dark Spots
Class Filter All Objects · People · Vehicles · Animals · People + Vehicles · People + Animals · Sports Equipment · Food & Drink · Electronics · Furniture · Bags & Accessories · Street Objects

02TrackingHolding identity

Control	Range	Default	What it does
Temporal Stability	off / on	on	The master switch. On, subjects are matched frame to frame and carry stable IDs, smoothed boxes and velocity. Off, they are simply ranked by size each frame.
Temporal Smoothing	0.0 – 1.0	0.5	Trades responsiveness for steadiness. 0 snaps to the detector; 1 glides and lags. Raise it for calm graphics, lower it for fast action.Needs Temporal Stability
Max Move Distance	0 – 500 px	0	How far a subject may travel between frames and still keep its ID. 0 sizes the gate automatically from each subject — usually the right answer.Needs Temporal Stability
Revive Frames	0 – 60	0	How long a subject’s ID survives while it can’t be seen. The occlusion control — at 0 a single missed frame ends the track and the subject returns renumbered. Try 10–20.Needs Temporal Stability
Confirm Frames	0 – 30	0	A subject must persist this many frames before it is drawn at all. The cure for one-frame false positives popping in and out.Needs Temporal Stability
Show Occluded (Ghost)	off / on	off	Keeps drawing a hidden subject’s predicted box, dimmed, while its ID coasts — far better on screen than a box that blinks out.Needs Temporal Stability, and Revive Frames above 0

03BoxesThe frame around each subject

Control	Range	Default	What it does
Show Tracking Boxes	off / on	on	Group master.
Tracker Style	7 options	Corner Brackets	The shape drawn around each subject.
Box Color Mode	3 options	Fixed Color	How boxes and centre markers are coloured: one colour, a distinct colour per object class, or warm-to-cool by distance.Needs Boxes or Markers to be on. By Depth also needs Enable Depth, otherwise it falls back to Fixed
Tracker Color	colour	green	The box colour in Fixed Color mode.
Box Line Width	0.5 – 6.0 px	1.0	Line weight, measured at HD.
Bracket Length %	5 – 100	70	How far the corner brackets reach along each edge. At 100 they meet and you get a closed rectangle.Corner Brackets and Reticle styles
Circle Gap	5 – 45°	15	Size of the four breaks in the targeting circle.Targeting Circle style
Tracker Opacity	0.0 – 1.0	1.0	Box opacity.

Tracker Style Corner Brackets · Rounded Rectangle · Targeting Circle · Hexagon · Diamond · Reticle · Target Rings
Box Color Mode Fixed Color · By Class · By Depth

04Center MarkersThe point, and where it’s heading

Control	Range	Default	What it does
Show Center Markers	off / on	off	Group master for the marker itself.
Marker Style	10 options	Dot + Ring	The marker shape.
Marker Color	colour	magenta	Marker colour — also used for the velocity arrow.
Marker Size	2 – 80 px	16	Marker size at HD; scales with the frame.
Velocity Arrow	off / on	off	Draws a motion vector from the centre of every moving subject. Its own overlay — it still draws with the markers switched off. For a numeric speed readout instead, use Labels › Speed.

Marker Style Dot · Plus · Cross · Crosshair · Ring · Target · Dot + Ring · Diamond · Square · Triangle

05ConnectionsLines between subjects

Control	Range	Default	What it does
Show Connections	off / on	off	Group master. Everything below needs it.
Connection Type	6 options	Sequential	Which subjects get linked. Proximity, Nearest Neighbor and Spanning Tree build sparse, readable webs — cleaner and faster than a full mesh when there are many subjects.
Max Connection Distance	0 – 2000 px	0	Link subjects closer together than this. 0 adapts automatically to how spread out the scene is.Proximity type
Nearest Links	1 – 8	2	How many neighbours each subject links to.Nearest Neighbor type
3D Tilt (Depth)	0.0 – 1.0	0.0	Tilts links into the scene: the end at the farther subject tapers thin and dim, and the line bows. 0 = flat.Needs Enable Depth
Line Style	6 options	Straight	Straight, organic curves, a 45°-dogleg circuit-board trace, or right-angled routes with sharp or filleted corners. Arrowheads follow the actual route.
Curve / Bend Amount	−1.0 – 1.0	0.5	How far Curved and Wave lines bow — negative bows the other way. Also sets the corner radius on Smooth Bend.Curved, Wave and Smooth Bend styles
Corner Position	0.0 – 1.0	0.5	Where along a routed path the right-angle turn happens.PCB Trace, Smooth Bend and Step Bend styles
Line Pattern	3 options	Solid	Solid, dashed or dotted. Dashes march when Animate Overlays is on.
Dash Length	5 – 30 px	10	Length of each dash or gap.Dashed and Dotted patterns
Show Arrows	off / on	off	Directional arrowheads riding the line.
Arrow Style	5 options	Filled Triangle	Arrowhead shape.Needs Show Arrows
Arrow Size	5 – 50 px	25	Arrowhead size at HD.Needs Show Arrows
Arrow Position	0.0 – 1.0	0.5	Where the arrowhead sits along the line.Needs Show Arrows
Line Color	colour	cyan	Connection colour.
Connection Width	0.5 – 5.0 px	2.5	Line weight at HD.
Line Opacity	0.0 – 1.0	1.0	Connection opacity.

Connection Type Sequential · Star · Mesh · Proximity · Nearest Neighbor · Spanning Tree (MST)
Line Style Straight · Curved · Wave · PCB Trace · Smooth Bend · Step Bend
Line Pattern Solid · Dashed · Dotted Arrow Style Filled Triangle · Outline Triangle · Angle · Double Angle · Circle

06LabelsWhat each subject says about itself

Control	Range	Default	What it does
Show Labels	off / on	on	Group master — off hides all label text and the plate behind it.
Object Name	off / on	on	The detected object name, such as PERSON or CAR. It follows the subject’s whole history, so it doesn’t flip when the detector wavers between two names.Smart Detect and Face Detect only; faces carry no name
Tracker ID	off / on	on	The stable identity number — the one you type into Subject 1–4 and Focus Tracker ID.
Coordinates	4 options	Off	The subject’s centre position in pixels: both axes, or just one.
Speed	off / on	off	Speed in pixels per frame. For the on-screen arrow instead, use Center Markers › Velocity Arrow.
Depth (Z)	off / on	off	Distance readout, 0 to 100, where 100 is nearest.Needs Enable Depth
Confidence %	off / on	off	How certain the detector is about this subject.
Text Size	0.01 – 0.05 (type to 0.10)	0.02	Label height as a fraction of the frame, so it holds across resolutions.
Text Color	colour	yellow	Label text colour.
Text Opacity	0.0 – 1.0	1.0	Label text opacity.
Background Plate	off / on	on	A dark, rounded plate behind each label so text stays legible over bright or busy footage.
Plate Opacity	0.0 – 1.0	0.55	How solid that plate is.Needs Background Plate

Coordinates Off · X Y · X only · Y only Order on screen NAME ID:n X:n Y:n V:n Z:n n%

07Overlay StyleOutline and glow for every layer

Control	Range	Default	What it does
Overlay Outline	0.0 – 1.0	0.5	A thin dark edge under boxes, markers, arrows, connections and label text, and around the label plate, so light overlays hold on bright footage. Higher is wider and darker. 0 = off.
Overlay Glow	0.0 – 1.0	0.0	Light around boxes, markers, arrows and connections, screen-blended onto the picture under the overlays. Labels take no glow, and the plate hides any glow behind it. In Overlays Only output the glow is part of what is exported. 0 = off.
Overlay Glow Width	2 – 80 px	16	How far the glow reaches, at HD; scales with the frame.Needs Overlay Glow above 0
Overlay Glow Color	colour	black	The glow colour. Left at black, each overlay glows in its own colour.Needs Overlay Glow above 0

08AnimationLife in the overlay

Control	Range	Default	What it does
Animate Overlays	off / on	off	Drives overlay motion from the timeline: dashes march, circular reticles spin, centre markers and the subject glow pulse, and the Hologram and Heat Ripple effects move. Driven by frame number, so it renders identically every pass.
Animation Speed	0.1 – 5.0×	1.0	Speed multiplier for all of the above.Needs Animate Overlays

09DepthFront, back and the air between

Control	Range	Default	What it does
Enable Depth	off / on	off	Estimates scene depth and reads a distance for every subject. Master for this group’s features plus the Z label and By-Depth colours — rack focus, depth export, depth mattes and the depth visualiser switch the engine on by themselves.
Occlude by Depth	off / on	off	Hides a subject’s overlays wherever something nearer is in front of it. Turns a flat HUD into a depth-aware one.Needs Enable Depth
3D Connection Depth	0 – 4000	0	How much distance counts when deciding which subjects are “near” each other. Above 0, the proximity graphs link true 3D neighbours.Needs Enable Depth
Fog Amount	0.0 – 1.0	0.0	Atmospheric haze graded by distance, edge-aware so it doesn’t bleed across subjects. 0 = off.Needs Enable Depth
Fog Color	colour	pale blue	The colour of the haze.Needs Enable Depth, and Fog Amount above 0
Fog Start (Near)	0.0 – 1.0	0.8	The distance at which fog begins. Anything nearer stays clear. 1 is nearest to camera.Needs Enable Depth, and Fog Amount above 0
Fog End (Far)	0.0 – 1.0	0.2	The distance at which fog reaches full strength. 0 is farthest from camera.Needs Enable Depth, and Fog Amount above 0

10Rack FocusDepth of field

Control	Range	Default	What it does
Focus	5 options	Off	What decides the focus distance. The automatic modes follow a subject and ease into the new distance over roughly half a second, so the pull looks operated rather than snapped. Together with a Blur Strength above 0 it switches the depth engine on by itself.
Focus Tracker ID	1 – 64 (type to 999)	1	Which subject to keep sharp. Read the number off its label.Follow Tracker ID mode
Focus Depth	0.0 – 1.0	0.5	The focus distance, set by hand. 1 is nearest to camera, 0 is farthest — keyframe it for a manual rack.Manual mode
Focus Range	0.0 – 1.0	0.15	How deep the sharp zone is around the focus distance. Small is a shallow lens; large keeps most of the scene sharp.Needs a Focus mode
Blur Strength	0.0 – 2.0	0.0	How strongly out-of-focus areas defocus. At 0 nothing happens at all — picking a Focus mode is not enough on its own.Needs a Focus mode

Focus Off · Manual (slider) · Follow Tracker ID · Nearest Subject · Farthest Subject

11Subject MaskWho is selected, and how closely the mask fits

Control	Range	Default	What it does
Subject 1	−1 – 64 (type to 999)	−1	Which subjects Region FX, Subject Glow and matte export apply to. −1 selects all of them, 0 selects none, and a number selects that subject. Clicking a subject in the viewer sets it for you.Used by Region FX, Subject Glow and matte export
Subject 2	0 – 64 (type to 999)	0	A second selected subject, added to the first. 0 means unused.
Subject 3	0 – 64 (type to 999)	0	A third selected subject. 0 means unused.
Subject 4	0 – 64 (type to 999)	0	A fourth selected subject. 0 means unused.
Select All Subjects	button	—	Selects every tracked subject: Subject 1 goes to −1 and the others to 0.
Clear Selection	button	—	Selects nothing: all four go to 0, and the subject effects wait until you click or type a subject.
Matte Source	3 options	Box	How tightly the mask hugs each selected subject. Shapes Region FX and matte export; Subject Glow always follows the AI outline. Falls back gracefully if a source isn’t available.Needs a Region FX or a matte export
Matte Quality	2 options	High (Per-Subject)	High zooms the AI into each subject for edge detail that scales with subject size; Standard analyses the whole frame once and is notably faster.Segmentation source, or any Subject Glow
Matte Range	0.02 – 0.5	0.15	How wide a band of distance counts as the subject. Tighter cuts closer to the body.Depth Silhouette source
Matte Feather	0.0 – 1.0	0.0	Softens the matte edge. The first thing to reach for when hair or motion blur shimmers. It also shapes the Subject Glow outline.Segmentation source, or any Subject Glow
Matte Grow / Shrink	−20 – +20 px	0.0	A true expand or choke of the matte. Positive swallows an edge fringe; negative pulls inside the subject. It also shapes the Subject Glow outline.Segmentation source, or any Subject Glow

Matte Source Box · Depth Silhouette · Segmentation (AI) Matte Quality Standard · High (Per-Subject)

12Region FXTreating the selected subjects

Control	Range	Default	What it does
Region Effect	20 effects	None	An image effect applied to the tracked subject regions — privacy blurs, censor mosaics, stylised looks, spotlighting.
FX Amount	0.0 – 1.0	0.5	The strength dial: blur radius, mosaic block size, desaturation, hue rotation, glitch intensity or dim amount, depending on the effect.Needs an effect other than None or Tracker Zoom
Zoom Level	1.0 – 5.0×	2.0	Magnification for the Tracker Zoom effect.Tracker Zoom effect
Apply To	2 options	Inside Subjects	Treat the selected subjects, or treat everything else.Needs an effect other than None, Tracker Zoom or Spotlight — those three ignore it

Region Effect None · Tracker Zoom · Spotlight (Dim BG) · Blur · Pixelate (Mosaic) · Solid Fill · Sharpen · Black & White · Invert · Colorama (Hue Shift) · Posterize (Cel Shade) · Thermal · Night Vision · Halftone · Ink Sketch · Dither (Retro) · ASCII Mosaic · Hologram · Chromatic Ghost · Glitch · Heat Ripple
Apply To Inside Subjects · Outside Subjects

13Subject GlowLight on the silhouette

Control	Range	Default	What it does
Glow Intensity	0.0 – 2.0	0.0	Group master. A rim and halo hugging the subject’s AI outline, blended as emitted light. Follows the Subject Mask selection and pulses with Animate Overlays. 0 = off.
Glow Width	4 – 60 px	18	How far the halo reaches beyond the silhouette, at HD.Needs Glow Intensity above 0
Glow Color	colour	cyan	Rim and halo colour.Needs Glow Intensity above 0

14OutputHanding channels downstream

Control	Range	Default	What it does
Output Mode	6 options	Normal	Exports the subject matte, the depth pass or the overlays on their own, for compositing. Matte modes use the Matte Source chain and follow the Subject Mask selection.
Invert Output	off / on	off	Flips the exported matte or depth, for hold-out and garbage-matte workflows.Needs a matte or depth Output Mode
Depth Black Level	0.0 – 1.0	0.0	The depth that maps to black in an exported depth pass — pull the range tight around your scene before it leaves the plug-in.Depth output modes
Depth White Level	0.0 – 1.0	1.0	The depth that maps to white.Depth output modes
Overlays Ignore Alpha	off / on	off	Forces alpha fully opaque wherever an overlay is drawn, so a HUD stays solid over a keyed or transparent background.

Output Mode Normal · Matte Only · Matte in Alpha (Cutout) · Depth Map · Depth in Alpha · Overlays Only

15DiagnosticsSeeing what the plug-in sees

Control	Range	Default	What it does
Visualize Detection Field	off / on	off	Replaces the image with what the detector sees: the thresholded field in the classical modes (red = above Threshold), the detections as class-coloured boxes weighted by confidence in the AI modes.
Visualize Depth Map	off / on	off	Replaces the image with the depth map, near white to far black. Switches the depth engine on by itself. For a deliverable pass use Output › Depth Map instead.
Visualize Segment Masks	off / on	off	Tints each subject’s AI outline in its own colour over the normal render, so you can check what the matte is actually cutting.
Show Performance Stats	off / on	off	A labelled readout in the bottom-right corner: detection mode, AI backend and frame; tracker count; per-stage times (detection, depth, matte) in milliseconds; and the state of every AI model (ready, preparing with a clock, failed, missing, not loaded). Frames replayed from the tracking history say CACHED instead of a time.

16LicenseActivation

Control	Range	Default	What it does
Status	read-only	—	Live license state. Shows DEMO — Watermark Active until you activate, then Activated ✓, and warns you in advance if an offline check-in is due.DaVinci Resolve and other OpenFX hosts
Website	read-only	nxcolor.com	Where to find your account and support.Final Cut Pro and Motion
License…	button	—	Opens the activation window when unlicensed, and the manage window — refresh, deactivate, machine ID — once activated.

Diagnostics

What the Plug-in Tells You

The performance readout

Switch on Show Performance Stats and a compact block appears in the bottom-right corner of the frame — one topic per row, with a fixed label column so the numbers line up and can be read at a glance:

NX NEXUS: the detection mode, where the AI is running (COREML on Apple hardware, DIRECTML on Windows GPUs, CPU if neither is available) and the frame number. TRACKED: subjects published out of the Max Trackers cap. TIMING: how long each stage took on this frame — CACHED when the frame was replayed from the tracking history without running the detector again, and depth and matte times only when those passes actually ran. MODELS, over two rows: every AI model’s state — READY, PREPARING with a running clock, FAILED, MISSING, or NOT LOADED for a model nothing has asked for yet. If you ever wonder why a shot is slow, or why nothing is happening yet, this block answers it in one glance.

On-screen model status

You don’t need the performance readout to know what the AI is doing. Whenever a model that your current settings need is not ready, a small message appears in the top-left corner on every host — there is no switch for it. Amber means wait; red means something needs your attention:

Message	What it means
TRACKING AI PREPARING... FACE AI PREPARING... DEPTH AI PREPARING... MATTE AI PREPARING...	That model is being prepared for your GPU. Normal on first use — the line beneath, TAKES A MOMENT ON FIRST USE, says so (the per-model clock lives in the performance readout, if you want to watch it). Boxes, depth or mattes appear the moment it flips to ready and the viewer refreshes by itself; the work is cached, so it happens once per machine, not once per project.
STILL LOADING...	Nexus never runs AI inference while another model is being prepared in the same host, so boxes or maps can disappear for a moment during a first-time preparation. They return by themselves when it finishes.
… FAILED - SEE NX NEXUS LOG	The model couldn’t start. The matching log file (see When Something Looks Wrong) names the reason on its first line; send it to support.
… MISSING - REINSTALL	The model file is not inside the installed plug-in. Reinstall Nx Nexus.
MATTE AI NO MASKS - SEE NX NEXUS LOG	The outline model is running but producing nothing for this frame. Try switching Matte Quality to Standard; if it persists, contact support with the segmentation log.
DEPTH AI NOT AVAILABLE - RENDERED WITHOUT IT MATTE AI NOT AVAILABLE - RENDERED WITHOUT IT	Your settings ask for depth, or for a matte, but this frame was rendered without one, and not because a model is still preparing or has failed. What you see is the plain picture, not the look you asked for. Render the frame again; if the message stays, send the matching log to support.
AI MODEL RUN FAILING - SEE NX NEXUS LOG	A model loaded fine but its per-frame run keeps failing on this machine. The log records each failure.

The notice only speaks about models your settings actually use — a matte model preparing quietly in the background is not mentioned until you reach for a matte control — and it never draws on a matte or depth export. In Overlays Only output it counts as an overlay and is exported with the rest.

Diagnostic views

Visualize Detection Field

Classical modes: the raw field, the fastest way to set Threshold. AI modes: every detection’s box in its class colour, brighter with confidence.

Visualize Depth Map

The depth pass as greyscale. Useful for sanity-checking fog ranges and focus distances.

Visualize Segment Masks

Tints each AI outline over the normal render, so you can confirm the matte is on the right subject before you commit to it.

Recipes

Known-Good Starting Points

Each of these takes under a minute from a fresh instance.

You want	Set
Blur every face in the shot	Detection Mode Face Detect · Region Effect Pixelate or Blur · Apply To Inside Subjects · Revive Frames 15 · switch off Boxes and Labels.
A clean sci-fi targeting HUD	Tracker Style Reticle · Show Center Markers on, Marker Style Crosshair · Show Connections on, Connection Type Nearest Neighbor · Line Style PCB Trace · Line Pattern Dashed · Animate Overlays on.
Isolate one actor with a rim light	Click the actor in the viewer (or type their number into Subject 1) · Matte Source Segmentation (AI) · Glow Intensity around 0.8 · Glow Width to taste. Add Region Effect Spotlight to dim the rest.
A cutout with a real alpha channel	Select the subject · Matte Source Segmentation (AI) · Matte Quality High · Output Mode Matte in Alpha (Cutout) · a little Matte Feather if hair shimmers.
Follow-focus on a moving subject	Focus Follow Tracker ID · enter the ID from the label · Blur Strength around 0.8 · Focus Range around 0.12. (Blur Strength must be above 0 or nothing happens.)
Push the background back	Enable Depth on · Fog Amount around 0.5 · Fog Start 0.75 · Fog End 0.15 · tint Fog Color toward your grade.
Track lights, stars or particles	Detection Mode Bright Spots · Threshold to taste · Blob Sensitivity down to catch small ones · Max Blob Size % down to reject the sky.
A depth pass for compositing	Output Mode Depth Map · set Depth Black and White Levels to bracket your scene · switch off every overlay group.
The HUD as its own layer	Duplicate the clip · on the upper copy, Output Mode Overlays Only · add a glow, a blur or a blend mode to that copy alone · the clean clip underneath shows through wherever nothing is drawn.
A crowd data-viz look	Class Filter People · Max Trackers 40 · Show Connections on, Connection Type Spanning Tree · Labels: ID and Speed only · Box Color Mode By Depth with Enable Depth on.
Maximum speed	A classical Detection Mode · Enable Depth off · Matte Source Box · Glow Intensity 0 · Max Trackers low. All the overlay styling stays available.

Troubleshooting

When Something Looks Wrong

No boxes at all on the first use

The AI model is still being prepared for your GPU. Give it a few seconds to a minute — this happens once per machine, then never again — and the viewer updates by itself when it is ready (in Premiere Pro, touch any control to refresh). If boxes never appear, check that your Detection Mode is what you expect and that AI Confidence isn’t set very high.

IDs keep changing

Raise Revive Frames. At its default of 0, a single missed detection ends the track, and the box blinks out with it. Try 15. If IDs still churn, raise Temporal Smoothing, and check that Max Trackers is comfortably above the number of subjects in frame.

IDs differ after scrubbing

A frame reached by jumping has no predecessor to build on. Play the shot through once and the numbering settles onto the correct sequential set as the playhead passes over each frame. After Effects works the earlier frames out for you, and Resolve does over short jumps — see The Same Every Time.

The first frame after a jump is slow

In After Effects, and over short jumps in Resolve, Nexus first tracks the frames leading up to the one you jumped to, and with a Segmentation matte it reads up to sixteen of them for the matte as well. That is what makes the frame match a playthrough. It happens once; the frames after it play normally.

Boxes clip the edge of the subject

In the classical modes, raise Blob Size Scale. In Smart Detect, the box comes from the detector; raising Temporal Smoothing slightly and lowering AI Confidence often gives a fuller box on a partly-occluded subject.

Too many detections

Raise AI Confidence, set a Class Filter, lower Max Trackers, and add a couple of Confirm Frames. In the classical modes, raise Threshold and Blob Sensitivity and lower Max Blob Size %.

The matte is a rectangle

Set Matte Source to Segmentation (AI). If it’s already set, look for MATTE AI PREPARING... top-left — the outline model is still preparing itself and the matte is temporarily falling back.

The matte edge shimmers

Add a little Matte Feather. A small negative Matte Grow / Shrink also helps by pulling the edge inside the subject where the model is most confident.

A status message is baked into a render

The render started while a model was still being prepared, which happens on first use. Nexus waits briefly for a model that is almost ready, but a first-time preparation can take longer than that. Let the top-left message clear in the viewer before you start a final render.

Rack focus does nothing

Blur Strength is at 0. Choosing a Focus mode alone has nothing to express — raise Blur Strength and the pull appears.

Depth features look inert

Fog, occlusion, 3D tilt, the Z label and By-Depth colours all need Enable Depth. (Rack focus, depth export, the depth matte and the depth visualiser don’t — they switch it on themselves.)

Playback is slow

Turn on Show Performance Stats and read where the time goes. Matte time dominant? Switch Matte Quality to Standard. Detection dominant? Try a classical mode, or lower Max Trackers. Depth dominant but unused? Switch Enable Depth off.

Clicking a subject does nothing

In DaVinci Resolve, the OpenFX overlay must be visible in the viewer. In After Effects, select Nx Nexus in the Effect Controls panel first, then click in the Composition panel with the Selection tool. In Final Cut Pro, select the clip and click Nx Nexus in the Video inspector. Premiere Pro doesn’t pass viewer clicks to effects — type the number from the subject’s label into Subject 1 instead.

The effect isn’t in the host at all

The host was running during installation. Quit it completely and relaunch — plug-ins are only loaded at launch.

Very small frames render untouched

Frames under 200 pixels on either side pass through unmodified. Thumbnails and tiny proxies are below the useful limit for detection.

Still stuck?

Email [email protected] with your host and version, your operating system, and a screenshot of the panel. If an AI feature is misbehaving, Nexus keeps small diagnostic logs (they never grow past a few megabytes) — on macOS in ~/Library/Logs/NxColor/ for every host, on Windows in %LOCALAPPDATA%\NxColor\, named nx_nexus_yolo.log, nx_nexus_depth.log, nx_nexus_face.log and nx_nexus_seg.log, plus nx_nexus_trace.log, which records model loading and cache activity for every host. Attaching the relevant one saves a round trip. If a message is showing in the top-left corner of the frame (What the Plug-in Tells You › On-screen model status), quote it word for word — it tells us which model, and whether it is still preparing or has actually failed.

Host Differences

The Same Engine, Four Panels

Every feature in this manual is available in every host — the detector, tracker, depth, mattes, Region FX, glow and output modes are one shared engine. What differs is what the surrounding application allows a plug-in panel to do.

The Nx Nexus panel in DaVinci Resolve — The same sixteen groups in the same order, whichever host you open them in. Resolve and the Adobe apps fold each group away; Final Cut lists them all in one scroll.

The Nx Nexus effect in Premiere Pro and After Effects — The same sixteen groups in the same order, whichever host you open them in. Resolve and the Adobe apps fold each group away; Final Cut lists them all in one scroll.

	Resolve / OpenFX	After Effects	Premiere Pro	Final Cut / Motion
Greys out inactive controls	Yes	Yes	No	No
Tooltips on controls	Yes	No	No	No
Colour controls	Colour well	R / G / B sliders	R / G / B sliders	Colour well
Click a subject in the viewer	Yes	Yes	No — type the Subject numbers	Yes
Fetches earlier frames after a jump	Up to a dozen	Yes	No	No
GPU rendering	Metal / CUDA / OpenCL	CPU (by design)	Metal · CUDA on Windows	Metal
Live license status text	Yes	In the dialog	In the dialog	In the window

Premiere Pro rendering

GPU rendering uses Metal on macOS and requires an NVIDIA card (CUDA) on Windows. On other GPUs, in OpenCL mode, or in Mercury Software Only, Premiere renders Nexus on the CPU — every feature still works, just more slowly, and the picture is identical.

After Effects

Nexus renders on the CPU in After Effects, by design: it is the path on which two Nexus layers in one project keep separate identities. Each frame’s pixel work is spread across the processor’s cores, and frames may be rendered in any order, several at a time — Multi-Frame Rendering gives the IDs, boxes and mattes of a straight playthrough.

Premiere Pro & Final Cut panels

Every control appears active, including ones that do nothing in the current mode — those hosts give plug-ins no way to grey a control out. Use the group structure, and the Control Reference in this manual, which notes under each control what it needs in order to do anything.

Colours in Adobe

Premiere’s GPU interface has no colour-picker type, so colours are three sliders. The values are identical to the colour wells in the other hosts.

Part V

The Foundation

What Nx Nexus is built on, and who to talk to.

Credits

Built on Open Research

Nx Nexus is built on published computer-vision research and permissively licensed open-source components. Each third-party component is distributed under its respective licence, with the applicable licence texts included alongside the plug-in.

Each was selected for the combination of accuracy, performance and practical deployment inside a real-time editing environment — and all of it runs on your own machine.

Nothing leaves your computer. There is no cloud service, no account, no telemetry and no upload. Your footage is analysed locally, every frame, on your own hardware.

What	Built on	Used for
Object detection	YOLOX — Megvii, 2021 Apache 2.0	Smart Detect. Recognises eighty classes of everyday object, trained on the COCO dataset.
Face detection	YuNet — OpenCV Zoo MIT	Face Detect. A very small, very fast dedicated face model.
Multi-object tracking	ByteTrack — Zhang et al., 2022	The identity layer: two-stage association, motion prediction, lifecycle. Extended here with camera-motion compensation, a colour signature for each subject and a fully deterministic per-frame history.
Assignment	The Hungarian algorithm — Kuhn, 1955	Matching predictions to detections optimally rather than greedily, which is what keeps IDs from swapping when subjects cross.
Depth estimation	Depth Anything V2 Apache 2.0	Monocular depth for fog, occlusion, 3D links, rack focus, depth mattes and the depth pass.
Segmentation	Segment Anything 2.1 — Meta AI Apache 2.0	The AI subject matte, prompted by the tracker’s own boxes.
Inference runtime	ONNX Runtime MIT	Running all of the above on Apple Neural Engine, Windows GPUs via DirectML, or CPU.

The table names the principal components and the research they come from; the complete third-party licence texts are installed alongside the plug-in.

Support

Getting Help

Support

[email protected]
Include your host and version, your OS, a screenshot of the panel and any message shown in the top-left of the frame.

Account & licence keys

nxcolor.com › My Account
Your key is also in your original order email.

Nx Nexus

Nx Nexus performs all analysis locally. No footage, frames or metadata are ever transmitted.