ANR AI Point Tracker

Every tracking control, under the toolbar row of transport buttons.

Add PointButtonanrPointTb_addPoint

Add an algorithm point, then click its feature in the viewer. Existing tracks stay in place.

|◀ButtonanrPointTb_trackStepBackward

Tracks one frame backward and stops.

◀ButtonanrPointTb_trackBackward

Tracks backward from the current frame one frame at a time, the viewer following, until the start of the Range or until you press ■.

■ButtonanrPointTb_stop

Stop ANR AI Point Tracker: ANR AI Point Tracker1

▶ButtonanrPointTb_trackForward

Tracks forward from the current frame one frame at a time, the viewer following, until the end of the Range or until you press ■.

▶|ButtonanrPointTb_trackStepForward

Tracks one frame forward and stops.

Center ViewerButtonanrPointTb_center

Centres the viewer once on the selected points (on the first point when none is selected).

FollowCheckboxanrPointTb_follow

Keeps the viewer centred on the selected points as the frame changes.

Default: off

Analyse RangeButtonanrPointTb_analyse

Same as Analyse Whole Range (Best Quality).

Refine OccludedButtonanrPointTb_refine

Same as Refine Occluded Frames: re-applies Post Tracking Refine to the tracks on the node.

StatusStatusstatus

Reports whether the inference models are installed and which device is in use.

Load ModelButtonloadModel

Load the chosen tracker into memory. Takes a few seconds. The node passes the picture through until this succeeds.

UnloadButtonunloadModel

Release the model and its VRAM. Useful before rendering something else heavy on the same GPU.

Tracking ModelMenutrackModel

Frame by Frame is included with the plugin, and fine for paid work.

Extra Heavy is a research-only model that does not come with the plugin. The first time you pick it, its licence is shown; accept it and choose a folder, and it downloads from its authors. It may only be used for non-commercial research, and the same goes for its tracks. Research Models... lists what is installed and removes it.

Frame by Frame Tracker (included): the AI tracker, fast, one frame at a time; track it with the transport buttons. Frame by Frame Extra Precision: the same AI tracker followed by sub-pixel matching against each point's reference patch, rejecting weak or ambiguous matches; slower. Classical Precision (CPU - No AI): a pattern tracker on the CPU with a pattern and search box per point.

Default: Frame by Frame Tracker (included)Options: Classical Precision (CPU - No AI) Frame by Frame Tracker (included) Frame by Frame Extra Precision Full Range Tracker (Extra Heavy) - research only

Track PointPositionSome modesclassicalCenter

Selected algorithm point. Add Point in Points, then click to place. Grab its marker to move it; empty clicks preserve existing tracks.

Default: 0 0

Pattern SizeSliderSome modesclassicalPatternSize

Inner box width in source pixels. Drag its corners in the viewer. Include a distinctive feature, not just a flat area.

Default: 61

Search SizeSliderSome modesclassicalSearchSize

Outer search box width in source pixels, centered at the last measured point. Enlarge for fast movement. Drag its corners in the viewer.

Default: 401

Match ThresholdSliderSome modesclassicalThreshold

Classical Precision: how well a match must agree with the point's reference pattern to be accepted, from 0.5 to 0.99. Below it the point is lost on that frame. Raise it to reject doubtful matches, lower it for soft or changing features.

Default: 0.75

Moving CropCheckboxSome modesclassicalMovingCrop

Process a crop around each algorithm point, including its full search box and optical-flow margin. Turn off to compare full-image processing. This does not enlarge the search box.

Default: on

Research ModelStatusanrResearchInfo

Extra Heavy is a research-only model you install yourself. This line says whether it is installed here.

Research Models...ButtonanrResearchManage

The research-only models installed on this machine: their licences, where they are, and Remove.

Extra Heavy Tracker Window (frames)SlidertrackWindow

How many frames the Full Range Tracker (Extra Heavy) looks at together on every step: the reference frame plus the most recent ones. Each step is a full pass over the window, so the time per frame grows with it.

Measured on a 4K face shot: 16 frames took 3.0 s a frame, 8 frames 2.4 s and 4 frames 1.2 s, and the tracks were no worse at 8 (4.6 against 4.5 px on a round trip) or even at 4 (1.4 against 1.9 px on the lips). A longer window gives the model more context through an occlusion; the reference frame is always in it, so re-detection does not depend on this. The window is also capped by what fits on the card. The Frame by Frame Tracker ignores it.

Default: 8

Points FromMenupromptMode

Where the points to follow come from.

Grid In Box -- drag a box and scatter an even grid inside it. For planar surfaces, warps and mesh work, where you want many points rather than chosen ones.

Click Points -- click them in the viewer, the way you would place trackers by hand. Best when you know which features matter.

From Matte -- scatter points inside the matte wired into the Matte input, so only the subject is tracked. Feed it an ANR AI Matte and the two nodes work together.

Default: Click PointsOptions: Grid In Box Click Points From Matte

PointsSlidertrackGrid

How many points to scatter, for Grid In Box and From Matte. They are laid out to suit the shape of the box, so 40 points in a wide box gives more columns than rows. Anything landing outside the box or off the matte is dropped, so the result can be fewer than asked for.

Default: 20

Subtract ModeCheckboxsubtractMode

When on, clicking in the viewer adds an exclude point instead of an include point. Ctrl-click does the same in hosts that forward modifier keys.

Default: off

Clear PromptsButtonclearPrompts

Remove every click point and box from the current frame.

Reference FrameSlidertrackRefFrame

The frame the points are placed on. Analyse runs forward to the end of the range and backward to its start, both from here, so a point can be placed where it is easiest to see rather than on the first frame.

Default: 1

Set to Current FrameButtontrackRefFromPlayhead

Set Reference Frame to the frame the viewer is on.

A new node already starts on the frame it was created at, so this is for after you have moved somewhere else and want to prompt from there instead.

RangeMenutrackRangeMode

Which frames to track. Input uses the range of the footage into this node, Global uses the project range, and Custom uses the two frame fields below.

Default: InputOptions: Input Global Custom

Custom FirstSlidertrackRangeStart

First and last frame to analyse, used when Range is set to Custom.

Default: 0

Custom LastSlidertrackRangeEnd

First and last frame to analyse, used when Range is set to Custom.

Default: 0

Analyse Whole Range (Best Quality)ButtontrackAnalyse

Solves the whole range as one clip with either Full Range Tracker: the model sees every frame of the range together, both directions, which is what those models were built for and many times faster than frame by frame. The Frame by Frame Tracker has no whole-clip mode, so the button is off while it is selected; track it with the transport buttons.

The final-quality pass. Seeds at Reference Frame, walks the range forward and backward from there, and writes a keyframe per frame.

This route ALWAYS feeds the model the plate at its own resolution -- Track Detail does not reduce it -- and it may use the two-pass moving crop, which is the only thing here that sees the whole shot before deciding where to look. It is the slow one, and it is the one to use on long occlusions and small subjects.

It REPLACES every key inside the range, including ones the interactive buttons wrote. Keys outside the range are left alone. Loads the model first if it is not up.

◀◀ButtontrackBackward

Interactive. Track from the CURRENT frame back to the start of the range. Same trade-off as the forward button: fast, resumable, whole-frame, and no future-aware crop.

|◀ButtontrackStepBackward

Track one frame backwards and stop.

▶|ButtontrackStepForward

Track one frame forward and stop, so you can watch it and correct it.

Carrying on from where you stopped keeps the model's memory, which is what carries a point through an occlusion. Jumping somewhere else first re-seeds from wherever the points are on that frame instead, which is what you want after a scrub.

▶▶ButtontrackForward

Interactive. Track from the CURRENT frame to the end of the range, writing a key per frame and carrying the viewer along so you can watch it. Stops at the range end, or when you cancel, and picks up again from where it stopped.

This is the quick-check workflow, not the final one. It has no first pass, so it cannot use the moving crop and it only ever sees the whole frame -- on a small subject in a wide plate that is exactly where Analyse Whole Range wins. What the model is fed here is set by Interactive Quality.

Clear TracksButtontrackClear

Delete every keyframe this node has written, and the points placed on it. Analyse overwrites rather than appends, so this is only needed to start over.

Tracking InputMenutrackInputSpace

How the pixels are encoded on their way to the model. This changes the TRACKING BUFFER only -- the picture this node renders is untouched, and no display transform is baked into anything.

Working Space is the default and what this node has always done: clamp the host's values to 0..1 and quantise. Measured level with sRGB on a normally exposed plate and better on a display-referred one, so it is the safe setting.

sRGB encodes the buffer the way the older Python tracker did -- it rendered sRGB JPEGs, so its tracker never saw a linear pixel. Worth trying on a DARK scene-linear plate, and only there: two and a half stops down it measured 21% more accurate, because clamping linear leaves 44 distinct code values where sRGB leaves 116. On material that is already display-referred it encodes twice, flattens the contrast and makes the model guess -- measured inventing a sighting 530 px from the truth.

Press Dump Tracking Frame on both settings and look at the two files: the right one is the one that looks like the shot.

Default: Working Space (no transform)Options: Working Space (no transform) sRGB (dark scene-linear plates) Working Space, Keep Highlights sRGB, Keep Highlights

Dump Tracking FrameButtontrackDumpFrame

Write the exact bytes the model would be fed for the current frame, as a PNG beside the tracks file, and put the path in Status. Nothing else reads it -- it exists so Tracking Input and Interactive Quality can be checked by eye instead of assumed.

Refine Low-Confidence FramesCheckboxtrackRefineLowConfidence

Include amber low-confidence frames as lost frames when applying Post Tracking Refine. Refinement starts at the first amber frame, not only at the first red frame. The yellow selection ring is unrelated. Choose a refinement method or press Refine Occluded Frames to apply. Turning this off affects future refinements; use Undo to revert an applied refinement.

Default: off

Post Tracking RefineMenutrackOccludedMode

What to do with a marker while it is hidden - behind something, or out of frame. Picking one applies it to the tracks already on the node.

Trajectory Matching (the default): every stretch a marker was not seen is rebuilt. While other markers are still seen it moves with them (their pan, rotation and scale) - forwards from where it was last seen, backwards from where it is seen again. Only while no marker at all is seen does it follow its own path, at its own speed and direction, and the two sides are blended there, so it meets the subject exactly when it is found again, with no jump. A marker that never comes back keeps its motion to the end. While tracking it keys the model's own guess, like Independent; the rebuild runs when you pick it or press Refine Occluded Frames.

Independent: the marker goes where the AI guesses it is. Each marker keeps its own identity, so a turning face or a deforming object comes back on the right spot. Best with the Full Range Tracker (Heavy), whose guess while hidden measured within half a pixel.

Follow Visible Markers: hidden markers move with the ones still visible, so a rigid cluster holds its shape. Use this with the Frame by Frame Tracker when a marker leaves the frame - that model cannot place a point outside the picture.

Interpolate Gap: no keys while hidden; the curve draws a straight line across.

Hidden frames never count as solved, whichever you pick. Change it after tracking and press Refine Occluded Frames to apply it without tracking again.

Default: Trajectory MatchingOptions: Follow Visible Markers Independent Interpolate Gap Trajectory Matching

Refine Occluded FramesButtontrackRefineOccluded

Re-apply Post Tracking Refine and Keep Moving Off-Screen to the tracks already on this node, without tracking again.

Every frame the model reported hidden is rewritten the way the current method says: carried on the visible markers, the model's own coordinate, or left to the curve. Frames the model saw are not touched. So track once with any model, then try the three modes on the result and keep the one that looks right. Works on any track made after this button existed; older tracks have no record of the model's own coordinate and use the keyed position instead.

Keep Moving Off-ScreenSlidertrackCoastFrames

How many frames the points carry on travelling after the LAST one has left the frame, before they stop and wait for the subject to come back.

While any point is still visible the hidden ones always follow the motion of the ones that can be seen. This is only about what happens once nothing is visible at all and there is nothing left to measure: the points coast on the last motion for this many frames, then hold.

0 stops them the moment the last one goes, which can look early - a subject part-way out of frame stops before it has finished leaving. Higher values carry it further out. None of these frames count as solved, so no match-move is ever built from them.

Default: 5

Draw MarkersCheckboxtrackShowMarkers

Draw each tracked point onto the output picture. Useful for checking a track and for rendering a reference; turn it off to pass the plate through untouched.

Default: on

DeviceMenudevice

Automatic picks the fastest available and is what you want. The other two force a specific backend for testing: CUDA is the GPU, CPU is many times slower and cannot run tracking on the ONNX engine at all.

Default: Automatic (fastest)Options: Automatic (fastest) CUDA CPU

TracksSlidertrackCount

How many tracks Analyse produced. Read-only.

Default: 0

The list of tracked points, and the exports to a Tracker or a CameraTracker.

anrClassicalPointListPanelanrClassicalPointList

The list of points, in two groups: AI points and Classical points. Selecting a group or a point makes that tracking mode active. Ctrl- or Shift-click selects several points to delete or export; Delete removes them. Classical points can be renamed, switched on and off, and given their own Pattern, Search and Match values in the list. The buttons below add a point, delete the selection, go to the reference frame, and track or analyse.

Default: empty

Export to TrackerButtonanrExportTracker

Build a native Tracker4 from these tracks. Occluded frames arrive marked as interpolated rather than solved.

Export to Camera TrackerButtonanrExportCameraTracker

Build a CameraTracker with these tracks as its User Tracks, ready to solve. Its table has no per-frame enable column, so occluded frames arrive as genuine GAPS in the curve rather than as keys marked unsolved. It does not press Solve -- that wants survey points and a lens model, and is your call.

T (translate)CheckboxanrExportT

Set this column for every exported track. Rotate and scale need at least two tracks to mean anything.

Default: on

R (rotate)CheckboxanrExportR

Set this column for every exported track. Rotate and scale need at least two tracks to mean anything.

Default: off

S (scale)CheckboxanrExportS

Set this column for every exported track. Rotate and scale need at least two tracks to mean anything.

Default: off

Separate tracker per maskCheckboxanrExportPerMask

OFF exports the usual way: one node holding every track.

ON builds a node per mask region instead, each holding only the points that sat on it -- for separate objects that each want their own match-move, rather than one solve averaging across things that are not moving together.

Only does anything when the points came From Matte and the matte has more than one separate region in it. A region is a connected white area of the picture on the Matte input: the node is handed a rendered matte, not the Roto's list of shapes, so two shapes that touch or overlap count as one.

Default: off

Export Selected Points OnlyCheckboxanrExportSelectedPoints

When enabled, both Export to Tracker and Export to Camera Tracker use only the selected rows in Points. Select the AI group to export all its points. With no selected tracks, nothing is exported.

Default: off