2D point tracking and 3D camera solves for VFX, on your own machine. Drop in a clip, solve it, and send the result straight to Nuke, After Effects, Blender or USD.
Each pipeline has its own tab, its own settings and its own export set. Both read from the same media pool and write into the same scene folder.
Meta's CoTracker3 follows hundreds of points through a shot at once, with confidence and velocity filtering to throw out the ones that slip.
A full structure-from-motion solve that recovers the camera path, lens, and a sparse point cloud of the set, with an optional environment mesh.
One window, two tabs. Settings on the left, the media pool or viewer on the right, and the licence line along the bottom.
One solve writes every format below into the scene folder. A Timeline Start control shifts all keys onto the shot's real frame range, and the Nuke tracker is copied to the clipboard so it pastes straight into a script.
| 2D tracks | What you get |
|---|---|
| Nuke .nk | Native Tracker4 node, one per layer. Corner-pin mode adds a CornerPin2D. Masks export as a Roto node. |
| After Effects .jsx | ExtendScript that builds tracked null layers, or a corner-pinned layer. |
| Blender .py | Empties following each point, or a quad surface for corner pins. |
| CSV / JSON | Raw per-frame positions and confidence for anything else. |
| Overlay .mp4 | Motion-trail preview for review. |
| 3D camera | What you get |
|---|---|
| Nuke .nk / .chan | Camera script and animation channel file. |
| Blender .py / .blend | One-click import of camera, image sequence, point cloud and a RANSAC ground plane. A baked .blend is written when Blender is installed. |
| Alembic .abc | Baked through headless Blender when it is found. |
| USD .usda | Pixar USD camera and points. |
| PLY / JSON | Sparse point cloud, optional environment mesh, and the full solve as JSON. |
Everything is bundled: Python, PyTorch, COLMAP, FFmpeg and the CoTracker3 weights. Nothing is downloaded at run time.
Windows 10 or 11, 64-bit. Installs to Program Files, so setup asks for admin once.
An NVIDIA GPU with CUDA makes both engines fast. Without one, tracking and solving fall back to the CPU.
Automatic VRAM chunking sizes the tracking window to free memory and halves it on an out-of-memory error.
About 6 GB for the install. Frame caches and solves are written next to your footage.
The installer is split into a small setup exe plus numbered .bin data files, because the bundled AI model and PyTorch runtime are larger than a single Windows installer can hold.
-1.bin, -2.bin, … file.Meta AI's transformer point tracker drives the 2D tab. The bundled weights are released under the Creative Commons BY-NC 4.0 licence, which permits non-commercial use only.
The structure-from-motion and multi-view stereo pipeline behind the 3D solve, including its vocabulary tree for loop detection.
Frame extraction, GPU inference and the interface. Blender, when installed, bakes Alembic and .blend exports.