Core Concepts
GPU and Dependencies
What ships inside the add-on, what downloads later, and how to check your machine is set up right.
Almost everything Dope Track needs is already inside the add-on. That's why the download is big, and why tracking and solving work the moment you enable it.
What's Bundled
| Component | Used for |
|---|---|
| COLMAP | Finding and matching features, solving cameras |
| GLOMAP | Global mapping inside Auto Track |
| OpenMVS | Densify and Photogrammetry |
| FFmpeg | Reading video and pulling frames out of it |
| OpenCV (CUDA build) | Image processing, lens calibration, stabilizing |
| rawpy / LibRaw | Camera RAW stills for photogrammetry |
| CUDA runtime | Shared by all of the above |
The bundled OpenCV is a custom build with CUDA support for several generations of NVIDIA cards, from Pascal up to Blackwell. You don't need to match a specific model.
What Downloads Later
Only the AI masking feature needs anything extra:
| Component | Size | When it downloads |
|---|---|---|
| AI runtime (PyTorch and friends) | about 2 GB | First time you use the AI Mask node |
| AI model | about 900 MB | First time you generate a mask |
Both have their own buttons on the AI Mask node, Install AI Runtime and
Download AI Model. Neither runs until you press it. If you never touch AI masking,
nothing downloads and nothing else is affected.
The GPU Toggle
Open the node editor sidebar and look at the Device panel. There's a global
Use GPU switch, along with the GPU and driver it detected.
Turn it on and nodes use CUDA where they can. Turn it off and everything falls back to the CPU. That's slower, but it's a useful test when you suspect the GPU is the problem.
With the switch on, a dropdown underneath picks which card does the work. If your machine has more than one, check it: the automatic choice is whichever card Windows enumerated first, which is often the display adapter rather than the fast one. See Device.
An NVIDIA card is strongly recommended. Solving still runs on CPU without one, but dense reconstruction and AI masking become too slow to be practical.
Checking Your Setup
Go to Edit > Preferences > Add-ons > Dope Track and press Run Diagnostics.
It takes about a minute and tests the whole machine: memory and disk space, your NVIDIA driver and VRAM, bundled OpenCV with real CUDA operations, rawpy, an FFmpeg encode, COLMAP and GLOMAP starting up, OpenMVS, and a real GPU feature extraction on generated frames.
You get a plain text report on your Desktop with a .json copy beside it. Each check
says PASS, FAIL, WARN, INFO or SKIP, so if something's broken the report
names which part and why.
Run it once after installing. Run it again before reporting any problem, and send the report along.