Repository navigation
First release of DBS_SplatTrain: a Nuke 17 node that trains a Gaussian splat from a shot you have already tracked in Nuke, and brings it back aligned with the track camera.
What it does
- One button, Export + Train. It renders every Nth frame of the plate, writes the CameraTracker camera and point cloud as a COLMAP dataset, trains the splat with Brush in a separate process and shows it in the node. Nuke stays usable meanwhile; progress shows in the panel and in the node's label.
- Inputs:
plate,cam(tracked camera),points(tracker point cloud) and an optionalmask(a roto of what moves, which Brush leaves out). - The node renders the splat through
cam, back in the working colour space, so it lines up with the plate. - Skip blurry frames (on by default): in each group of N frames it exports the sharpest one, which leaves motion blur out of the training.
- Create 3D setup: puts the splat next to the node as GeoImport -> SplatRender -> OCIOColorSpace, so you can edit it in 3D (e.g. remove parts with a FieldShape).
- Proxy: a lighter copy of the splat (500k / 100k / custom) for a faster viewer. Turn it off for the final render.
- Step-by-step buttons (1. export, 2. train), a stop button, and the command line tools for running it by hand.
- Works in Nuke Non-commercial.
Requirements
- Nuke 17 (it uses the new splat nodes), Windows, and a GPU that runs Brush (tested on an RTX 4060 Ti, 8 GB).
- Brush v0.3.0 (Apache 2.0), downloaded separately. Use v0.3.0: later versions changed their command line.
Install
- Download the zip below (Source code) and unzip it, e.g. to
C:/nuke_tools/DBS_SplatTrain. - Unzip Brush's
brush-app-x86_64-pc-windows-msvc.zipso that you getDBS_SplatTrain/bin/brush-v0.3.0/brush_app.exe(or set the node'sbrushknob to wherever you keep it). - Add to
C:/Users/<you>/.nuke/menu.py:nuke.pluginAddPath("C:/nuke_tools/DBS_SplatTrain/nuke") import DBS_SplatTrain DBS_SplatTrain.install_menu()
- Restart Nuke. The node is under DBS > DBS_SplatTrain.
Full notes: README · CHANGELOG
MIT licence. Only tested on Windows so far.