Train
Train the splat with the built-in Brush trainer, or hand the alignment to LichtFeld Studio or Postshot.
Training is the final step: it turns the sparse model from Alignment into a Gaussian splat you can view and share. For the short version, see the First Run walkthrough.
What alignment hands off
A finished alignment is a sparse COLMAP model — camera poses plus a sparse point cloud — ready to be handed to a 3D Gaussian-splatting trainer. Nothing about the reconstruction is baked in yet; the trainer builds the actual splat from these camera positions and your masked frames.
Choosing a trainer
Pick the trainer at the top of the Train tab:
| Trainer | Notes |
|---|---|
| Brush (default) | Bundled with the app — nothing to install. Free and open source. Runs on any GPU. |
| LichtFeld Studio | One-time payment, or free build from source. NVIDIA only. |
| Postshot | Easy GUI, monthly subscription. NVIDIA only. Needed for HDR training. |
| No Train | Stop after alignment. Everything a trainer needs is in 0_train_data. |
LichtFeld Studio and Postshot need CUDA, so on AMD and Intel GPUs only Brush is offered.
Brush
Brush comes with the app and trains inside the Train tab. With Live view on, you watch the splat form while it trains. Between runs, the same panel shows the clip's newest splat: orbit with the mouse, or click into it and fly with WASD. The ‹ / › button hides the settings to give the viewer more room.
| Setting | Default | What it does |
|---|---|---|
| Total Steps | 30 k | Training length, in thousands of steps. |
| SH Degree | 3 | Colour detail by view angle. Lower = smaller file, less shine and reflection. |
| Max Resolution | 1920 px | Training images are scaled down to this size. |
| Max Splats | 3000 k | Upper limit on the number of splats (3000 k = 3 million). |
| Export Every | 5 k | Save a .ply every N thousand steps. |
| Export Name | gaussian_splat_{iter}.ply |
{iter} becomes the step number. |
| Live view | On | Watch the splat train. Off trains a little faster. |
If Brush runs out of GPU memory, lower Max Splats — the log suggests a value — and train again.
LichtFeld Studio
Point the app at your own LichtFeld Studio install. Version 0.5.4 or newer is recommended: it adds Exposure correction (replacing per-camera appearance and the bilateral grid), extra export formats (.sog, .spz and an .html viewer) and Normal loss for flatter walls and floors. Older versions still work; the app detects the version and only shows the options it supports.
After training
The trained .ply (or the trainer's own format) is saved in the project's
training_output folder. View it in any Gaussian-splat viewer, embed it on the
web, or bring it into other 3D tools.
To train on fewer frames without aligning again, use the Camera filter.
Project folder
Each clip gets a project folder next to the input video, named after the clip.
From 1.2.0, the root holds only what you need afterwards. Everything else is
kept in pipeline/:
| Item | What it is |
|---|---|
0_train_data/ |
Training images, masks and the sparse model. Point any trainer here. |
0_train_data_hdr/ |
HDR (EXR) training images, only when HDR is on. |
mesh/ |
The collision mesh, only when it's on. |
training_output/ |
The trained splats. |
pipeline_log.txt |
The processing log. |
run_settings.json |
The clip's settings. The command line reads and writes it too. |
pipeline/ |
Working files: extracted frames, masks, the COLMAP database and intermediate models. |
Once you're happy with a splat, you can delete pipeline/ to free most of the
disk space. Re-aligning, Run Mesh, and adding detail photos or stills then
need a fresh extract. Don't delete it in 3DGUT projects, where the training
images live inside pipeline/.
Projects made with older versions are not moved — they keep their working files in the root and carry on as before.