How it works

A Gaussian splat describes a scene as many small, soft, coloured blobs, each with a position, size, orientation, opacity and a colour that can change with the viewing direction. Training fits those blobs to pictures of the scene until the splat looks like the pictures from every angle. SplatPilot takes those pictures in Blender, trains on your graphics card and gives you the result as a native Blender 5.3 splat object.

Create Splat runs four steps: views, rendering, training, and a splat object in your scene.

1. Views: where the cameras go

A view is one camera position and direction that SplatPilot renders. You choose where views come from with Source in the Views panel:

• Automatic (the default): SplatPilot plans the views from your subject.
  • An object (a model, a prop, a character) is seen from all around: views sit on three spheres around it, close up, framing it and wider, all looking at it.
  • An environment (a room, a building, a street) is seen from the inside: views are placed by the surfaces themselves, wherever the scene camera can get to, so that everything is seen at the same detail; some stand at eye height on floors and the ground and look around. Cheap geometry views show where the surfaces are, and the views rendered in full are picked among them so that together they see every surface from as many directions as possible, as many views as it takes to see the space fully.
• Rigs: camera rigs you place and shape yourself.
• Cameras: the cameras of a collection you pick.

A few views are kept back as test views: the splat is not trained on them, so comparing it with them tells how well it holds up from angles it never saw. See Views for the details.

2. Render: pictures of your scene

SplatPilot starts a second Blender in the background and renders the views there, with the scene’s render engine (Cycles or EEVEE; Cycles when the scene uses another) and the render devices from your preferences. That Blender opens your saved file, or a temporary copy when the file has unsaved changes, and sets its render settings in its own copy only: your file and your scene’s settings stay as they are.

Each view becomes a picture plus the data training reads: linear colour and every pixel’s position and normal. They are rendered with a transparent background, without motion blur or depth of field, at the current frame. For an object chosen by selection or collection, the views show only the subject by default (Subject Only): other objects still light it and show in its reflections. Rendered views are kept, so a later run with the same views and render settings does not render them again.

3. Train: a splat on your GPU

SplatPilot’s own training engine runs next to Blender, in a separate process, on your graphics card through Vulkan. It starts from points on the rendered surfaces and grows the splat up to its splat budget (Max Splats) over the training steps. Training is tuned for the way Blender draws splats.

While it trains, you see:

• a preview of the splat in the viewport, refreshed every 15 seconds;
• the score so far in dB (PSNR): how closely the splat matches the views it trains on;
• the time left.

The engine saves a checkpoint every 5 minutes. If the training fails (a GPU reset, out of memory), Resume goes on from the last checkpoint.

4. Splat in your scene

The finished splat is imported with Blender’s own PLY importer into a Splats collection as a Gaussian splat object, which Blender 5.3 draws and renders like any other object. SplatPilot reads part of it back to check that it arrived exactly as trained. The import is one undo step: Ctrl+Z removes the splat again.

The splat’s object keeps a record of its training: its subject, its views, its checkpoint, its steps, its score and its recipe (the quality and settings it was made with, also written as a .json beside its trained file). That record is what lets you Improve it, Train Again with other settings or Use Its Settings later. See Results.

One step at a time

Create Splat runs all of this in one go. The three buttons below it run one step each:

Button	What it does
Views	Places the views and shows them in the viewport, so you can check them before rendering. An environment renders its geometry views first (a minute or a few); an object’s views appear at once.
Render	Renders the views (placing them first if needed), then stops.
Train	Trains a new splat on the views rendered last, whatever is selected now, with the current Training settings. The splat is named after the subject those views were rendered for.

Each step reuses what an earlier one made for the same settings: change the training steps and Train again without rendering again, or change the samples and Render again. Create Splat skips the steps that are already done. The buttons show where each step stands: a check mark when it is done for the current settings, red when it was done for other settings (see Where each step stands). After Render, the panel’s status line reads, for example, “100 views of Suzanne rendered”; once the settings differ from those views, “Train uses the 100 views of Suzanne rendered last, not the views set now”.

What SplatPilot adds to your file

• A Splats collection with your splats (and the preview while one trains).
• A SplatPilot collection for the helpers you create: rigs, focus points, baked cameras, and the SplatPilot View camera that Look Through uses.

On disk, SplatPilot keeps rendered views, training runs and trained splats in a splatpilot folder next to your file or in your Blender Project. See Storage and Blender Projects.