Training settings
The Training panel (closed by default) sets how the splat is trained. Create Splat, Train and Train Again use these settings. Create Splat takes SH Degree, Opaque Splats, Start Detail, Focus Detail and Preview Every when you click it, so changing them while the views render applies to the next run. Settings marked Advanced show with Advanced Settings on (Setup panel), and also with it off once they are set differently from their default.
Settings
Section titled “Settings”| Setting | Default | What it does |
|---|---|---|
| Steps | Unticked: the quality’s | Training steps. Unticked, the field shows the quality’s: 7,000 for Draft, 30,000 for Standard and High. More steps refine detail and take longer. |
| Max Splats (k) | Unticked: the quality’s | The splat budget in thousands: 1500 means 1.5 million splats. Unticked, the field shows the quality’s: 300, 1,000 or 2,500. |
| SH Degree Advanced | Auto | How much a splat’s colour may change with the viewing direction. See below. |
| Opaque Splats Advanced | Auto | Whether training pushes every splat to fully opaque or gone. See below. |
| Start Detail Advanced | 0.5 | Where the splats start: on small faces and fine visible detail (ropes, leaves, edges), or spread evenly over the surfaces at 0. See below. |
| Focus Detail Advanced | 3 | How much harder training works around focus points. Available when the scene has focus points. |
| Preview Every Advanced | 15 s | How often the splat so far is shown in the viewport while it trains. 0 shows it only at the end. |
Steps and Max Splats (k) are heading checkboxes: ticking one or editing its number makes it yours, unticking returns to the quality’s. See Override any number.
The panel ends with a line on what training does for this capture:
| Capture | For example |
|---|---|
| An object | “Objects: SH degree 3” |
| An environment | “Environments: SH degree 3 · opaque splats · 4,667 more steps on geometry views” |
The SH degree is named only while SH Degree is Auto, opaque splats only while Opaque Splats is Auto, and the extra steps only with geometry views. With nothing to name (an object with an SH degree of your own), there is no line. Environments train opaque splats: training pushes every splat to fully opaque or to nothing. See Opaque Splats.
Steps and the splat budget
Section titled “Steps and the splat budget”Each splat starts at a point on a surface the views show (see Start Detail). An environment trained for 20,000 steps or more (Standard and High) starts with its whole budget, up to a million splats; any other training starts with up to 100,000 and grows them to the budget. Every step refines them: training moves them, reshapes them and moves the ones that do nothing to where detail is missing. A larger budget can hold finer detail but needs more GPU memory and time; the estimate in the Create Splat panel shows both.
In an environment with geometry views (the default), training also spends steps on them, on top of the steps you set: they make up 40% of the total. A Draft splat of a room therefore trains 11,667 steps (7,000 + 4,667), as the estimate’s Splats row and the Results panel show.
SH Degree
Section titled “SH Degree”A splat’s colour can change with the viewing direction, described by spherical harmonics (SH). The degree sets how much:
| SH Degree | Colour per splat |
|---|---|
| Auto (default) | 3: objects and environments are seen from all around |
| 0 | One colour, the same from every direction (the smallest file) |
| 1 | Gentle changes with the direction |
| 2 | Stronger changes |
| 3 | The most: sheen and soft reflections |
Higher degrees need every surface seen from many directions. With views from only a few places, they make up colour for directions no view saw, which shows as flares when you look from new angles. An environment’s views are placed by its surfaces, which they see from all around, so Auto uses 3 there as for objects: on a harbour model degree 3 did 1 dB better from around it and 1.4 dB from down at the water.
Raise it when a room’s views see its shiny parts from many sides. Lower it when new angles show colours that do not belong.
Opaque Splats
Section titled “Opaque Splats”| Opaque Splats | Splats |
|---|---|
| Auto (default) | Opaque for environments, free for objects |
| On | Every splat fully opaque or gone |
| Off | Splats keep the opacity training gives them |
Opaque splats render solid in EEVEE: in our test room they scored 1.5 dB higher there. On objects they cost 0.1 to 0.4 dB, so Auto leaves objects free. See-through materials (glass, fog, smoke) cannot stay see-through as opaque splats: in the Barbershop (fog and glass), free splats scored 0.2 dB higher on views training did not see, at Draft. Off suits an environment full of them. A setting other than the kind’s own makes a training of its own: Train turns red after the change.
Start Detail
Section titled “Start Detail”Every splat starts at a point on a surface the views show. Up to half of the points come from the faces of the objects the views show: drawn on the faces, kept where a view sees them, moved onto the surface as rendered (displacement included) and coloured as the views show it there. The rest come from the rendered pixels. Start Detail decides where both gather:
| Start Detail | Points from faces | Points from pixels |
|---|---|---|
| 0 | Spread over the faces by area | Spread evenly over what the views show |
| 0.5 (default) | Half by area, half the same for every face | More where the views show detail |
| 1 | The same for every face, whatever its size | Most where the views show detail |
Small faces, such as ropes, cables, railings and a finely modelled object’s detail, take few pixels, so points taken from pixels alone seldom land on them; drawn per face, they get their own. Detail without faces, such as leaves on cards, textures and displacement, shows in the pixels: colour contrast, edges against what lies behind, and bends in the surface.
When an environment starts with its whole budget, each splat is a flat disc lying on its surface, so walls, floors and the ground are covered from the first steps. Smaller starts use small round splats, and training grows more where detail is missing: a short training (Draft) or an object does better that way. The points on faces are drawn when the views render, by area and by face, and Start Detail mixes the two when training starts: changing it needs no new render.
At Standard, our harbour scene scored 32.1 dB on the test views against 31.8 dB when training started from 100,000 points taken from pixels, with 0.4 dB more on the bushes’ leaves, and our test room 40.5 against 40.1 dB. A setting other than 0.5 makes a training of its own: Train turns red after the change.
Focus Detail
Section titled “Focus Detail”With focus points in the scene, the pixels that show a focus area count this many times, and more of the splats gather there. Focus areas come out sharper and the rest a little softer. In our test room, a focus area came out 1.4 dB sharper at 3 with nothing lost elsewhere. 1 turns it off. It is separate from the Views panel’s Focus Weight, which picks views.
Preview Every
Section titled “Preview Every”While a new splat trains, a preview object named <name> (training) shows the splat so far in the
Splats collection. It is replaced every 15 seconds by default and removed when the finished
splat arrives. The refresh button at the end of the training’s progress bar shows the current
state at any time.
While a splat trains on (Improve), it stays as it is until the new result arrives.
Training further
Section titled “Training further”| Action | Where | What it trains |
|---|---|---|
| Train | Create Splat panel | A new splat on the views rendered last, with these settings, named after the subject they were rendered for |
| Improve… | Results | The same splat from its checkpoint, with more views (Automatic views only), more steps, or both; with Keep the Current Splat, into a new splat |
| Train Again | Results, the splat’s menu | A new splat on that splat’s views with these settings, named after it (Suzanne Splat.001); the old one stays to compare |
Checkpoints and failures
Section titled “Checkpoints and failures”The engine saves a checkpoint every 5 minutes, and when you pause. If a training fails, for example after a GPU reset or when memory runs out, the panel says why, with Open Log at the end of the line, and if a checkpoint was saved, Resume goes on from it.
If memory runs out, the message reads “not enough memory (GPU or system)”: lower Max Splats, the view Count or the image size, or close other programs while it trains. A VRAM limit in the preferences makes the estimate warn before you start.
GPU memory
Section titled “GPU memory”Training keeps the splats and the views’ data on the GPU. When the views would take more than 1 GB there, SplatPilot keeps them in system memory instead and copies each step’s view to the GPU as it goes, at the same speed. Training a room in our tests took at most about 0.7 GB of GPU memory at Draft, under 1 GB at Standard and 2.8 GB at High. The estimate’s VRAM row shows what your settings need, out of your card’s memory once Test GPU has found it (“0.6 of 12 GB”).
Training time
Section titled “Training time”The estimate’s training time comes from your GPU’s speed: Test GPU in the preferences measures it, and every finished training refines it. The first training on a machine also compiles the engine’s shaders, a minute or two at “Preparing the engine…”, unless Test GPU did that already.