0:00 · Exploring a Gaussian-splat reconstruction
A Hypercliq title introduces Splat Viewer. The walkthrough begins inside a Gaussian-splat reconstruction of an unfinished apartment, looking around walls, exposed ceiling details, and a balcony doorway.
Explanatory cards describe real-time viewing using Rust and Vulkan, WASD movement, mouse look, and adjustable movement speed, field of view, and splat size. A visible-splat count is highlighted as the viewpoint changes. Another card attributes visibility culling and depth sorting to the GPU.
The rendering switches from a photographic-looking splat to individual points and back. The overview and position readout show the camera’s location, direction, and the extent of the capture.
0:35 · Notes and object display options
A list of viewpoint notes appears and a note is selected. Cards explain that notes retain the camera pose so the user can return to that view, and that notes may be typed or dictated. The dictation card states that Whisper transcribes on the device without sending data away.
The object-overlay menu then offers flat 2D boxes, 3D edges, and shaded boxes for displaying detections.
0:53 · Capturing panoramas and labelled views
Capture menus show panorama resolution choices and labelled dataset export. Cards describe panoramas up to 8K and captured views containing RGB images, per-pixel point indices, and world XYZ coordinates. The File menu also presents loading options for point clouds, bounding boxes, camera paths, and reference photographs.
At approximately 1:10, a panorama captured from the viewer fills the screen. The view pans across the apartment’s walls, ceiling, floor, and balcony doorway. Its explanatory card describes an equirectangular PNG output up to 8192 × 4096 pixels.
1:19 · Loading and exploring a LiDAR scan
A loading progress dialog appears, followed by navigation through a dense LiDAR point cloud. Cards identify the example as a 974 MB scan containing 36.5 million points and describe loading on a background worker.
The scan is shown as surface elements whose spacing, according to the overlay, is derived from the captured data. The rendering switches to raw points to inspect the sampling.
At approximately 1:44, detected floor cells are highlighted in green as a walkable-area overlay, showing potential standing locations. The overlay is subsequently removed.
1:53 · Following the recorded capture route
An orange recorded camera trajectory appears in the scene. Playback follows the sensor’s route through the scan. Cards explain that a route can be played at an adjustable speed or scrubbed to a chosen moment.
A reference-photo panel changes with the viewpoint. Its card describes matching the nearest photograph from a COLMAP dataset.
2:22 · Managing a larger scan
Another scan is introduced as containing 216 million points on an 8 GB GPU. Cards explain that the full scan is held in host memory while octree streaming supplies the current view to GPU memory.
A highlighted GPU-memory budget control changes the resident point set. The explanatory text says the finest level of detail is prioritised where the camera looks.
2:34 · Extracting building geometry
Detected walls, floors, and ceilings receive orange, blue, and cyan class colours. Bordered translucent polygons outline the surfaces, and the overview becomes a floor plan.
Wall faces are then shown with yellow door frames and cyan window frames. Cards describe each wall as a single quadrilateral extending between corners. The room stage describes enclosed spaces with their areas, ceiling heights, and boundary polygons.
3:02 · Exporting and viewing the floor plan
The export menu is shown. An explanatory card describes DXF output with wall thicknesses, dimensions, door swings, and room labels, and also mentions JSON output.
At approximately 3:10, the exported DXF floor plan fills the screen. It contains walls, dimensions, door swings, windows, and labelled rooms with areas. The accompanying card describes opening the output in CAD software.
3:19 · Measuring distances and areas
A laser-style tool displays the distance to the surface at which the view is aimed. Selecting two points creates a labelled measurement line, first between walls and then from floor to ceiling.
Cards explain horizontal and vertical measurement components and intersection with fitted planes. One card claims sub-centimetre accuracy on flat surfaces.
Measurements then snap to detected wall corners. The card specifies a 10 cm snapping range. Further points form a boundary; closing the boundary produces an area and perimeter readout.
The measurements remain drawn in the scene and listed in the panel. At approximately 3:46, a card says they are saved alongside the scan and included in exports as a DXF MEASURE layer and IFC annotations.
3:52 · Training a Gaussian splat
The Gaussian-splat training command opens a configuration dialog using the scan and posed photographs. Cards identify Brush as the trainer and describe initialising flat Gaussians on the scanned surfaces at 5 cm spacing, rather than starting from sparse structure-from-motion points.
A training progress panel shows iteration, splat-count, and image-quality information. A card says the scan is streamed out of GPU memory during training.
At approximately 4:16, a later card describes a result after twenty minutes, names 20,000 iterations as the default, and says the result can be loaded with one click.
4:27 · Exploring the trained result
The trained Gaussian splat opens and the camera explores the apartment again. A card explains that the result uses the scan’s coordinate frame and that a haze control suppresses faint Gaussians.
The final message connects three outputs from one capture: LiDAR for measurement, a floor plan for CAD and BIM, and a splat for a photographic walkthrough.