Gauss-Opt: WebGPU 3DGS Optimizer

Gauss-Opt

3D Gaussian Splatting in the browser: video or photos → cameras → training

Initializing WebGPU...

Keys & mouse (H or Esc to close)

Camera
drag	look around
wheel / pinch	move forward / back
↑ / ↓	fly forward / backward
← / →	turn left / right
E / Q	climb / descend (released: hold altitude)
Space	re-center: level, brake, face the scene
[ / ]	level: turn the picture 1° (⚙ view > Level)
Demo camera (G to toggle)
	slow flight along the capture path, wandering a little
drag / wheel / flight key	take over
Cameras
click frustum	show its photo
double-click	fly to that view
General
G	demo / drone camera
H or ?	this help
drop a file	video, gs/ folder, .ply / .splat / .gosnap, .gosfm

Start from a video or a set of photos (JPEG), or open a gs/ folder (sparse/ + images/) or a checkpoint. Files can also be dropped on the page.

Open gs/ folder… Open checkpoint…

How to shoot a photo set

Walk, don't pivot: take each photo from a new spot. Photos turning from one spot give no depth.
Small steps: from one photo to the next, turn at most 30° and move at most half the distance to the subject. Around an object, that is 12 photos per full turn.
Overlap: every part of the scene in 3 photos or more: the more photos see a part, the better it looks from in between.
Lenses and zoom can be mixed (a camera model each), and so can portrait and landscape; a lens with fewer than 5 photos is not calibrated on its own.
Something to match: avoid photos of only a plain wall, a window or the sky.

Open a video or photos (Project, or drop them on the page).

Open a video first.

What every camera sees alike, wherever it stands (no parallax: the sky, far away) goes onto a dome behind the splats, made from the photos before the training and fixed while it runs; the zenith and what no camera saw are filled in. The splats are left for what is near. Indoors, or with little sky: no dome. Its fill: the sky only, or a panorama of all the cameras saw (the far things as billboards).

Splats

The frame timeline of a video: sharpness of the sampled frames, keyframes by quality, loops.

sharpness (bright: keyframe, red: too blurry) good doubtful bad not registered (faded: separate map) selection (drag / click): what Repair acts on loop suspect loop: relocated on the map, one frame lands far from its pose

Log