Learn — what is gaussian splatting?

A place, rebuilt out of light.

No triangles. No texture maps. A gaussian splat is millions of oriented ellipses hanging in space — each one holding a position, a size, a tilt, a color, and how solid it is. Together they carry the light your camera saw, and the room comes back exactly as it stood.

01 · How it trains 02 · Splats and meshes 03 · The ecosystem

GAUSSIANS318 DENSIFYING DRAG TO ORBIT

FIRST GUESS SHARP

Drag the rail: same scene, more gaussians. The right end is nine thousand of them. A real capture runs into the millions, and keeps getting sharper the whole way.

Scroll

Step · 01 — the loop

It tunes itself against every frame.

Nothing measures a splat into existence. MacSplat solves where your camera stood, scatters a rough cloud of ellipses into the scene, renders a guess from one of those viewpoints and holds it against the real frame. The difference tells every ellipse exactly how to improve— a little left, a little wider, a little greener, a little more solid. Then it goes again. Tens of thousands of times, until the render and the photograph are the same picture.

Why — the renderer runs backwards

Splatting a gaussian onto a pixel is a smooth bit of arithmetic, so it can be differentiated: for every pixel that came out too red, you can compute exactly how much each ellipse's position, size, tilt, color and opacity contributed to that redness. Slide all of them a hair down that slope and the next render is closer. That's the whole trick — the picture is the supervision, and nobody has to model anything.

FRAMES SFM — CAMERAS SOLVED SEED CLOUD TRAIN · DENSIFY · PRUNE SPLAT

ITER0 GAUSSIANS9 PSNR0.0 dB TRAINING

This is the real loop, at 80×80 pixels.

Nine fat ellipses, one target image, and gradient descent doing the rest — running here, not a video. Watch the error panel drain and the fit climb past 35 dB. A blob flashes green when it splits: an ellipse too coarse for the detail it covers gets cut in two, invisible ones get deleted. That is how a splat gets from a few thousand to a few million without anyone deciding where they go.

Where the first ellipses come from

Structure from motion matches features across your frames, solves the camera path, and drops a sparse point cloud. Every one of those points becomes a gaussian — that's the guess training starts from.

TOO COARSE → SPLIT IN TWO

Split, and prune

An ellipse that is blurring over real detail gets cloned and shrunk; one that ended up transparent or enormous gets thrown away. The count is never fixed — the scene decides it.

COLOR PER ANGLE

Color that knows where you stand

Each gaussian stores its color as spherical harmonics — a little function over directions — so a wet floor or a chrome kettle brightens as you move, instead of staying a flat painted patch.

Step · 02 — two approaches

Two ways to rebuild the same room.

Photogrammetry and splatting start from the same footage and the same solved cameras, then part ways. A mesh commits to geometry: it decides where the surface is, stretches triangles over it and bakes color into them — something you can edit, collide with, animate, print. A splat never builds a surface at all; it fits translucent ellipses until the render matches the photograph — the light of the place, kept as it was. Different logic, different technology, different answers. Which one is right depends entirely on what you plan to do next.

{{ meshChip }} {{ splatChip }} SURFACE — SOLID, EDITABLE, DECIDED LIGHT — SOFT, VIEW-DEPENDENT, OBSERVED

One capture, two representations.

Move your pointer across to slide the divider; drag to orbit. Same frames, same cameras. On the left, geometry gets decided: clean, solid, light enough to hand to a game engine — and the curve of the handle, the rim of the glass and the glaze on the jug all get resolved into faceted surface, because surface is what a mesh has. On the right, nothing gets decided: the ellipses simply agree on what the frames showed, reflections and all.

Where they differ Mesh Splat

Made oftriangles + baked textureoriented 3D gaussians

Flat walls, floors, boxes✓clean and cheap~can look grainy

Leaves, hair, wire, mesh fabric~resolved into surface✓kept, wisp for wisp

Glass, chrome, wet stone~baked as one fixed color✓shifts as you move

Physics, collision, 3D print✓it is a surface✗nothing to collide with

Editing, rigging, animation✓every tool on earth~young, improving fast

Weight on disk✓megabytes~hundreds of megabytes

Looking like the real place~convincing✓often indistinguishable

Capture to result~modelling and cleanup✓one pass, no cleanup

Step · 03 — the ecosystem