What is Gaussian Splatting? — one analogy

1The analogy

Building a mosaic mural means setting colored tile pieces down one by one. A single piece is just a blob of color, but once tens of thousands of them find their places, stepping back turns them into a single picture.

Gaussian splatting sticks this same kind of tile not on a wall but floating in empty space. Every piece has its own position, size, tilt, and color. The edges aren't cut clean with a knife — they fade out toward the outside. Because those faded edges overlap each other, no gaps show between the pieces.

Move your viewing spot, and the pieces overlap in a new way. Front pieces hide back pieces, and a piece that looked flat turns elongated as you circle around. Nobody re-glues the tiles — it's the same pile of tiles, just seen from a different spot.

2In detail

A pile of pieces comes out of several photos

It starts with dozens of photos of one subject, shot from many angles. The photos get compared against each other to work out where each one was taken from — the same corner showing up across several photos is the clue used to trace back the camera positions.

Once the camera positions are fixed, pieces get scattered roughly through space to start. At first they sit in the wrong places with the wrong colors. From this state, the system renders a picture of the pile from the same spot the original photo was taken, and compares it against the real photo.

Wherever they differ, pieces get moved, resized, and recolored by that much. Repeat this tens of thousands of times, and the pieces settle snugly onto the object's surface. Nobody places the pieces by hand — the photos decide where each piece belongs.

What one piece holds

A piece isn't a sphere — it's a squashed ellipse. It carries values for where it sits, how big it is and in which direction it's stretched, what color it is, and how much shows through from behind.

A flat surface like a wall or floor only needs a handful of broad, flattened pieces. A spot with lots of fine detail, like leaves or wire mesh, gets packed with tiny pieces instead. Since a piece can reshape itself however it needs to, simple surfaces use few of them and complex ones use many.

Drawing means pressing and painting over

To display a scene, every piece gets flattened toward the viewer and stamped onto the screen. Then they're painted on in order, farthest piece first. If the front piece is translucent, the color behind shows through and blends in; if it's opaque, the color behind stays hidden.

This approach doesn't have to trace the path light took to every point on the screen, so it's very fast. That's why spinning a scene around with the mouse doesn't stutter. Being viewable in real time is exactly why this approach caught on so quickly.

Piece count grows, shrinks, and gets refined

The number of pieces keeps shifting while the scene is being built. Wherever a piece doesn't match the photos well, it gets split in two, or a new one gets added. Pieces that go nearly transparent and stop contributing anything get deleted.

By the time it's finished, the piece count runs from the hundreds of thousands into the millions depending on the scene. Since each piece carries several values, the files get fairly large. But because the pieces are a visible pile, you can later select and delete or move just one part of the scene.

Where it splits from NeRF

Both approaches build a scene viewable from any angle out of several photos. Where they split is what they hold the result as. Splatting holds a visible pile of pieces; NeRF holds a computational rule that returns a light value given a spot and a direction in space.

So how each one draws to the screen differs too. Splatting flattens and paints its pieces on; NeRF stretches a line of sight through every point on the screen and asks the values along it over and over. Whichever asks more questions runs slower.

3More precisely

A single piece is a three-dimensional Gaussian — a bell-shaped distribution, dense at the center and fading toward the edges. Because the edges taper off gradually instead of cutting off sharply, just a few overlapping pieces are enough to make a surface look smooth. When drawing to the screen, this 3D distribution gets flattened toward the camera into a 2D one, then color and transparency get mixed and stacked from farthest to nearest.

A piece's color isn't fixed to one value. Because it also carries direction-dependent values that let it shift color slightly with viewing angle, even reflections on metal or glass come through to some degree.

The analogy breaks down in places. A person glues mosaic tiles by following a design, but here the pieces move themselves, shrinking the gap between the photos and the rendered picture. Tiles are opaque and don't overlap, but pieces are translucent and stack up in many layers. And a tile mural is one fixed picture on a wall, while a pile of pieces is a scene redrawn fresh depending on which angle you view it from.

4Try it yourself

5Common misconceptions

It's easy to think this is a laser 3D scan that plots points one by one, but actually it's built by comparing ordinary photos and moving, growing, and deleting pieces.
It's easy to think it looks perfect from every angle, but actually angles that were never photographed show the pieces splitting apart into blotches or floating specks.
It's easy to think this is just another name for NeRF, but actually both what it holds the scene as and how it draws to the screen are different.

7One-line summary

In shortGaussian splatting seats millions of color blobs in space from several photos, and draws them so they overlap correctly from whatever angle you look — that's its way of holding a 3D scene.

Did this analogy help?

Spotted an error or have a better analogy? Suggest an edit · Last updated2026-09-02