A previous article covered how to use Geometry. This article is about BufferGeometry. BufferGeometry is generally faster to start and uses less memory but can be harder to setup.

In the article on Geometry we went over that to use a Geometry you supply an array of Vector3 vertices (positions). You then make Face3 objects specifying by index the 3 vertices that make each triangle of the shape you're making. To each Face3 you can specify either a face normal or normals for each individual vertex of the face. You can also specify a face color or individual vertex colors. Finally you can make a parallel array of arrays of texture coordinates (UVs), one array for each face containing an array of UVs, one for each vertex of the face.

BufferGeometry on the other hand uses named BufferAttributes. Each BufferAttribute represents an array of one type of data: positions, normals, colors, and uv. Together, the added BufferAttributes represent parallel arrays of all the data for each vertex.

Above you can see we have 4 attributes: position, normal, color, uv. They represent parallel arrays which means that the Nth set of data in each attribute belongs to the same vertex. The vertex at index = 4 is highlighted to show that the parallel data across all attributes defines one vertex.

This brings up a point, here's a diagram of a cube with one corner highlighted.

Thinking about it that single corner needs a different normal for each face of the cube. It needs different UVs for each face as well. This points out the biggest difference between Geometry and BufferGeometry. Nothing is shared with BufferGeometry. A single vertex is the combination of all of its parts. If a vertex needs any part to be different then it must be a different vertex.

The truth is when you use Geometry three.js transforms it into this format. That is where the extra memory and time comes from when using Geometry. Extra memory for all the Vector3s, Vector2s, Face3s and array objects and then extra time to translate all of that data into parallel arrays in the form of BufferAttributes like above. Sometimes that makes using Geometry easier. With BufferGeometry it is up to us to supply the data already turned into this format.

As a simple example let's make a cube using BufferGeometry. A cube is interesting because it appears to share vertices at the corners but really does not. For our example we'll list out all the vertices with all their data and then convert that data into parallel arrays and finally use those to make BufferAttributes and add them to a BufferGeometry.

Starting with the texture coordinate example from the previous article we've deleted all the code related to setting up a Geometry. Then we list all the data needed for the cube. Remember again that if a vertex has any unique parts it has to be a separate vertex. As such to make a cube requires 36 vertices. 2 triangles per face, 3 vertices per triangle, 6 faces = 36 vertices.

const vertices = [
  // front
  { pos: [-1, -1,  1], norm: [ 0,  0,  1], uv: [0, 0], },
  { pos: [ 1, -1,  1], norm: [ 0,  0,  1], uv: [1, 0], },
  { pos: [-1,  1,  1], norm: [ 0,  0,  1], uv: [0, 1], },

  { pos: [-1,  1,  1], norm: [ 0,  0,  1], uv: [0, 1], },
  { pos: [ 1, -1,  1], norm: [ 0,  0,  1], uv: [1, 0], },
  { pos: [ 1,  1,  1], norm: [ 0,  0,  1], uv: [1, 1], },
  // right
  { pos: [ 1, -1,  1], norm: [ 1,  0,  0], uv: [0, 0], },
  { pos: [ 1, -1, -1], norm: [ 1,  0,  0], uv: [1, 0], },
  { pos: [ 1,  1,  1], norm: [ 1,  0,  0], uv: [0, 1], },

  { pos: [ 1,  1,  1], norm: [ 1,  0,  0], uv: [0, 1], },
  { pos: [ 1, -1, -1], norm: [ 1,  0,  0], uv: [1, 0], },
  { pos: [ 1,  1, -1], norm: [ 1,  0,  0], uv: [1, 1], },
  // back
  { pos: [ 1, -1, -1], norm: [ 0,  0, -1], uv: [0, 0], },
  { pos: [-1, -1, -1], norm: [ 0,  0, -1], uv: [1, 0], },
  { pos: [ 1,  1, -1], norm: [ 0,  0, -1], uv: [0, 1], },

  { pos: [ 1,  1, -1], norm: [ 0,  0, -1], uv: [0, 1], },
  { pos: [-1, -1, -1], norm: [ 0,  0, -1], uv: [1, 0], },
  { pos: [-1,  1, -1], norm: [ 0,  0, -1], uv: [1, 1], },
  // left
  { pos: [-1, -1, -1], norm: [-1,  0,  0], uv: [0, 0], },
  { pos: [-1, -1,  1], norm: [-1,  0,  0], uv: [1, 0], },
  { pos: [-1,  1, -1], norm: [-1,  0,  0], uv: [0, 1], },

  { pos: [-1,  1, -1], norm: [-1,  0,  0], uv: [0, 1], },
  { pos: [-1, -1,  1], norm: [-1,  0,  0], uv: [1, 0], },
  { pos: [-1,  1,  1], norm: [-1,  0,  0], uv: [1, 1], },
  // top
  { pos: [ 1,  1, -1], norm: [ 0,  1,  0], uv: [0, 0], },
  { pos: [-1,  1, -1], norm: [ 0,  1,  0], uv: [1, 0], },
  { pos: [ 1,  1,  1], norm: [ 0,  1,  0], uv: [0, 1], },

  { pos: [ 1,  1,  1], norm: [ 0,  1,  0], uv: [0, 1], },
  { pos: [-1,  1, -1], norm: [ 0,  1,  0], uv: [1, 0], },
  { pos: [-1,  1,  1], norm: [ 0,  1,  0], uv: [1, 1], },
  // bottom
  { pos: [ 1, -1,  1], norm: [ 0, -1,  0], uv: [0, 0], },
  { pos: [-1, -1,  1], norm: [ 0, -1,  0], uv: [1, 0], },
  { pos: [ 1, -1, -1], norm: [ 0, -1,  0], uv: [0, 1], },

  { pos: [ 1, -1, -1], norm: [ 0, -1,  0], uv: [0, 1], },
  { pos: [-1, -1,  1], norm: [ 0, -1,  0], uv: [1, 0], },
  { pos: [-1, -1, -1], norm: [ 0, -1,  0], uv: [1, 1], },
];

We can then translate all of that into 3 parallel arrays

const positions = [];
const normals = [];
const uvs = [];
for (const vertex of vertices) {
  positions.push(...vertex.pos);
  normals.push(...vertex.norm);
  uvs.push(...vertex.uv);
}

Finally we can create a BufferGeometry and then a BufferAttribute for each array and add it to the BufferGeometry.

  const geometry = new THREE.BufferGeometry();
  const positionNumComponents = 3;
  const normalNumComponents = 3;
  const uvNumComponents = 2;
  geometry.setAttribute(
      'position',
      new THREE.BufferAttribute(new Float32Array(positions), positionNumComponents));
  geometry.setAttribute(
      'normal',
      new THREE.BufferAttribute(new Float32Array(normals), normalNumComponents));
  geometry.setAttribute(
      'uv',
      new THREE.BufferAttribute(new Float32Array(uvs), uvNumComponents));

Note that the names are significant. You must name your attributes the names that match what three.js expects (unless you are creating a custom shader). In this case position, normal, and uv. If you want vertex colors then name your attribute color.

Above we created 3 JavaScript native arrays, positions, normals and uvs. We then convert those into TypedArrays of type Float32Array. A BufferAttribute requires a TypedArray not a native array. A BufferAttribute also requires you to tell it how many components there are per vertex. For the positions and normals we have 3 components per vertex, x, y, and z. For the UVs we have 2, u and v.

That's a lot of data. A small thing we can do is use indices to reference the vertices. Looking back at our cube data, each face is made from 2 triangles with 3 vertices each, 6 vertices total, but 2 of those vertices are exactly the same; The same position, the same normal, and the same uv. So, we can remove the matching vertices and then reference them by index. First we remove the matching vertices.

const vertices = [
  // front
  { pos: [-1, -1,  1], norm: [ 0,  0,  1], uv: [0, 0], }, // 0
  { pos: [ 1, -1,  1], norm: [ 0,  0,  1], uv: [1, 0], }, // 1
  { pos: [-1,  1,  1], norm: [ 0,  0,  1], uv: [0, 1], }, // 2
-
-  { pos: [-1,  1,  1], norm: [ 0,  0,  1], uv: [0, 1], },
-  { pos: [ 1, -1,  1], norm: [ 0,  0,  1], uv: [1, 0], },
  { pos: [ 1,  1,  1], norm: [ 0,  0,  1], uv: [1, 1], }, // 3
  // right
  { pos: [ 1, -1,  1], norm: [ 1,  0,  0], uv: [0, 0], }, // 4
  { pos: [ 1, -1, -1], norm: [ 1,  0,  0], uv: [1, 0], }, // 5
-
-  { pos: [ 1,  1,  1], norm: [ 1,  0,  0], uv: [0, 1], },
-  { pos: [ 1, -1, -1], norm: [ 1,  0,  0], uv: [1, 0], },
  { pos: [ 1,  1,  1], norm: [ 1,  0,  0], uv: [0, 1], }, // 6
  { pos: [ 1,  1, -1], norm: [ 1,  0,  0], uv: [1, 1], }, // 7
  // back
  { pos: [ 1, -1, -1], norm: [ 0,  0, -1], uv: [0, 0], }, // 8
  { pos: [-1, -1, -1], norm: [ 0,  0, -1], uv: [1, 0], }, // 9
-
-  { pos: [ 1,  1, -1], norm: [ 0,  0, -1], uv: [0, 1], },
-  { pos: [-1, -1, -1], norm: [ 0,  0, -1], uv: [1, 0], },
  { pos: [ 1,  1, -1], norm: [ 0,  0, -1], uv: [0, 1], }, // 10
  { pos: [-1,  1, -1], norm: [ 0,  0, -1], uv: [1, 1], }, // 11
  // left
  { pos: [-1, -1, -1], norm: [-1,  0,  0], uv: [0, 0], }, // 12
  { pos: [-1, -1,  1], norm: [-1,  0,  0], uv: [1, 0], }, // 13
-
-  { pos: [-1,  1, -1], norm: [-1,  0,  0], uv: [0, 1], },
-  { pos: [-1, -1,  1], norm: [-1,  0,  0], uv: [1, 0], },
  { pos: [-1,  1, -1], norm: [-1,  0,  0], uv: [0, 1], }, // 14
  { pos: [-1,  1,  1], norm: [-1,  0,  0], uv: [1, 1], }, // 15
  // top
  { pos: [ 1,  1, -1], norm: [ 0,  1,  0], uv: [0, 0], }, // 16
  { pos: [-1,  1, -1], norm: [ 0,  1,  0], uv: [1, 0], }, // 17
-
-  { pos: [ 1,  1,  1], norm: [ 0,  1,  0], uv: [0, 1], },
-  { pos: [-1,  1, -1], norm: [ 0,  1,  0], uv: [1, 0], },
  { pos: [ 1,  1,  1], norm: [ 0,  1,  0], uv: [0, 1], }, // 18
  { pos: [-1,  1,  1], norm: [ 0,  1,  0], uv: [1, 1], }, // 19
  // bottom
  { pos: [ 1, -1,  1], norm: [ 0, -1,  0], uv: [0, 0], }, // 20
  { pos: [-1, -1,  1], norm: [ 0, -1,  0], uv: [1, 0], }, // 21
-
-  { pos: [ 1, -1, -1], norm: [ 0, -1,  0], uv: [0, 1], },
-  { pos: [-1, -1,  1], norm: [ 0, -1,  0], uv: [1, 0], },
  { pos: [ 1, -1, -1], norm: [ 0, -1,  0], uv: [0, 1], }, // 22
  { pos: [-1, -1, -1], norm: [ 0, -1,  0], uv: [1, 1], }, // 23
];

So now we have 24 unique vertices. Then we specify 36 indices for the 36 vertices we need drawn to make 12 triangles by calling BufferGeometry.setIndex with an array of indices.

geometry.setAttribute(
    'position',
    new THREE.BufferAttribute(positions, positionNumComponents));
geometry.setAttribute(
    'normal',
    new THREE.BufferAttribute(normals, normalNumComponents));
geometry.setAttribute(
    'uv',
    new THREE.BufferAttribute(uvs, uvNumComponents));

+geometry.setIndex([
+   0,  1,  2,   2,  1,  3,  // front
+   4,  5,  6,   6,  5,  7,  // right
+   8,  9, 10,  10,  9, 11,  // back
+  12, 13, 14,  14, 13, 15,  // left
+  16, 17, 18,  18, 17, 19,  // top
+  20, 21, 22,  22, 21, 23,  // bottom
+]);

Just like Geometry, BufferGeometry has a computeVertexNormals method for computing normals if you are not supplying them. Unlike the Geometry version of the same function, since positions can not be shared if any other part of a vertex is different the results of calling computeVertexNormals will be different.

Here are 2 cylinders where the normals were created using computeVertexNormals. If you look closely there is a seam on the left cylinder. This is because there is no way to share the vertices at the start and end of the cylinder since they require different UVs. Just a small thing to be aware of. The solution is to supply your own normals.

We can also use TypedArrays from the start instead of native JavaScript arrays. The disadvantage to TypedArrays is you must specify their size up front. Of course that's not that large of a burden but with native arrays we can just push values onto them and look at what size they end up by checking their length at the end. With TypedArrays there is no push function so we need to do our own bookkeeping when adding values to them.

In this example knowing the length up front is pretty easy since we're using a big block of static data to start.

-const positions = [];
-const normals = [];
-const uvs = [];
+const numVertices = vertices.length;
+const positionNumComponents = 3;
+const normalNumComponents = 3;
+const uvNumComponents = 2;
+const positions = new Float32Array(numVertices * positionNumComponents);
+const normals = new Float32Array(numVertices * normalNumComponents);
+const uvs = new Float32Array(numVertices * uvNumComponents);
+let posNdx = 0;
+let nrmNdx = 0;
+let uvNdx = 0;
for (const vertex of vertices) {
-  positions.push(...vertex.pos);
-  normals.push(...vertex.norm);
-  uvs.push(...vertex.uv);
+  positions.set(vertex.pos, posNdx);
+  normals.set(vertex.norm, nrmNdx);
+  uvs.set(vertex.uv, uvNdx);
+  posNdx += positionNumComponents;
+  nrmNdx += normalNumComponents;
+  uvNdx += uvNumComponents;
}

geometry.setAttribute(
    'position',
-    new THREE.BufferAttribute(new Float32Array(positions), positionNumComponents));
+    new THREE.BufferAttribute(positions, positionNumComponents));
geometry.setAttribute(
    'normal',
-    new THREE.BufferAttribute(new Float32Array(normals), normalNumComponents));
+    new THREE.BufferAttribute(normals, normalNumComponents));
geometry.setAttribute(
    'uv',
-    new THREE.BufferAttribute(new Float32Array(uvs), uvNumComponents));
+    new THREE.BufferAttribute(uvs, uvNumComponents));

geometry.setIndex([
   0,  1,  2,   2,  1,  3,  // front
   4,  5,  6,   6,  5,  7,  // right
   8,  9, 10,  10,  9, 11,  // back
  12, 13, 14,  14, 13, 15,  // left
  16, 17, 18,  18, 17, 19,  // top
  20, 21, 22,  22, 21, 23,  // bottom
]);

A good reason to use typedarrays is if you want to dynamically update any part of the vertices.

I couldn't think of a really good example of dynamically updating the vertices so I decided to make a sphere and move each quad in and out from the center. Hopefully it's a useful example.

Here's the code to generate positions and indices for a sphere. The code is sharing vertices within a quad but it's not sharing vertices between quads because we want to be able to move each quad separately.

Because I'm lazy I used a small hierarchy of 3 Object3D objects to compute sphere points. How this works is explained in the article on optimizing lots of objects.

function makeSpherePositions(segmentsAround, segmentsDown) {
  const numVertices = segmentsAround * segmentsDown * 6;
  const numComponents = 3;
  const positions = new Float32Array(numVertices * numComponents);
  const indices = [];

  const longHelper = new THREE.Object3D();
  const latHelper = new THREE.Object3D();
  const pointHelper = new THREE.Object3D();
  longHelper.add(latHelper);
  latHelper.add(pointHelper);
  pointHelper.position.z = 1;
  const temp = new THREE.Vector3();

  function getPoint(lat, long) {
    latHelper.rotation.x = lat;
    longHelper.rotation.y = long;
    longHelper.updateMatrixWorld(true);
    return pointHelper.getWorldPosition(temp).toArray();
  }

  let posNdx = 0;
  let ndx = 0;
  for (let down = 0; down < segmentsDown; ++down) {
    const v0 = down / segmentsDown;
    const v1 = (down + 1) / segmentsDown;
    const lat0 = (v0 - 0.5) * Math.PI;
    const lat1 = (v1 - 0.5) * Math.PI;

    for (let across = 0; across < segmentsAround; ++across) {
      const u0 = across / segmentsAround;
      const u1 = (across + 1) / segmentsAround;
      const long0 = u0 * Math.PI * 2;
      const long1 = u1 * Math.PI * 2;

      positions.set(getPoint(lat0, long0), posNdx);  posNdx += numComponents;
      positions.set(getPoint(lat1, long0), posNdx);  posNdx += numComponents;
      positions.set(getPoint(lat0, long1), posNdx);  posNdx += numComponents;
      positions.set(getPoint(lat1, long1), posNdx);  posNdx += numComponents;

      indices.push(
        ndx, ndx + 1, ndx + 2,
        ndx + 2, ndx + 1, ndx + 3,
      );
      ndx += 4;
    }
  }
  return {positions, indices};
}

We can then call it like this

const segmentsAround = 24;
const segmentsDown = 16;
const {positions, indices} = makeSpherePositions(segmentsAround, segmentsDown);

Because positions returned are unit sphere positions so they are exactly the same values we need for normals so we can just duplicated them for the normals.

const normals = positions.slice();

And then we setup the attributes like before

const geometry = new THREE.BufferGeometry();
const positionNumComponents = 3;
const normalNumComponents = 3;

+const positionAttribute = new THREE.BufferAttribute(positions, positionNumComponents);
+positionAttribute.setUsage(THREE.DynamicDrawUsage);
geometry.setAttribute(
    'position',
+    positionAttribute);
geometry.setAttribute(
    'normal',
    new THREE.BufferAttribute(normals, normalNumComponents));
geometry.setIndex(indices);

I've highlighted a few differences. We save a reference to the position attribute. We also mark it as dynamic. This is a hint to THREE.js that we're going to be changing the contents of the attribute often.

In our render loop we update the positions based off their normals every frame.

const temp = new THREE.Vector3();

...

for (let i = 0; i < positions.length; i += 3) {
  const quad = (i / 12 | 0);
  const ringId = quad / segmentsAround | 0;
  const ringQuadId = quad % segmentsAround;
  const ringU = ringQuadId / segmentsAround;
  const angle = ringU * Math.PI * 2;
  temp.fromArray(normals, i);
  temp.multiplyScalar(THREE.MathUtils.lerp(1, 1.4, Math.sin(time + ringId + angle) * .5 + .5));
  temp.toArray(positions, i);
}
positionAttribute.needsUpdate = true;

And we set positionAttribute.needsUpdate to tell THREE.js to use our changes.

I hope these were useful examples of how to use BufferGeometry directly to make your own geometry and how to dynamically update the contents of a BufferAttribute. Which you use, Geometry or BufferGeometry, really depends on your needs.