Math

Prefer Three.js math classes (Vector3, Euler, Quaternion, Matrix4) over custom implementations: they are well-tested, handle edge cases, and integrate seamlessly with Three.js objects. The same goes for measurement primitives like Box3 and coordinate helpers like Spherical. The throughline is the reference frame — world space versus local space — which is where even the built-ins mislead if you ignore it.

This document covers four areas. Spherical Coordinates and Trigonometry replaces manual sin/cos. Rotations covers choosing between Euler and Quaternion. Billboards covers camera-facing quads, which must be oriented in world space. Measuring an object's size (Box3) covers why measuring a skinned/animated character with setFromObject is corrupted, and what to do instead.

Spherical Coordinates and Trigonometry

For spherical coordinates or anything involving sin/cos, use Euler, Vector3, and Quaternion rather than hand-rolled trigonometry. For an orbit position, build a Spherical and read it into a vector. For a direction from accumulated angles, build an Euler and apply it to a forward vector.

// Instead of manual sin/cos for orbit position:
const spherical = new Spherical(radius, polarAngle, azimuthalAngle)
position.setFromSpherical(spherical)

// For direction from angles:
const euler = new Euler(pitch, yaw, 0, 'YXZ')
const direction = new Vector3(0, 0, -1).applyEuler(euler)

Rotations

Rotations need careful handling, and the representation is chosen by use case. Use Euler with 'YXZ' order for orbital and FPS cameras, where pitch and yaw accumulate separately. Use Quaternion for general rotations, for interpolation (slerp), and to avoid gimbal lock. When converting between representations, account for the Euler order. Staying in a single representation reduces bugs. Three.js objects expose both .rotation (an Euler) and .quaternion (a Quaternion), and keep them synced automatically.

Billboards (labels, health bars, icons that face the camera)

A billboard quad must be oriented in world space. If it is a child of a rotating object (an enemy, a turret), copying the camera's orientation into its local rotation is wrong — the parent's world rotation still applies on top, so the quad tilts and spins with the parent. This local-rotation mistake is the usual cause of a health bar that is rotated incorrectly.

Fix A: orient in world space, compensating for the parent's world rotation. Fix B (simpler): do not parent the label to the moving object at all — keep it in world space and position it directly.

// A — orient in world space, compensating for the parent's world rotation:
mesh.quaternion.copy(camera.quaternion) // desired WORLD rotation
const parentWorld = mesh.parent.getWorldQuaternion(new Quaternion())
mesh.quaternion.premultiply(parentWorld.invert()) // local = parentWorld⁻¹ · desiredWorld

// B (simpler) — don't parent it to the moving object; keep it in world space:
scene.add(label)
label.position.copy(targetWorldPosition).add(offset)
label.quaternion.copy(camera.quaternion)

THREE.Sprite always faces the camera automatically and avoids the problem entirely — prefer it for simple health bars and icons.

Measuring an object's size (Box3) — beware skinned/rigged meshes

new Box3().setFromObject(object) is correct for static meshes. It is wrong and unstable for skinned, rigged, or animated objects such as characters, for two reasons. First, it unions every descendant — including the skeleton bones (which sit far outside the visible body) and any reference or clone sub-objects a character system keeps. Second, it expands to bind-pose geometry, so the box changes every frame as the animation poses the bones.

The wrong result is routinely 2–3× (or more) the visible body height, and it silently breaks anything sized from that number. The classic failure: scaling one character to match another's measured height produces a giant, because the measured character was live and animating (inflated) while the other was measured fresh (clean).

Use one of three correct options.

• Measure the SkinnedMesh's geometry.boundingBox — the stable bind-pose mesh extent (mesh.geometry.computeBoundingBox(), then read mesh.geometry.boundingBox). Caveat: that box is in local geometry space; for a world-unit height, transform it by the mesh's world matrix — mesh.geometry.boundingBox.clone().applyMatrix4(mesh.matrixWorld) — which folds in scale from parent nodes, where a glTF's import scale usually lives. Reading it raw, or multiplying only by mesh.scale, measures a scaled character wrong.
• Measure the raw model once on load, before it is rigged, animated, or wrapped.
• Avoid measuring entirely: size against a known constant — a target character height you decide.

Don't setFromObject a live character every frame and trust the height.