Placement, anchors and the camera

One convention covers 2D and 3D, single objects and instances: position is the world position of the anchor, and width, height and depth are the target size — not scale factors. The anchor is the geometric centre of the object's own bounding box.

The members you use

The payoff is that you can size an imported asset without knowing how the artist exported it. Width = 1f means one metre wide, whatever the GLB claims, and two models from two artists placed side by side are actually the same height.

The world matrix, written out

Every 3D control builds its matrix the same way, and the order is the whole point:

Matrix4x4.CreateTranslation(-AnchorLocal)
    * Matrix4x4.CreateScale(ComputedScale)
    * GetRotationMatrix()
    * Matrix4x4.CreateTranslation(PosX, PosY, PosZ)

Under the row-vector convention that is ((p - A) * S) * R + Pos. The anchor A lands exactly on Pos, the rotation pivot is the anchor, and therefore rotating an object changes neither its position nor its size. That property is what makes the object editor in the reference application possible without any compensation maths.

Placing an origin instead of a centre

Sometimes you want the model's own origin pinned to a world point — a character exported standing on the ground plane, for instance. The identity p*S*R + t == ((p-A)*S)*R + Pos gives you the conversion directly:

// Pin the model's local origin to a world-space target.
var target = new Vector3(4f, 0f, 12f);
var offset = robot.AnchorWorldOffset;   // recompute each frame

robot.PosX = target.X + offset.X;
robot.PosY = target.Y + offset.Y;
robot.PosZ = target.Z + offset.Z;

Recompute the offset each frame rather than caching it: it already has the current scale and rotation folded in, so the identity keeps holding as the object changes. Set size and rotation first, then position.

Three behaviours worth knowing before they surprise you

• An unset dimension settles to the asset's own size. When bounds are established after loading, any Width, Height or Depth that is null or 0 is filled in from LocalSize, so a model with no size set appears at its exported scale rather than collapsing to nothing.
• Flat axes are not scaled. If a local dimension is effectively zero — below 1e-6, or more than four orders of magnitude smaller than the largest axis — that axis stays at scale 1. Without the guard, thickness noise in an exported plane produces scales around 1e5 and tears the mesh into offset fragments. This is a real asset, not a hypothetical one.
• Animated models have two bounding boxes. GetWorldBounds() returns the conservative box used for culling and shadows, inflated for skinned motion. GetWorldBoundsRaw() matches the rendered body. Picking and selection boxes use the raw one; using the conservative box for a selection box leaves visible empty space above and below the model.

The camera

BaseApp.Camera is a single Camera3D shared by every backend and every pass, and it is the only source of truth for view, projection and frustum. CameraPos and CameraTarget forward to Camera.Position and Camera.Target.

CameraTarget = new Vector3(0f, 3.5f, -2.5f);
CameraPos = CameraTarget + new Vector3(0f, 0.5f, -3.5f);

Camera.FovY = MathF.PI / 4f;   // radians
Camera.Near = 0.1f;
Camera.Far = 1300f;

With CameraPos.Z negative the camera looks toward the origin and world +X is screen right — the arrangement every sample uses, and the one the 2D overlay coordinate system agrees with.