Models, animation and instancing

glTF 2.0 is the only model and animation format. Not the preferred one — the only one. That decision removes an abstraction layer, an import pipeline and a class of bugs where two formats disagree about bind poses.

A single model

Model places a GLB in the world using the engine's uniform placement convention: position is the world position of the anchor, and width, height and depth are the target bounds.

var robot = new Model
{
    Name = "Assets/3DGodotRobot.glb",
    PosX = -2f, PosY = 1f, PosZ = 6f,
    Width = 1f, Height = 1f, Depth = 0.5f,
    Rotation = MathF.PI / 2f
};
AddControl(robot);

Placement, anchors and the camera covers the anchor maths, including AnchorWorldOffset for the case where you want to pin the model's own origin rather than the centre of its bounds.

Animation

Skinning, morph targets and named clips all come from the file. Clips are addressed by name, and the names are the ones the artist used.

// After Load has completed:
foreach (var name in robot.GetAnimationNames())
    app.AddLog(LogType.Info, name);

robot.PlayAnimation("Walk");

// Or cycle, which is what a debug key usually wants:
robot.SwitchToNextAnimation();

// Advance the clock yourself; Time is the animation clock.
robot.Update(time);

Many models, one control

An InstancedModel is declared once and bound to a list of instance transforms. Skinning and morph targets are per instance, so twenty characters can each play a different clip from one draw. This is the pattern the reference application uses for robots, birds, rocks and the beach.

robotField = new InstancedModel { ModelName = "Assets/3DGodotRobot.glb" };
AddControl(robotField);

for (var i = 0; i < 10; i++)
{
    var person = new Person
    {
        PosX = -2f, PosY = 1f, PosZ = 6f + i * 3f,
        Width = 1f, Height = 1f, Depth = 0.5f,
        Animation = animations[i]
    };

    robotField.Instances.Add(person);
}

Person derives from MeshInstanceTransform, so the object in your list is the instance. Editing its fields takes effect on the next frame with no copy step, and inserting or removing in the middle of the list preserves order.

Procedural geometry

Not everything comes from a file. Mesh3D takes a list of Surface objects — vertices, indices and a material — and behaves like any other control from there. The sea, the beach tiles and the cube fields in the reference application are built this way.

Culling, and why animated bounds are inflated

Frustum culling and light-space shadow culling both run on the CPU against bounding volumes. The subtlety is that a skinned mesh can move outside its own bind-pose bounds, so the bounds used for culling are grown by AnimatedBoundsScale, 1.5 by default.

An animation that swings a limb further than 50 per cent past the bind pose will pop at the frustum edge. Raise AnimatedBoundsScale or clear CullingEnabled on that control; both are cheap compared with the alternative of computing exact skinned bounds every frame.

Lights that arrive with the model

A glTF file can carry KHR_lights_punctual lights, and Model surfaces them as ImportedPunctualLights, appending them into the frame's lighting through AppendWorldLights.

Two things to remember: the frame budget is eight lights total, so an asset carrying twelve of them will crowd out your own; and glTF intensities are photometric, which is why Model.LightIntensityScale and RenderQuality.KhrLightIntensityScale exist. See Lighting and shadows.

Known gaps

• Absent Animation blending and state machines. One clip at a time, switched as a cut.
• Absent Level of detail. No automatic LOD selection or mesh simplification.
• Absent Indirect draw and GPU-driven culling. Instances are culled and submitted from the CPU. Roadmap Track A.
• Absent Occlusion culling. Frustum and shadow-frustum only; a wall does not cull what is behind it.