Melonjs 3d Agent Skill

3D and 2.5D in melonJS

> melonJS is a 2D engine that grew a real 3D tier: perspective cameras, meshes,

> instancing, glTF scenes and 3D lights. Most of what a model "knows" about 3D

> graphics comes from the OpenGL convention, and melonJS's is the inverse.

> Read the conventions section before writing any 3D code.

Coordinate conventions — the inverse of OpenGL

This is the single most important thing on this page.

| | melonJS | OpenGL convention |

|---|---|---|

| vertical | Y-down — higher y is *lower* on screen | Y-up |

| depth | +Z forward — higher z is *farther* away | −Z forward |

Camera3d exposes three rotation angles: camera.pitch

(X, look up/down), camera.yaw (Y, look left/right) and camera.roll (Z, bank

the horizon). The view is R(yaw) ∘ R(pitch) ∘ R(roll) inverted, and the frustum

planes come off that same matrix — so culling follows a banked view.

The basis those angles produce is readable and writable: `getBasis(right,

up, forward) reads it, setBasis(right, up, forward)` poses the camera from

one.

camera.rotate() is NOT the way to bank a 3D camera. It writes the inherited

currentTransform, which a 3D view never reads, so the call is silently inert.

Set roll. (On a Camera2d it is the other way round: roll IS that transform's

rotation, which is why worldToLocal / localToWorld compensate for it.)

The payoff is that 2D code translates directly — anywhere you used pos.x /

pos.y, add pos.z and the maths still holds. The cost is that every

OpenGL-shaped instinct is backwards.

Consequence that bites immediately: "behind everything" is the *largest*

depth when the camera looks along +Z. Put a backdrop on the wrong side and it

either paints over the whole scene (it is nearer than everything else) or

disappears without trace (it falls in front of the near plane and the GPU clips

it). Which sign means "far" depends on which side the camera sits, so check the

camera before picking a backdrop depth.

glTF assets are authored Y-up right-handed. level.load converts them by

rotation (rightHanded: true is its default), so you do not apply a flip

yourself. The Mesh constructor defaults rightHanded to false, which

bridges by mirroring instead — geometry you hand to a Mesh directly comes in

flipped left/right unless you pass rightHanded: true. See

melonjs-3d-assets.

Opting in

3D starts with the camera class, and nothing else has to change. Set it once on

the Application:

import { Application, Camera3d } from "melonjs";

const app = new Application(1024, 768, {
    parent: "screen",
    cameraClass: Camera3d,      // ← the opt-in
});
await app.init();

That flips the world's sort to "depth", switches the broadphase to a 3D index,

and renders sprites and meshes with perspective projection. Camera3d extends

Camera2d, so follow, fade, shake and post effects all still work.

Keep the camera, do not fish it back out

app.viewport is the active stage's default camera, so it is typed

Camera2d: a stage overrides the application's cameraClass, and the

built-in loading screen does exactly that. Reading your Camera3d back out of

it therefore needs a cast, and a cast that is wrong while the loader is up:

const camera = app.viewport as Camera3d;   // works, until it does not

Own it instead. Pass the camera in and keep the typed reference, and nothing

downstream ever narrows:

class GameStage extends Stage {
    readonly camera: Camera3d;

    constructor(app) {
        const camera = new Camera3d(0, 0, app.renderer.width, app.renderer.height);
        super({ cameras: [camera] });
        this.camera = camera;
    }
}

A freshly built camera is named "default", and a stage registers its

cameras by name, so this IS the stage's default camera: app.viewport

returns the same instance.

The two patterns differ in LIFETIME, and that decides which you want

This is the part that bites. cameraClass builds a **fresh camera for each

stage**, which the engine does deliberately so "Camera3d state shouldn't

bleed across stages". An owned camera passed through super({ cameras }) is

built once and lives as long as the stage object, which for a stage you

register with state.set() means as long as the application.

So anything you attach to the camera now outlives the visit:

// in onResetEvent — runs EVERY time the stage is entered
app.viewport.addPostEffect(bloom);

With cameraClass that lands on a fresh camera each time and is harmless.

With an owned camera it stacks: a second bloom on the second visit, a third

on the third. The symptom is text and edges appearing to double or smear,

because the scene is being post-processed once per accumulated effect.

Note that removePostEffect DESTROYS the effect rather than detaching it, so

"just remove it on the way out" is not a free undo.

Choose on lifetime, not on typing:

| your effects / camera state are | use |

|---|---|

| set up per stage visit, in onResetEvent | cameraClass — fresh camera per stage |

| set up once, in the stage constructor | an owned camera in super({ cameras }) |

If you want the typed reference but your setup is per-visit, keep

cameraClass and narrow once on reset:

override onResetEvent(app) {
    const viewport = app.viewport;
    if (!(viewport instanceof Camera3d)) {
        throw new Error("this stage needs a Camera3d");
    }
    this.camera = viewport;   // typed, checked, and still fresh per visit
}

That is what the Earth Defender example does, and why: it attaches post

effects on every reset and never removes them.

Narrowing beats asserting wherever you read the shared slot: the language

verifies instanceof, and fails loudly if the application was never given a

cameraClass, where a cast fails silently at the first 3D call.

Per-stage opt-in also works, which is useful when the loading screen is 2D:

class GameStage extends Stage {
    constructor() {
        super({ cameras: [new Camera3d(0, 0, 1024, 768, { fov: Math.PI / 3 })] });
    }
}

Camera traps

• There is no sky. The frame clears to renderer.backgroundColor (opaque

black unless the renderer was created transparent), and world.backgroundColor

clears over it when you set one — both are flat colours. Anything more is a

backdrop renderable, screen-fixed or placed at a far depth.

• Set the clip planes for your scene scale. Defaults are near 0.1 / far 1000.

Objects beyond the far plane clip or project with bad w-divides; a near that

is too small wastes depth precision and distant props z-fight. Use

camera.setClipPlanes(near, far).

• camera.pos.set(x, y) is 2-argument and zeroes z. Use

camera.setPosition(x, y, z), which leaves the depth untouched when z is

omitted; camera.depth is the z accessor if you only want that.

• lookAt(target) solves pitch and yaw only, so on its own it cannot

express "up is this direction" and it leaves roll alone. Pass an up to

bank the horizon too: camera.lookAt(target, up), where up is what should

appear UP ON SCREEN. This engine renders Y-down, so a level horizon is

(0, -1, 0). There is deliberately no default, because giving one would

silently level every camera that had banked its own horizon.

• When the game already has all three axes, hand them over:

camera.setBasis(right, up, forward), the writable counterpart of

getBasis. The case that needs it is a view over a curved surface, where up

is the surface normal and differs at every point, so no pitch and yaw pair

expresses it and integrating towards one drifts and gimbals at the poles.

setBasis absorbs a triple that has drifted off orthonormal rather than

leaving the camera holding something that is not a rotation, and it does not

care what the axes are scaled to: a direction is a direction, so anything

from a denormal to 1e300 resolves the same way. Note its up is the camera's

local +Y column, which under Y-down points DOWN the screen: the opposite

sense to lookAt's, and one negation apart.

• worldToLocal does not project. It is a 2D camera's offset subtraction.

To pin a label or marker to a point in the scene use

camera.worldToScreen(x, y, z), which applies the projection and returns

null behind the camera. See melonjs-camera-and-drawing.

Where an up vector earns its keep

Three cases. In the first two, lookAt(target) on its own is already right and

an up would be noise. The third is the one it exists for.

// 1. ordinary third-person: world up never changes, so say nothing
camera.setPosition(x, y, z).lookAt(player);

// 2. a deliberate dutch angle: this is a ROLL, not an up vector
camera.lookAt(player);
camera.roll = Math.PI / 32;

// 3. a view over a curved surface: up is the surface normal, and it is
//    different at every point on the globe. No pitch and yaw pair expresses
//    it, and integrating towards it drifts and gimbals over the poles.
const up = player.pos.clone().sub(planetCentre).normalize();  // surface normal
camera.setPosition(eye.x, eye.y, eye.z).lookAt(player, up.negate());

Note the negate() in case 3. The surface normal points away from the planet,

which is the direction the player's head points, and up wants what should

appear UP ON SCREEN. This engine renders Y-down, so (0, -1, 0) is a level

horizon and a world-space "up" direction has to be handed over in that sense.

Get it backwards and the view is upside down, which is the single most likely

mistake with this argument.

If the game already computes all three axes, skip the solve entirely:

// a surface frame the game is already maintaining
camera.setBasis(right, up, forward);

setBasis takes the camera's OWN columns, where up is local +Y and so points

down the screen: the opposite sense to lookAt's, one negation apart. It also

absorbs a frame that has drifted off orthonormal, which one integrated over

thousands of steps will have.

**A camera posed this way and a world rotated into a fixed camera's frame are

the same picture.** If a game already transforms its world into view space by

hand, there is nothing to gain by switching; the value of setBasis is not

having to write that transform in the first place.

Turning a direction: slerp, not lerp

lerp walks the straight line between the two tips, which passes INSIDE the

sphere. The result points the right way at both ends, lingers near them, rushes

the middle, and comes out shorter than either input as it crosses the midpoint.

On a camera swing that reads as a lurch; on a body walking over a globe it dips

into the surface.

Vector3d#slerp turns along the great circle instead, at a constant angular

rate, and carries the length across linearly, so two unit vectors give back a

unit vector and nothing needs renormalizing.

// a turret, a camera, a ship: a facing chasing another facing
facing.slerp(toTarget, 0.25);

// frame-rate independent, the same shape as math.damp
facing.slerp(toTarget, 1 - Math.exp((-6 * dt) / 1000));

ObservableVector3d has it too and fires its callback ONCE, so

renderable.pos.slerp(...) is one notification rather than three.

Vector2d has the same method, and it is not a cut-down copy: in 2D the

whole turn is one signed angle, so there is no sine to divide by and no

degenerate near-parallel case at all. What it buys a 2D game is the wrap: a

heading damped as a NUMBER from 170 degrees to -170 degrees sweeps the long 340

instead of the short 20, and an arc between directions cannot do that.

Two degenerate pairs are handled inside rather than left to the caller, and

both turn up in ordinary game code:

• the same direction, where every path between them is the same path.
• opposite directions, where infinitely many great circles pass through

both and none is the shorter way round. The textbook formula divides by a

vanishing sine here and hands back NaN, which propagates into a position or

a matrix and the object disappears. slerp picks a stable perpendicular and

turns through half a circle instead.

That second case is why a hand-rolled version is a liability rather than a

shortcut: anything orbiting a planet reaches the antipode of where it started,

every run.

math.damp for a scalar, slerp for a direction. Same job, different

space; what damp cannot do is keep a vector on the sphere.

Depth sorting

Under Camera3d the world sorts on "depth" every frame (recursively), so

unlike 2D you *can* assign depth after addChild and it will reorder. That is

the one place where the usual z-ordering rule is relaxed — see

melonjs-renderables for the 2D rules, which still apply everywhere else.

You will often *have* to. A Container has autoDepth: true by default, so

addChild(child) with no explicit z **overwrites pos.z with the child's

index** — a real world depth in a 3D scene, and never the one you wanted. Pass

it (world.addChild(mesh, z)), or set mesh.depth afterwards. The glTF

importer turns autoDepth off on the container it loads into for this reason.

floating does not opt out of this sort. It skips the camera *transform*,

not the depth *order*. A floating child is ordered by |pos.z| alone — its

pos.x/y are screen pixels, and the camera does not move relative to it — so

the magnitude is the distance and the sign is ignored:

world.addChild(hud, -150);       // small -> nearer than anything -> on top
world.addChild(skybox, -10000);  // large -> farther -> behind everything
world.addChild(skybox, 100000);  // equally far: sign does not matter

Both hold at any camera position. A HUD given the huge z that would put it on

top in 2D lands at the far end of the level instead, with the scenery drawing

over it.

Transparency

A mesh fades by setting its opacity — there is nothing else to switch on:

ghost.setOpacity(0.4);

Meshes render in two phases. The opaque pass writes depth in sort order; the

transparent pass replays afterwards, back-to-front, blending and writing no

depth. A draw lands in the second whenever its alpha is fractional.

That default matters because the opaque path writes premultiplied colour with

blending off, so before this a faded mesh came out darkened toward black

rather than see-through — the background contributed nothing.

transparent: true when the transparency is in the TEXTURE rather than the

opacity — a soft-edged glow, smoke, a glTF material with alphaMode: "BLEND".

The automatic check reads the draw's alpha and cannot see into a texture. Watch

alphaCutoff here: it discards texels *before* blending sees them, so a soft

edge needs a low cutoff (Sprite3d drops its own default to 1/255 when you

set transparent: true). The cutoff thresholds the MATERIAL's alpha, not the

drawn alpha, so a fading cutout mesh keeps its shape instead of disappearing at

its own threshold.

The glTF loader does not set this for you: one loaded mesh can merge several

materials and the flag routes the whole mesh, so a "BLEND" material sharing

geometry with an opaque one would drag the opaque half into the transparent pass.

transparent: false pins a mesh to the opaque pass however it is faded.

depthTest: false for something with no surface to be occluded on

transparent decides the PASS, and the transparent pass turns depth *writing*

off on its own so overlapping blended draws blend instead of fighting. Whether

geometry in front of a mesh hides it is a separate question, and

depthTest is where you answer it. The default, true, is what you want for

anything genuinely in the scene — a muzzle flash behind a wall should still be

behind the wall, and this is not a synonym for "additive".

Set it false for a mesh standing in for a screen-space effect: an additive

glow carrying a world position only so it can sort and move with the thing it

belongs to. A world-space caption billboarded over a prop wants it for the

same reason, since it has no surface for the prop's own foliage to be in front

of. Left depth tested, a flat billboard is sliced along a hard straight

line the moment any geometry is nearer at some pixel. That shows up worst

around something round — a glow on a rock near a planet's limb gets cut,

and no offset escapes it, because the near surface bulges further toward the

camera than any offset you would dare apply.

const halo = new Sprite3d(0, 0, {
    image: glowTexture,
    width: rock.radius * 3, height: rock.radius * 3,
    billboard: "spherical",
    blendMode: "additive",
    transparent: true,     // the blended pass
    alphaCutoff: 0,        // do not cut the soft falloff
    depthTest: false,      // and do not let the world slice it
});

Honoured in the transparent pass on both GPU backends; the Canvas renderer has

no depth buffer and ignores it.

Blending uses the renderable's existing blendMode, so a glow is one property.

The advanced modes ("overlay", "difference", and the rest that need a

compositing pass) fall back to "normal" here, on both backends:

const glow = new Mesh(0, 0, {
    ...quad, texture: glowTexture,
    transparent: true, blendMode: "additive", alphaCutoff: 0,
});

| | |

| --- | --- |

| sorting | per object, by distance from the camera |

| intersecting transparent meshes | may pop as the camera moves — split them, or accept it |

| InstancedMesh | sorts as one object; instances draw in buffer order |

| needs | a GPU backend and a Camera3d |

Ground shadows ride the same pass — a blob is a decal, and decals are its first

client rather than a feature of their own.

Trail3d is the transparent pass's other obvious client: a camera-facing

ribbon for an exhaust streak, a speed trail or a tyre mark. It is a Mesh, so

it sorts and fogs with everything else here, and like the rest of this tier it

needs a Camera3d. Trail is the 2D one and stays flat under a perspective

camera. See melonjs-particles-and-trails.

Distance fog

Off until you ask for it, and one call on the camera:

camera.setFog({ near: 2000, far: 7000 });   // linear: name the two distances
camera.setFog({ mode: "exp2", density: 4e-4 }); // or one density
camera.setFog(null);                         // off

It is the cheapest thing that stops a 3D scene reading as flat cut-outs, and it

lets props arrive at the far plane without a visible edge.

Every parameter is optional, and the omitted ones track live. Distances

default to the camera's own clip planes, so fog cannot silently disagree with

them after a later setClipPlanes. The colour defaults to

renderer.backgroundColor and follows it, so geometry dissolves into whatever

sky you already set — including through a day/night fade. Pass color only

when the fog should deliberately differ from the backdrop:

camera.setFog({ far: 5000, color: "#8899aa" });

A Color is held by reference, so mutating it animates the fog.

Height falloff makes mist pool in low ground instead of hanging at every

altitude equally — the difference between fog reading as weather and as a global

desaturation:

camera.setFog({ near: 1200, far: 7000, fogHeight: 0, heightFalloff: 0.0015 });

heightFalloff defaults to 0, which is uniform fog — not a special case,

the same integral with the dial at zero, so leaving it out changes nothing.

Render space is Y-down, so fogHeight is the floor and density rises

*below* it; every published form of this formula assumes Y-up and has the

opposite sign.

Fog is measured radially from the camera and applied per fragment, so

it does not slide as the camera turns and does not band across large triangles.

It lives on the camera, so a split-screen or minimap view fogs independently —

and a Camera2d never fogs at all.

Per object: fog: false exempts a mesh however far away it is — for a

waypoint or objective marker that has to stay readable. It exempts the mesh and

not the ground shadow it casts: a blob is a mark on the floor and fogs with the

floor. Emissive surfaces fog

like everything else (light travelling through fog is attenuated too), so a

neon sign that should punch through wants fog: false, not a brighter

emissive.

Only meshes fog. 2D content, HUDs and floating renderables never reach the

mesh shaders, so a screen-space overlay stays clean with no work.

A custom mesh shader is not fogged unless it asks to be. Fog is compiled

into the engine's own mesh programs — #define FOG on WebGL, an enable_fog

overridable constant on WebGPU — and a shader you supply is yours: the engine

binds it as written and never substitutes a fogged variant. So a mesh given its

own GLShader keeps full contrast while the scene around it recedes. It is

safe — nothing throws, and the camera's fog is simply not applied — but it is

usually surprising.

A hosted ShaderEffect is the other case and needs nothing: its body is

spliced into the engine's own mesh program as a colour hook, and fog is applied

to the hook's result, so such a mesh fogs like any other.

To opt in, declare the same uniforms and the engine will feed them, because the

fog values are pushed to any mesh program that declares them rather than only to

the built-in ones:

uniform vec3 uFogColor;   // straight (unpremultiplied) fog colour
uniform vec4 uFogParams;  // x = mode (0 off / 1 linear / 2 exp2),
                          // y = near, z = 1/(far - near), w = density

Your vertex stage computes the distance itself — `length((uViewMatrix *

uModelMatrix * vec4(aVertex, 1.0)).xyz)`, radially so it does not swim as the

camera turns — and the blend must scale the fog colour by the fragment's own

alpha, mix(uFogColor * a, rgb, f), because vColor arrives premultiplied.

Mixing toward the unscaled colour haloes every alpha-cutout edge.

The flip side is the reason fog costs nothing when unused: with no camera fog,

the mesh programs are compiled without any of it, on both backends. It is not a

branch that is skipped at runtime — the code is not there.

Meshes

const cube = new Mesh(x, y, {
    model: "cube",          // a preloaded OBJ name — no built-in primitives
    texture: atlas,         // TextureAtlas, image, or a preloaded image name
    width: 64, height: 64,
    cullBackFaces: true,    // the default
    lit: true,
});

A mesh pivots about its model origin (0, 0, 0), not an anchor point. Under

Camera3d it opts out of the anchor offset entirely, so anchorPoint is inert

there — place the origin where you want the pivot at authoring time, or nest the

mesh under a transformed parent. The anchor is only honoured on the legacy

2D-camera path.

InstancedMesh, and when it is the wrong tool

InstancedMesh draws one mesh many times in a single draw call — the

difference between a hundred trees and a hundred thousand. glTF scenes using

EXT_mesh_gpu_instancing load as one automatically; by hand it is a Mesh

with a count:

const trees = new InstancedMesh(0, 0, { ...treeGeometry, instanceCount: 400 });
const at = new Matrix3d();               // one scratch, reused
for (let i = 0; i < trees.instanceCount; i++) {
    at.identity().translate(x, y, z);
    trees.setInstance(i, at);
}
world.addChild(trees, 0);

trees.visibleInstanceCount = 120;        // draw fewer, without re-uploading

It is not a free upgrade. One InstancedMesh is **one geometry and one

material**, and four things move from per-object to per-group:

| | with Mesh | with InstancedMesh |

| --- | --- | --- |

| depth sort | each object sorts on its own pos | the whole set has one sort key |

| ground shadow | one blob per object | one instanced draw for the set |

| removal | removeChild, indices unaffected | removeInstance(i) swaps the last instance into the hole, so any index you were holding is now wrong |

| colour | tint per object | needs instanceColors: true and setInstanceColor(i, …) |

So the question is not "how many are there" but **"does the game address them

individually"**:

• Scenery — instance it. Trees, rocks, grass, debris: the game never asks

about one of them.

• Collision-tested props — still fine. You test against positions you

already own; instancing only changes how they are *drawn*.

• Collectibles and enemies — usually not. Anything removed one at a time

makes removeInstance's swap your problem: you have to keep an index↔object

map and repair it on every removal. At small counts a pooled Mesh each is

less code and no slower.

Under a few hundred objects the draw-call saving is not what limits you

anyway — reach for it when the count is in the thousands, or when the objects

are pure scenery and it costs nothing to.

Normals are generated for you

A lit mesh needs per-vertex normals for the shader to light with. Supply them

if you have them; omit them and the engine computes them from the geometry:

const mesh = new Mesh(x, y, { vertices, uvs, indices, lit: true });
// normals derived from the triangles — nothing else to do

You do not choose flat or smooth, because the geometry already decides. Face

normals accumulate into their vertices weighted by area, so where faces share

a vertex they average and the surface shades smoothly, and where every triangle

carries its own three vertices — a triangle soup, which is how most

hand-built geometry comes out — each vertex belongs to one face and the result

is that face's normal, so it shades flat. Want faceted edges: duplicate the

vertices. Want smooth: share them.

An explicit settings.normals always wins, and an unlit mesh gets none — there

would be nothing to read them.

A lit mesh with no normals used to render fullbright, which looks like the

lighting is broken rather than absent. If an older scene suddenly picks up

shading, that is why.

Colouring a mesh

There are four levels, and picking the wrong one is the usual reason a colour

"does not apply". They all multiply together.

| level | how | use for |

|---|---|---|

| whole object | mesh.setTint("#ff8800"), or mesh.tint.setColor(r, g, b) | flash on hit, team colour, fading one object |

| per vertex | settings.vertexColors, or mesh.setVertexColor(i, color) | a gradient *within* one mesh — distance haze, a darker crease |

| per material | textureGroups, from a multi-material OBJ + MTL | a model whose parts differ, in one draw call |

| per instance | new InstancedMesh(…, { instanceColors: true }) then setInstanceColor(i, color) | a thousand copies that differ |

tint is per object. That is the trap: build a terrain as one big mesh and

you can tint the whole valley or none of it. Anything that varies *across* a

single mesh is per-vertex.

// fade a procedural terrain toward the sky the further out it goes
const ground = new Color(217, 230, 244);
const sky = new Color(207, 230, 247);
const haze = new Color();
for (let i = 0; i < mesh.vertexCount; i++) {
    const t = Math.min(1, mesh.originalVertices[i * 3 + 2] / 6000);
    mesh.setVertexColor(i, haze.copy(ground).lerp(sky, t));
}

Supply the whole array at construction when you already have it —

vertexColors takes a packed Uint32Array (the form the batchers read, so no

conversion) or one Color per vertex. A length that does not match

vertexCount throws; it is not padded, because a short array would leave

the tail of the mesh mis-coloured and that reads as a lighting bug.

Mutating the array directly is fine, but say so afterwards:

mesh.vertexColors[i] = color.toUint32(color.alpha);
mesh.needsUpdate = true;   // the retained Camera3d path uploads once

setVertexColor does that for you. Skip it and the colour applies under a 2D

camera and silently does not under Camera3d.

On a lit mesh the colour multiplies the lit result, so it behaves as albedo

rather than as an emissive override — a vertex colour will not make an unlit

face bright.

alpha hides a mesh as it hides anything else: at 0 the draw is skipped. There

is no *partial* mesh transparency though — the mesh path renders opaque, so a

mesh at alpha = 0.5 draws fully opaque rather than half see-through. Fade a

mesh out and it will stay solid until it vanishes.

Moving a mesh in 3D

Mesh and Sprite3d inherit the ordinary Renderable transform helpers, and

both are already 3D:

mesh.rotate(Math.PI / 2, new Vector3d(1, 0, 0));   // axis overload
mesh.scale(2, 1, 0.5);                             // z is a real argument

rotate(angle) alone is the 2D case (about Z); pass a Vector3d and it

rotates about that axis. Both write to currentTransform, a Matrix3d,

which Mesh folds into its model matrix — so orienting a mesh costs a matrix

update, not new geometry.

This matters most for the case it is least obvious for: a horizontal plane.

Sprite3d with billboard: false is a quad in the sprite plane — upright —

and the way to lay one flat (water, a road, a shadow decal) is to rotate it a

quarter turn about X, not to hand-build vertices. The one reason to build

geometry instead is tiling: a Sprite3d bakes its UVs 0..1 from its atlas

frame, so a surface that needs textureRepeat across many world units still

wants an explicit vertices/uvs mesh.

The mesh pass is opaque

Worth stating plainly, because the symptom does not look like a blending

problem: the mesh pass disables blending. A soft-edged texture — a glow, a

sun, a halo, anything with a gradient alpha — composites as a hard-edged disc

with a grey rim, which reads as a broken texture rather than as a missing

feature.

transparent: true routes it through the transparent pass, and `blendMode:

"additive"` on top of that makes a light source add to the sky instead of

sitting on it. This is the same feature a translucent water surface needs.

Textures under a perspective camera

A tiling texture on a large ground plane is seen at a grazing angle, and that

changes which detail survives:

• Nothing straight and parallel to the direction of travel. Every such run

converges on the vanishing point, so a few current bands on a river become a

fan of rays radiating out of the horizon. Narrowing them and tiling more

often only multiplies the rays. Short, scattered, direction-free detail

minifies into an even shimmer instead.

• textureFilter: "linear" for a floor — "nearest" opts out of mipmaps

and anisotropy, and detail aliases into a wash exactly where the camera looks

most.

Ground must outlast the props standing on it

Scenery scattered up a slope is placed by its own rule, and terrain is built by

another. When the planting reaches further than the mesh does, a low camera

sees trees standing over open sky at the corners of the frame. Build the ground

wider than the widest thing planted on it — past the profile's clamp it is a

flat plateau, so the extra columns cost two triangles and no silhouette.

Recycling has the same trap in Z: an endless runner that recycles tiles a fixed

distance behind the *player* forgets that the camera trails further back still,

and the bottom of the frame falls off the world.

Sprite3d and billboards

Sprite3d is the 2.5D workhorse: a flat sprite living at a real depth, with

billboard controlling how it faces the camera.

| billboard | behaviour |

|---|---|

| false (default) / "none" | no rotation — a flat plane in the world |

| true / "cylindrical" | yaws to face the camera, stays upright (characters, trees) |

| "spherical" | always fully faces the camera (particles, impostors) |

"cylindrical" is what you want for paper-thin characters in a 2.5D game.

Billboarding needs a Camera3d drawing the frame; under a 2D camera the quad

renders fixed-orientation. **Any other string falls through to the spherical

branch**, so a guessed value like "upright" silently gives you spherical.

A billboard takes its orientation from the camera, so rotate() has no

effect on one: a card already turned to face the camera has no free rotation

left to give. Everything that is not an orientation still applies, scale()

and meshScale included, so animating the size of a shockwave or a pickup is

the ordinary call and the cull bounds grow with it:

const wave = new Sprite3d(x, y, { image: "ring", width: 64, height: 64,
                                  billboard: "spherical", transparent: true });
wave.scale(1 + 3 * t);          // grows; `scale` accumulates, so drive it
                                 // from the authored size, not per frame

For an in-plane rotation, a bolt streak or an exhaust, use billboard: false

and orient the quad yourself.

Lighting

Light3d takes one argument — the options object (it has no x, y, z

pair; a light carries its own position). Types are "directional" (the

default), "ambient", "point" and "spot".

world.addChild(new Light3d({ type: "directional", direction: [0.3, 1, 0.2] }));
world.addChild(new Light3d({ type: "ambient", intensity: 0.3 }));
// a Vector3d works too, wherever the game already has one
world.addChild(new Light3d({ type: "spot", position: torch.pos, range: 400 }));

direction and position take an [x, y, z] array or a Vector3d, and

color takes a Color, a CSS string or an [r, g, b] array — so a value the

game already holds can go straight in. Each is read, not retained: move your

vector afterwards and the light stays where it was.

Two things still bite, both silently:

• new Light3d(0, 0, {…}) — the constructor takes options alone. JavaScript

drops the extra arguments, so options becomes the number 0 and every

setting in your literal is discarded. The light still appears, on pure

defaults: a type: "ambient" written this way is a second DIRECTIONAL light,

and the scene looks plausible enough that nobody checks. TypeScript stops

checking a literal once the argument count is wrong, so nothing inside it is

verified either.

• direction is where the light GOES, and this is a Y-down space — so a sun

overhead travels *downward* and its Y is positive. Get the sign wrong and

the scene is lit from underneath: faces that should be in shade are bright,

the ground glows and the sky-facing surfaces go dark.

  direction: [-0.35, 0.8, 0.45]   // sun overhead, late afternoon
  direction: [-0.35, -0.8, 0.45]  // lit from below — almost never what you want

position follows the same convention: a lamp above the floor has a

smaller y than the floor.

Use both halves: with a key light but no ambient, the shadow side of a mesh

goes black. With no Light3d in the world at all, a lit: true mesh falls back

to a white ambient and renders fullbright — indistinguishable from unlit, which

is why "I added lit: true and nothing changed" is the usual first report.

Other defaults worth knowing: intensity 1, color white, range 1000 (point

and spot, a stylised quadratic falloff, not inverse-square),

innerConeAngle 0 and outerConeAngle π/4 (spot).

Light2d is 2D-only and produces visible artifacts under perspective

projection. Do not combine it with Camera3d.

lit: true only does anything under a Camera3d. Lighting a mesh needs

world-space normals and a world-space fragment position, and the 2D-camera path

has neither: its vertices are already the projected output, so there is no world

space left to light in. Such a mesh degrades to unlit and warns once, naming the

fix. If you want lit 3D props in an otherwise 2D game, set

cameraClass: Camera3d on the application. (For 2D sprites, lighting is a

separate system that works under a 2D camera as normal, see the lighting skill.)

A lit mesh takes a tangent-space normal map, so relief comes from the

material rather than from geometry — settings.normalMap (a loader key or any

image-like source), or MTL map_bump / bump / norm, which the MTL loader

fetches and applies per material. The tangent frame is derived per fragment, so

the model needs no tangent attribute, and an unmapped mesh is unaffected. This

is the 3D counterpart of the Sprite normal maps in the lighting skill; they

are separate systems and a Mesh uses this one.

What the ground-shadow tier can and cannot do

Check this table before claiming a setting is unsupported. Every one of these

is read every draw, so all of them can be animated.

| setting | Mesh / Sprite3d / GLTFModel part | InstancedMesh |

|---|---|---|

| castGroundShadow | yes | yes |

| shadowGroundY | yes, the caster's own floor | yes, one plane through the SET's origin |

| shadowGroundPoint | yes on a Mesh or Sprite3d; on a GLTFModel part only by setting it after the load, since it is neither a model nor a level.load option | no, warns and ignores |

| shadowGroundNormal | yes | yes, one plane for the set |

| shadowOpacity | yes | yes |

| shadowScale | yes | yes |

| shadowOffset | yes | yes, one throw for the set |

| shadowStretch | yes | yes, one stretch for the set |

| shadowDirectionX / Z | yes | yes |

| shadowLight | yes | yes |

| height fade as the caster rises | yes | no |

| a different value per instance | n/a | no, and there is no plan for one |

The last two "no"s are the same limitation seen twice: a set gets ONE shadow

plane and ONE shared quad, so the instance's own height is dropped and nothing

can vary between instances. Everything else is one value for the whole set,

which is what a set means. A scatter over curved ground is therefore one set

per flat facet, and a flying scatter wanting a height fade has to be

per-object meshes.

shadowGroundPoint (20.9) is the one genuinely per-object setting, and the

reason is structural rather than missing work: the instanced tier builds its

plane THROUGH the set's own origin so the translation's x and z drop out of

the intercept, which is then groundY + lift / nY. An arbitrary point does

not drop out, and that same nY divisor is why the 75 degree tilt ceiling

cannot be lifted there either. InstancedMesh warns and ignores rather than

storing it, so a game cannot think it took.

Apart from that one, if a claim that a setting "does not work on an

InstancedMesh" is in your head, it is out of date: shadowOffset and

shadowStretch both reached the instanced tier in 20.8.

Ground shadows are on by default (the castGroundShadow application

setting), and need a GPU backend and a Camera3d. As a blanket default they

skip geometry with no vertical extent — a ground plane. Per object,

castGroundShadow: true/false overrides the app setting and is obeyed as

given, safeguard included; shadowGroundY names the floor the blob lands on,

shadowOpacity how dark it is and shadowScale how big it is.

On ground that is not flat, name the floor with shadowGroundPoint instead:

a point and shadowGroundNormal are a plane, and the blob leaves the caster's

vertical column to land on it. shadowGroundY cannot express that, because

the blob is always directly below the caster in world Y, which is only "below"

on level ground. It is also what lifts the tilt ceiling, since that ceiling

measures tilt from world up and so assumes a camera looking down the Y axis.

The height fade follows whichever floor you named. With a point and a

normal it is measured along that normal; with a point alone it is the

straight-line distance to the point. So on a sphere the blob fades with

ALTITUDE rather than with the vertical drop, which is the whole reason the

point exists. Two consequences worth knowing, because they fail in opposite

directions: a shadowGroundNormal facing inward makes the height negative, so

the fade clamps to full strength and the blob never fades at all; and a point

on the wrong side of the caster reads as a large height, so the blob

disappears outright rather than thinning to a ring the way a backwards

shadowGroundY does. It is also the one shadow setting that THROWS rather

than clamping or hiding: a non-finite component is rejected on the spot.

The blob is an ellipse sized to the caster's own footprint and placed at the

caster's x/z, or at shadowGroundPoint when one is set. So a tall or narrow object (a character, a tree, a pickup) shows

its shadow clearly, while a wide, flat-bottomed one resting on the floor covers

its own completely from a camera looking down at it. That is the shadow

behaving correctly, not a bug.

Do not chase it by raising shadowGroundY. Lifting the plane does not slide

the blob out from under the object, it floats the blob *up* — and past a few

units it projects over the top of the caster as a dark halo ringing it.

The simplest fix is a bigger blob, so its edge shows past the object:

mesh.shadowScale = 1.5;        // 1 is the default; 0 hides the shadow

shadowScale works everywhere, InstancedMesh included. Changing it on a

set rewrites the shared quad's four corners in place rather than rebuilding

it, as does a shadowLight that moves, so an animated sun costs no

allocation.

Or, for the look of a sun that is not overhead, use the shape controls:

mesh.shadowOffset = 1.5;       // slide it out, in blob radii
mesh.shadowDirectionX = 1;     // the ground direction it is cast along
mesh.shadowDirectionZ = 0;
mesh.shadowStretch = 2;        // and lengthen it along the same line

shadowOffset is measured in multiples of the blob's own radius, not in

world units, so one value serves a boulder and a pebble and every part of a

glTF model. About stretch - 1 is what keeps the shadow attached to its

caster. It is honoured only when the floor has been named, by

shadowGroundY or by shadowGroundPoint, because the

blob is a flat quad on one named plane: sliding it across a plane the game has

not named puts it where there may be no floor. shadowStretch is clamped to 3

and the blob fades as it pulls, so an extreme value degrades to nothing rather

than to a smear.

An InstancedMesh honours both, as one offset and one stretch for the

whole set, which is all a shared set can carry. Both ride the quad the blobs

share rather than the matrix that also places them: a stretch in the matrix

would scale the instance positions too and smear the scatter instead of

lengthening its blobs. So a scatter can sit next to loose props under the same

sun without looking lit from somewhere else.

shadowLight takes the direction from a Light3d you name, re-read every

draw, so a moving sun carries the shadows with it. Nothing is inferred: the

engine never picks a dominant light, because a scene with several has no

non-arbitrary answer and one with none has no answer at all.

A floor that is not level

shadowGroundY says where the floor is; shadowGroundNormal says which way

it faces:

prop.shadowGroundY = floorY;
prop.shadowGroundNormal = [Math.sin(0.35), -Math.cos(0.35), 0];  // a 20 deg slope

World up is (0, -1, 0), because render space is Y-down, and that is the

default. The blob is rotated onto the plane rather than projected, so it

keeps its size; it sits directly under the caster, since the normal turns it

rather than moving it, unless shadowGroundPoint says where instead. It needs

the floor named, by shadowGroundY or shadowGroundPoint, for the same

reason the offset does: tilting the bounds fallback has no anchor to turn

about.

The tilt is clamped at 75 degrees alongside shadowGroundY, past which an

edge-on blob has nothing left to show. That ceiling measures tilt from WORLD

up, which stands in for "tilted away from the viewer" only while the camera

looks down the Y axis, so it is lifted alongside shadowGroundPoint, where a

point and a normal state the plane outright.

On an InstancedMesh one plane serves the whole set: it passes through

the set's origin at shadowGroundY, faces the normal, and every blob is

projected onto it whatever height its own instance sits at, exactly as the

level case puts every blob at one height. A scatter over curved ground is

therefore one set per flat facet, each with the facet's own plane, not one

set with an averaged normal.

All three compose on a set: the stretch turns the shared quad, the offset

slides the whole set along the light, and both are carried onto the tilted

plane rather than applied flat, so a bank's planting lies along its slope

instead of hovering across it.

Give blobs to the plants that stand apart. Dense low ground cover is better

left without: a thousand blobs each as wide as its own fern sum to a darkened

hillside, not to a thousand shadows, and ground cover is the ground.

These are art direction, not a projection. A stretched ellipse is not a

silhouette and has no contact with terrain, so on ground that is not the plane

you named it will not lie on it. Direction-correct shadows want a shadow map,

which this tier does not do.

Get the sign right: the floor is a GREATER y

Render space is Y-down, so the floor an object stands on is a *larger* y

than the object itself. Nudging a shadow plane below a surface is therefore a

plus:

mesh.shadowGroundY = WATER_LEVEL + 8;   // just under the surface — correct
mesh.shadowGroundY = mesh.pos.y - 8;    // eight units ABOVE its own base — wrong

Get it backwards and the blob becomes a horizontal quad slicing through the

caster's own body. It is still drawn at full strength — the height fade sees a

*negative* height and clamps to 1, so nothing warns you — but the depth test

hides every part of it that falls inside the caster's silhouette. What survives

is a thin ring, only as wide as the blob overhangs the object.

The symptom is distinctive and easy to misread: **shadows appear to switch on as

casters approach the camera.** That ring is a few pixels on the nearest object

and sub-pixel further out, so a scene reads as "only nearby things have

shadows" — which looks like culling, an LOD stage or a fog problem, none of

which it is. Before chasing any of those, dump shadowGroundY against the

caster's own y and check which one is larger.

Collision in 3D

Use a body with a depth shape, not a hand-rolled distance check. Two shapes

carry depth — Box3d and Sphere — and any pair of them is resolved in three

dimensions; the response carries overlapNZ / overlapZ for the depth axis.

Pick by whether the thing turns. Box3d is world-axis-aligned and cannot be

oriented, so it fits walls, floors and crates and misfits anything that

tumbles. Sphere has no orientation at all, which makes it the shape for

spinning debris, projectiles, and anything spread over a curved surface: on a

globe each object's own "up" points a different way, so one axis-aligned box

cannot fit any of them.

A hand-rolled check is the usual mistake here and it costs twice — it throws

away the penetration vector the response already carries, and it is invisible

to the debug plugin's hitbox overlay, which draws from body.shapes. If the

red boxes never appear, the reason is usually that there is no body to draw.

Against a planar shape, either 3D shape degrades to its XY footprint with the

planar side read as unbounded along Z, so a sphere or box body still collides

with an ordinary collision layer, just without a depth result. And mind the

normal: a box pair separates along one world axis, a sphere along the line

between two centres, so a sphere contact fills the 2D and depth halves of the

response at the same time. See the physics skill for the full contract.

glTF / GLB scenes

Loaded through the same level director as everything else:

await loader.preload([{ name: "diorama", type: "glb", src: "data/diorama.glb" }]);
level.load("diorama", { scale, castGroundShadow, shadowGroundY, onLoaded });

// every ground-shadow setting travels except shadowGroundPoint, which is
// per object and dropped by a scene load: a loaded scene can
// be told where its sun is, and `shadowLight` is re-read every draw
level.load("diorama", {
    scale, castGroundShadow: true, shadowGroundY: 0,
    shadowLight: sun, shadowOffset: 0.6, shadowStretch: 1.8, async: true,
});

.obj / .mtl are also supported loader types.

An animated asset instantiates as a GLTFModel, which is a game object as

much as a scene: construct it straight from the parsed descriptor, and place it

with pos / depth / rotate / scale like any other renderable — the

placement drives the whole rig and composes with the playing clip.

const model = new GLTFModel(loader.getGLTF("boat"), { scale: 40, lit: false });
model.setCurrentAnimation("paddle", { loop: true });
model.pos.set(x, y);
model.depth = z;

To scatter an authored mesh yourself, read the geometry off the descriptor and

hand it to an InstancedMesh — loader.getGLTF(name).nodes[0] carries

vertices / uvs / normals / indices. That wants a single merged

primitive exported at the origin; see melonjs-3d-assets.

melonjs-3d-assets covers the rest: every level.load option, which material

features are imported, imported lights and their intensity units, node-TRS

animation (and the skinning that is out of scope), and OBJ/MTL.

Requires a GPU backend

The whole 3D tier needs WebGPU or WebGL 2. The Canvas renderer has no depth

buffer, no perspective path and no drawMesh, so a Camera3d scene does not

render correctly there — usually a black canvas. Light3d and shader effects

are inert too. A Mesh under a *2D* camera still renders, CPU-projected by

painter's algorithm and unlit.

Application does warn: constructing with a cameraClass whose

defaultSortOn is "depth" on a renderer with no depth buffer logs a

console.warn. But video.AUTO falls back silently to Canvas, so the strong

gate is to ask for a GPU backend by name — renderer: video.WEBGL (or

video.WEBGPU) makes await app.init() reject instead of misrendering:

const app = new Application(1024, 768, {
    parent: "screen",
    renderer: video.WEBGL,      // init() rejects if WebGL 2 is unavailable
    cameraClass: Camera3d,
});
await app.init();

To branch rather than fail, read app.renderer.supportsDepthBuffer after

init().

Symptom → cause

| symptom | cause |

|---|---|

| scene lit from underneath | direction Y sign — this is Y-down, so a sun overhead is +Y |

| a lit mesh renders fullbright | it had no normals — supply them, or let the engine generate them |

| a gradient across one mesh is impossible | tint is per object — use vertexColors / setVertexColor |

| a mesh stays solid as you fade it out | meshes render opaque; only alpha 0 (hidden) and 1 differ |

| a glow is cut along a hard straight line, worst near a round object's limb | a flat billboard being depth tested against the geometry behind it — depthTest: false |

| vertex colour applies under a 2D camera but not Camera3d | wrote the array directly without setting needsUpdate |

| Mesh: vertexColors has N entries, expected M | one colour per *vertex*, not per triangle or per index |

| nothing renders, or a backdrop covers everything | wrong depth sign — "far" is *larger* z when looking along +Z |

| distant objects vanish or warp | scene exceeds the default far plane; setClipPlanes |

| distant surfaces z-fight | near too small for the scene scale |

| black canvas under Camera3d | Canvas renderer (no depth buffer) — check the console.warn |

| everything flat and unlit | lit: true with no Light3d in the world (falls back to fullbright), or a mesh under a 2D camera |

| a faded mesh goes dark instead of see-through | transparent: false on it, or a 2D camera — the transparent pass needs a Camera3d |

| a soft-edged glow has hard edges | alphaCutoff discarded the soft texels; lower it |

| a glTF alphaMode: "BLEND" material draws opaque | the loader does not set transparent — one mesh can merge several materials, so set it yourself |

| two transparent objects pop as the camera moves | per-object sorting cannot order intersecting geometry |

| a floating HUD draws behind the scenery | a large \|z\| is *far* under Camera3d — use a small depth |

| fog hangs in the sky as thickly as in the valley | uniform fog — add heightFalloff so it pools low |

| mist sits on the ridges instead of the valley floor | the fogHeight sign — Y is DOWN here, density rises below it |

| distant geometry pops in against the sky | no fog — camera.setFog({}) picks up the clip planes and background colour |

| fog does not match the sky after a background fade | an explicit color was passed; omit it to track renderer.backgroundColor |

| geometry clips before it has finished fading | fog far beyond the clip far — omit the distances and they default to the clip planes |

| one marker must stay readable in fog | fog: false on that mesh |

| an object casts no visible shadow | wide and flat-bottomed — its own blob is underneath it; raise shadowScale (or slide it out with shadowOffset), not shadowGroundY, which haloes it instead of revealing it |

| a dark ring around the top of an object | shadowGroundY lifted too far, floating the blob up into the caster |

| a tilted blob never fades as its caster rises | shadowGroundNormal faces inward, so the height came out negative and clamped to full strength |

| a blob vanishes completely after setting shadowGroundPoint | the point is on the wrong side of the caster; the fade reads that as height and hides it |

| shadowGroundPoint throws | it is the one shadow setting that rejects a bad value instead of clamping: all three components must be finite |

| a scatter ignores shadowGroundPoint | InstancedMesh shares one plane across the set and warns; use one set per facet, or per-object meshes |

| a mesh sits at the wrong depth after being added | autoDepth overwrote pos.z with the child index — pass addChild(mesh, z) |

| a mesh sits half its size off | anchorPoint — only on the 2D-camera path; a Camera3d mesh pivots on its model origin |

| a billboard tips over when the camera looks down | "spherical", or a mistyped mode string falling through to it — use true / "cylindrical" |

| rotate() does nothing on a billboard | the camera owns its orientation; billboard: false and orient the quad yourself |

| the camera will not tilt its horizon to an arbitrary up | lookAt(target) solves pitch and yaw only; pass the up as the next argument, or hand all three axes to setBasis |

| the view is upside down after lookAt(target, up) | up is the direction that should appear UP ON SCREEN, and this is Y-down, so a level horizon is (0, -1, 0) |

| setBasis throws "forward must have a non-zero length" | forward is all zeros, the one case that names no direction; a short one is fine, so a surface frame that shrank as it drifted still works |

| roll is ignored when the target is straight up or down | no rotation about the view axis is determined there, so lookAt keeps the roll it had and setBasis falls back to the right axis |

| an object jumps to the camera plane | Vector3d.set(x, y) / camera.pos.set(x, y) zeroed its z |

| lighting looks wrong under a 3D camera | Light2d used instead of Light3d |

Related skills

• melonjs-renderables — the 2D z-ordering rules, anchors, custom draw
• melonjs-getting-started — Application settings and the renderer fallback