1. Transitioning towards Rendering

Introducing threlte 5

Transitioning towards Rendering

November 2022, by Grischa Erbe

Hi. If you're reading this, you are probably looking for answers to why a component or hook from @threlte/core is on deprecation notice. The simple answer is that @threlte/core is evolving towards acting more like a renderer for a simpler, faster, and more flexible developer experience that works with everything you throw at it.

Why? Up until this transition, threlte has been a wrapper component library for three.js, and as such it wouldn't be able to keep up with three.js's development speed. The wrapper approach would always keep threlte one step behind.

But let's dive a bit deeper.

The history of @threlte/core

@threlte/core wrapped the most basic building blocks like meshes, lines, lights and cameras by mimicing the class system of three.js as nested Svelte components. This worked for a number of components and provided great DX and type-safety. One of the very basic three.js classes is the Object3D. It manages transformations, visibility, scene hierarchy, matrix calculations, and other fundamental things.

So naturally, there has been a component called <Object3DInstance> which uses an instance of a THREE.Object3D and with the help of some props, you can reactively manage an object instance that is part of your scene graph, convenient! Now we already established that a large part of three.js's classes base on this class, so threlte did that as well. Which meant there's for example a <MeshInstance> component that forwards all props belonging to the domain of a THREE.Object3D to its child <Object3DInstance> component and adds features that are part of a THREE.Mesh. It sounds like a sane and future-proof approach, no? But it also means hard-wiring and typing all props, events and bindings. Right from the start I gave using $$props and restProps a try but the performance drop was drastic. The architecture of nested components proved to be the natural enemy of the catch-all $$props property. So I resided with individually typed and declared props and a strong type system. Every new component meant extending and adding types, fleshing out the component itself and adding some bits and pieces that would make it feel a little bit magic. At some point I wanted to add a property userData to <Object3DInstance> which meant adding it to every single @threlte/core component by hand. It was at that point that I was worried this architecture of nested components wouldn't hold up for long.

Svelte does not have custom renderers

In the meantime Paul Henschel – the person behind react-three-fibercriticised threlte for being a wrapper instead of a renderer (or in react terms a reconciler) and while he was totally right I was still busy wrapping stuff instead of finding a way to make a renderer happen in Svelteland. Svelte does not have something like custom renderers, a functionality that react-three-fiber extensively uses to dynamically use jsx elements as three.js classes: With react-three-fiber, <mesh> is dynamically transformed to a new instance of THREE.Mesh in the root context of a <Canvas> component. But what Svelte does have is the superpower of an AOT compiler and preprocessors that can parse the markup, script and style blocks of a component and inject everything that you can think of.

The new @threlte/core at a glance

So here's the new threlte way of rendering three.js classes with the new preprocessor:

	import { T } from '@threlte/core'

	<T.BoxGeometry />
	<T.MeshStandardMaterial />

First impressions

While fleshing out the details of the component <T> and <Three>, I was constantly testing them. In an admittedly pretty naive benchmark they proved to be 2-3x faster in updating props. They're extremely flexible and compatible with existing threlte packages like @threlte/extras or @threlte/rapier. Here's the result of that testing, have a go and post your score on our Discord server!

A quick FAQ at this point

  • Why not just <mesh />?

The three namespace consists of elements that cannot be differentiated from DOM elements (e.g. <path>), also introducing types for IDE autocompletion is hard and the VS Code Svelte extension doesn't like binding to a web component.

  • Why not just <Mesh />?

We also thought about a code generator that would transform every export of three to a standalone component but this approach is prone to a bug that vite has with large Svelte component libraries. With currently over 450 exports, this did not seem like the solution.

  • What's <T>?

This component will not end up in your bundle after compilation and will be stripped by the preprocessor. It's there to provide types for autocompletion for your IDE and is the marker by which the preprocessor knows where to replace.

So how does threlte's preprocessing work?

  1. The preprocessor looks for inline components that start with the set prefix, which is T by default.
  2. It then looks up what's following that prefix and if it's existing in its import catalogue. This catalogue is the root three namespace by default. So if you can import something from three, it's there. This catalogue can be extended.
  3. If there's no match it's ignoring the component and moving on.
  4. If there is a match, the component is replaced by the component actually in charge with providing all functionality: <Three /> from @threlte/core. It also places the required imports in the script block.

Basic Example

The preprocessor enables you to quickly outline a scene graph:

	import { T } from '@threlte/core'

<T.Mesh position.x={5}>
	<T.BoxGeometry args={[1, 2, 1]} />
	<T.MeshStandardMaterial color="hotpink" />


	import { Mesh, BoxGeometry, MeshStandardMaterial } from 'three'
	import { Three } from '@threlte/core'

<Three type={Mesh} position.x={5}>
	<Three type={BoxGeometry} args={[1, 2, 1]} />
	<Three type={MeshStandardMaterial} color="hotpink" />

Extending the preprocessor

To use a module other than three, simply extend the preprocessor with the option extensions where the key is a module name and the value is an array of import names:

const config = {
	preprocess: preprocessThrelte({
		extensions: {
			// import { OrbitControls } from 'three/examples/jsm/controls/OrbitControls'
			'three/examples/jsm/controls/OrbitControls': [OrbitControls],

			// import { TransformControls } from 'three/examples/jsm/controls/TransformControls'
			'three/examples/jsm/controls/TransformControls': [TransformControls],

			// import { CustomGrid } from '$lib/CustomGrid'
			'$lib/CustomGrid': [CustomGrid]

export default config

They can now be used in your components:

	import { T, useThrelte } from '@threlte/core'
	const { renderer= useThrelte()

<T.PerspectiveCamera let:ref>
	<T.OrbitControls args={[ref, renderer.domElement]} />

Do I need to use the preprocessor?

Strictly speaking, no. You can go ahead and use the component <Three> from @threlte/core directly. It provides the same functionality and the same (shared) types but the output looks odd and is less accessible to some because of the ever repeating <Three> component. The preprocessor is just syntactic sugar on top of <Three> and brings threlte's syntax closer to react-three-fiber to be able to use its ecosystem and tooling like gltfjsx more easily. In fact, there is an intention from both sides to bring react-three-fiber tooling to threlte. I'm super excited about that!

Props FAQ

  • What props are allowed?

    • Primitive properties like numbers, strings and boolean values accept just that: <T.Mesh visible={false}> is equal to mesh.visible = false.
    • Object properties like THREE.Vector3 or THREE.Color which have a set function accept any value that you can pass to that function. So <T.Mesh position={[1, 2, 3]} > is equal to mesh.position.set(1, 2, 3) and <T.MeshStandardMaterial color="hotpink" /> is equal to color.set('hotpink').

This means that every property available on a three.js class is available (and fully typed!) on <T>/<Three>. On top of that there are props to handle the disposal of objects (dispose) and camera-specific props (manual & makeDefault).

  • How would I set the x-axis of the position?

Props can be used with dot-notation to pierce into objects. mesh.position.x = 5 translates to <T.Mesh position.x={5} />. Be aware that dot-notated props do not provide autocompletion or type checking as of now.

  • What is args?

In three.js objects are classes that are instantiated. These classes can receive one-time constructor arguments, e.g. new THREE.SphereGeometry(1, 32). With threlte, constructor arguments are always passed as an array via args. If args change later on, the object must naturally get reconstructed from scratch. This is an expensive operation and should be avoided.

  • Do the prop types change?

Yes, some prop types change. In the past it was possible to define THREE.Vector3 props as objects. Because the position prop on a e.g. THREE.Mesh has a setter function, that type is the accepted type and nothing else. This means you will need to use an array for setting the position from now on: <T.Mesh position={[1, 2, 3]} />. It is my opinion that the pierced props provide so much comfort, that you will not miss those object-based transform props: <T.Mesh position.x={1} />. The compatibility with react-three-fiber and svelte-cubed is a bonus.

  • What is let:ref?

This is what Svelte calls a Slot Prop. In regular projects they're not that much used but they prove to be extra helpful when dealing with the new rendering pattern:

<T.PerspectiveCamera let:ref={camera}>
	<T.OrbitControls args={[camera, renderer?.domElement]}>

The slot prop ref holds a reference (that's why it's called ref) to the instantiated object that children can consume. Sometimes this spares you from using bind to pass variables up and down. With bind:ref you can of course access that object instance in your script as well.

	let camera
	$: console.log(camera)

<T.PerspectiveCamera bind:ref={camera} />

Attaching things

Now there are situations where you'd want to attach some object to a property. Consider the following example:

	material={new MeshStandardMaterial()}

Looks a bit silly, no? Because we now can use any three.js class, we can attach a material to its parents' material property:


This shows that objects which cannot be part of the scene graph (THREE.MeshStandardMaterial is not extending THREE.Object3D) can be attached to parent properties and therefore allow for easy composition of objects.


All three.js imports ending with "Material" receive attach="material", and all imports ending with "Geometry" receive attach="geometry" automatically. You do not strictly have to type it out!

Attaching to nested properties

Just like with props, you can also attach the ref of a <T>/<Three> component to a nested property of its parent:


What about event handling?

A lot of three.js classes like the OrbitControls are extending THREE.EventDispatcher and offer a lot of events which you might want to listen and react to. Setting up event forwarders in wrapper components has been a manual chore for the most part. While unfortunately three.js's types do not yet contain strong type definitions for events, you can now listen to any event that a module has to offer:

	on:change={(e) => {
		console.log('The times they are A-changin')
		console.log('event details:', e.detail)

One thing to note is that <T>/<Three> does not offer interaction or viewport awareness out of the box but it can be easily implemented using trait components.

What happens to all the components in @threlte/core?

Some of the components in @threlte/core will move to @threlte/extras because that's where they belong. They are abstractions on top of three.js classes or add functionality that is special to how Svelte or Threlte does things.

Other components will be removed after a certain transitional period. This gives you time to refactor your app to the new standard while keeping things running. Be aware that "old" @threlte/core components like <Mesh> or <PerspectiveCamera> will work in perfect harmony with the new approach!

The period in which the wrapping components will be marked as deprecated will be used to collect feedback, educate and let concerns and ideas flow back into the project.

Here's a preview of what goes where:

These components provide additional value and will be moved to @threlte/extras:

  • <AudioListener>
  • <Audio>
  • <PositionalAudio>
  • <OrbitControls>
  • <TransformControls>
  • <SpotLight>
  • <InstancedMesh>
  • <Instance>
  • <Line2>

These components will be removed:

  • <OrthographicCamera>
  • <PerspectiveCamera>
  • <PositionalAudioHelper>
  • <AmbientLight>
  • <DirectionalLight>
  • <PointLight>
  • <HemisphereLight>
  • <Fog>
  • <FogExp2>
  • <Mesh>
  • <Group>
  • <Object3D>
  • <Line>
  • <LineSegments>
  • <MeshInstance>
  • <Object3DInstance>
  • <LightInstance>
  • <CameraInstance>
  • <LineInstance>

I'm undecided on these components:

  • <Layers>
  • <Pass> – We want to open up the frameloop to allow custom rendering calls and therefore also custom postprocessing pipelines. This component might be obsolete then anyway.

Using Trait Components

Threlte's way of composing functionality with trait components works very well in harmony with the new components <T> and <Three>. Be aware that these new components are supposed to be as thin of a layer as possible and do not provide event handling or viewport awareness by default. It can easily be composed in though:

<T.Mesh let:ref>

The trait components might be slightly renamed in the future to fit this use-case a bit better.

What's next?

You will see that the documentation still is using mostly wrapper components. We need your help to make that transition as smooth as possible and transfer as many examples as possible to the new rendering approach. It's important to note that while preprocessing works great for your local setup, there's no definitive roadmap of how to show code using <T> in the documentation yet.

Wrapping it up

@threlte/core is changing. And I'm sorry for that. I know that some people might not like it in the beginning. I know that some educational content (that just came out) is at least a bit deprecated. But I would not do that if I wouldn't think that this is the way forward. All of a sudden so much more is possible with just compositing Svelte components while still benefitting from tree-shaking, extensibility, type-safety and reactivity. I'm truly excited and looking forward to what you build with it.

Don't forget to leave your comments and feedback at our Discord server!