[compat] GlowLayer (selective glow for emissive meshes)

[sdoleg2011]

Missing API

GlowLayer from @babylonjs/core — constructor options (blurKernelSize), intensity, isEnabled, customEmissiveColorSelector, customEmissiveTextureSelector, addIncludedOnlyMesh / addExcludedMesh, dispose.

Current behavior

It exists but throws LiteCompatError ("Effect layers are not implemented in Babylon Lite").

Use case

A kitchen configurator (Babylon.js 9, ~1700 meshes in a large project) that I am moving to Babylon Lite through lite-compat. LED strips under wall cabinets and inside carcasses are thin unlit emissive boxes; a GlowLayer with a colour selector gives them their halo, and everything that is not a strip acts as an occluder, so a strip behind a door does not glow through it. Without the layer the lighting feature looks broken, and a full-frame bloom is not a substitute: it is not selective and the scene has plenty of bright non-LED surfaces.

Minimal example

const glow = new GlowLayer("ledGlow", scene, { blurKernelSize: 32 });
glow.intensity = 0.6;
glow.customEmissiveTextureSelector = () => null;
glow.customEmissiveColorSelector = (mesh, _subMesh, material, result) => {
    if (mesh.metadata?.isLed) {
        const c = material.emissiveColor;
        result.set(c.r, c.g, c.b, 1);
    } else {
        result.set(0, 0, 0, 1); // occluder
    }
};

Additional context

I have this working in my app-side shim on 1.30.0, built only from native frame-graph tasks and without touching the scene's default render task, so I think it is feasible for the compat layer. The composition, in graph order after the scene task:

1. createGeometryRendererTask (one VIEW_DEPTH attachment) — used only for its geometryDepthTexture.
2. createRenderTask({ rt: <rgba16float, no depth>, depth: geo.geometryDepthTexture, autoMirror: false, clr: true, clrColor: transparent }) with the emitting meshes added through addMesh. Because the geometry depth is _eager, it is loaded rather than cleared, which gives exact occlusion for free.
3. A half-size copy, then createBlurPostProcessTask H+V at half size and H+V again at quarter size with kernel = blurKernelSize / 2 — the same two radii Babylon.js uses.
4. An additive merge of the two blurred targets onto engine.scRT (alphaMode: 1, clear: false), scaled by intensity.

Tasks are appended with addTask + buildFrameGraphTask once the first emitter shows up, and toggled with executionEnabled, so a scene without emitters pays nothing. Side by side with Babylon.js the halo size and brightness match closely.

Three things were awkward and may be worth a look regardless of whether GlowLayer itself lands:

• No public custom full-screen pass over a RenderTarget. createPostProcessTask is internal and the createEffectWrapper API binds Texture2D, not the canvas-sized RenderTargets that frame-graph tasks produce and rebuild on resize. For the half-size copy I used a blur task with direction: { x: 0, y: 0 }; for the merge (two sources and an intensity uniform) I had to reach into the blur task's _shader.
• No way to remove a task from a scene's frame graph (there is addTask* but no counterpart), and disposeRenderTarget is not exported, so explicit targetTextures have to be destroyed by hand.
• GeometryRendererTask had two problems for this kind of scene; I opened fix(geometry): emit valid normal outputs for unlit Standard materials #747 (unlit Standard material with a normal attachment produced invalid WGSL) and fix(geometry): load a material family bridge that first appears after preload #748 (a material family that first appears after _preload throws on every frame).

I would be glad to turn the shim code into a PR against babylon-lite-compat if that is a direction you want.