Global illumination and ambient occlusion

Two techniques at two scales. DDGI carries one bounce of indirect light across the room; GTAO darkens the creases that irradiance probes are far too coarse to resolve. Both are fully dynamic, and both are compute effects you can switch off without touching scene code.

DDGI, at a glance

Dynamic diffuse global illumination over a signed-distance-field voxelisation of the scene. Probes trace rays against the SDF, accumulate radiance into an irradiance atlas, and use Chebyshev occlusion — a mean and variance of hit distance per probe — to reject light leaking through walls. It is fully dynamic: move a wall and the bounce follows.

Ground-truth ambient occlusion runs in the AfterScene compute phase and is applied to indirect diffuse only, so it composes with glTF occlusion maps instead of fighting them.

Copy

// In the app constructor, before
// graphics initialization:
RenderQuality.DefaultGlobalIllumination
    = GiMode.Ddgi;

// At runtime, through the live tier:
RenderQuality.Current.GiIntensity = 0.6f;
RenderQuality.Current.AoIntensity = 0.8f;

GlobalIllumination is a mode read at initialization — it bakes shader variants and allocates the probe chain — so switching it later changes intensity, not cost. The engine documents which knobs are live and which are locked rather than hiding the difference.

DDGI · GTAO · TAA

How the probe volume is put together

The defaults describe a 16×8×16 grid — 2048 probes — over a 32-unit cube, each probe casting 128 rays, with half the grid refreshed per frame. Those numbers are chosen so DDGI is affordable on a laptop GPU, not so it is the best it could be.

The volume follows the camera, so a 32-unit cube is a moving window on indirect light rather than a bound on scene size. Large open scenes still lose distant bounce — a single cascade of probes is the honest limit here.

Leak control and temporal behaviour

Most of what separates usable DDGI from unusable DDGI is leak rejection and how fast probes forget. These are the knobs for both.

What the SDF is built from: 64 proxies

The probes do not trace against your triangles. They trace against a signed distance field built from proxy volumes that controls contribute during the frame — and GiProxies.MaxProxies is 64.

A control joins the field by overriding CollectGiProxy and calling TryAdd or TryAddSphere; both return false rather than throwing when the budget is spent. The two surface properties that matter are on Control:

The practical consequence: proxy the walls, floor and large blockers, and leave the clutter out. Sixty-four coarse volumes describing the room beats sixty-four volumes describing the furniture.

GTAO

Ground-truth ambient occlusion, computed from scene depth and normals in the AfterScene phase and published as compute://gtao/ao. It handles the scale DDGI cannot: the darkening where a chair leg meets the floor sits well inside one probe cell.

AO is applied to indirect diffuse only. Applying it to direct light is a common shortcut that produces dark rims on lit surfaces, and it is not what happens here — which also means AO will look weaker in a scene with no ambient or GI to occlude.

Seeing what the GI is doing

Two debug effects exist for exactly this. Sdf3DViewEffect renders a slice through a distance field of the same kind the probes trace, and DdgiEffect publishes its own slice under compute://ddgi/sdfslice alongside the irradiance and depth atlases. Both are addressable textures, so putting one on screen is a Sprite2D away. If GI looks wrong, the SDF is usually where the answer is: an object without a proxy is simply not in the field.

Post-processing and compute effects covers how those debug views are registered, and Apps/Engine shows them wired to on-screen toggles.