Fog a shadow can fall through

Three compute dispatches that fill a 3D texture laid out along the camera's frustum, and one permutation on the fog pass that reads it. Each cell of that texture — a froxel, a frustum-shaped voxel — holds how much light the air in it scatters toward the camera and how much it stops.

The !StandardFrame expansion emits both, named Volumetrics and Air, when the tier asks for them.

The fog every renderer starts with is a function of distance and altitude. That is enough for a horizon that fades, and it is structurally unable to do the thing people actually want fog for: it cannot know that a wall is between this patch of air and the sun. A valley under an analytic fog has no beams in it, and no amount of tuning puts them there — the information is not in the model.

Marching a volume is what buys that. Every froxel asks what light reaches it, so a froxel behind a wall is dark and the one beside it is bright, and the boundary between them is a shaft of light.

⚠ It does not replace the analytic falloff, it stands in front of it. The volume covers its own range — sixty-four metres or so — and the world does not stop there. Beyond the volume's far plane the same exponential falloff runs that always did, measured from the far plane rather than from the camera, so the two do not fog the same metres twice.

The simplest route is the tier. !StandardFrame emits the whole arrangement:

High and Epic fill the volume and shadow it; Low and Medium leave the analytic fog alone. See Choosing how hard the frame works for the four knobs — post.volumetricFog, post.volumetricShadows, post.volumetricSlices and post.volumetricFar — and how a project overrides them.

⚠ post.volumetricShadows is the one to reach for when the feature costs too much before dropping post.volumetricFog altogether — but understand what you are keeping. Off, the volume is still filled and still composited, and what goes is the beams. That is a defensible trade only where the slice count is already low; otherwise it is paying for the marching and not for the reason to march.

Authored by hand, it is two nodes:

⚠ sunDirection and sunColour are the fallback, not the answer. A host that hands the compositor a lighting feature — every host with lights in it does — gives the medium the scene's own directional light, and these two are what a document with no such host gets. Authoring them is for a frame that wants its air lit by something other than its sun.

⚠ sunColour is an illuminance in lux and ambientColour a radiance in cd/m². They are not tints. The lighting pass computes σs · p(θ) · sunColour and the composite adds the marched result straight into a frame lit in cd/m², so a value near one is a sun a hundred thousand times too dim. This shipped as [1.0, 0.9, 0.7] for several releases: the volume was filled, marched and composited correctly every frame, and on sample 3 at 512 frames the whole in-scatter came to 0.04/255 mean channel against a 0.03/255 run-to-run floor — a pass arithmetically indistinguishable from one that never ran. See Light that behaves like light; a clear midday sun is around 90 000 lux and a clear zenith around 4 000 cd/m².

⚠ volumeFar and volumeSlices must be the numbers the dispatch used. The composite finds a pixel's slice by inverting the grid's depth distribution, so a far plane or a slice count that disagrees reads the wrong slice for every pixel — a fog that is smooth, plausible, and wrong everywhere. The standard frame takes both from the same tier for exactly this reason.

The number worth understanding before any other. Raising far does not make the fog reach further; the analytic term already covers everything past it. What it changes is how the slices are spent.

The slices are distributed geometrically, so each is the same ratio deeper than the last. At sixty-four metres and sixty-four slices the nearest slice is about a centimetre deep and the furthest about four metres — which is the right shape, because a beam's edge is something you see near you and a hundred metres of haze is something you see the average of. Stretch the same sixty-four slices over a kilometre and the near slices become metres deep, which is coarse enough that a shadow edge crossing the volume reads as a staircase.

Fog that no shadow falls through is a glow. It is brighter toward the sun and it thins with altitude, and both of those the analytic falloff already did for a fraction of the price. A beam is the absence of light behind a caster, and nothing derivable from a pixel's distance and height can produce one — so the lighting pass asks, per froxel, whether the sun reaches it.

It turns itself on. There is no knob for the atlas. The pass looks for two things and uses them if the frame has both:

Either alone is useless: an atlas with no matrices cannot be projected into, and matrices with no atlas have nothing to sample. A frame missing either fills the volume unshadowed, which is a real answer rather than a failure — the height gradient and the phase peak are still worth marching.

⚠ A hand-authored frame that renames things must say so. shadowAtlas: and scenePass: on the node are how; the defaults match what the standard frame emits. A name the frame never declared is indistinguishable from a frame with no sun, so the fog goes quietly unshadowed.

The cascade selection is shared with the ground's lit path rather than restated — containment rather than depth, the two-column atlas fold, the blend across the last tenth of a cascade, the fade at the shadow distance. A froxel that picked its cascade by a second copy of that rule would shadow the air on one side of the terrain boundary and the ground on the other.

The tap is a 3×3 filter, and a froxel near a cascade edge blends two cascades — so up to eighteen depth comparisons per froxel, against one grid of 160 × 90 × slices. At sixty-four slices that is about 0.9 M froxels, and it is the dominant cost of the whole feature by a wide margin: the injection and the march are each one cheap pass over the same grid.

The levers, cheapest first:

• slices. Cost is linear in it and so is the shadow work. Dropping 64 → 32 halves it.
• far. Does not change the cost at all — it changes where the resolution is spent. Free.
• Grid width and height are fixed at 160 × 90 and deliberately not tiered. See below.

The original argument was that a shaft's edge comes from a slice boundary crossing a shadow edge — the depth axis — so lateral resolution buys sharpness the trilinear read gives straight back. That was asserted before the volume reprojected anything, and reprojection is exactly the thing that could have made lateral resolution matter more. It does not. Measured:

A froxel at view depth d is d · 2·tan(fovX/2) / 160 wide and d · ((far/near)^(1/slices) − 1) deep. At the High tier — 91° horizontal, 64 slices from 0.5 m to 64 m — that is 0.013 d wide and 0.079 d deep: the depth axis is six times coarser. At Epic (128 slices to 96 m) it is still 3.3×; at Medium (32 slices to 48 m), 12×. Whichever axis the camera's motion drags the history along, the one that was already several times coarser is the one that shows.

And reprojection made width more expensive, not less. The scattering volume is now a pair, so the frame holds four volumes rather than three: at rgba16f, 160 × 90 × 64 is 7.4 MB each and about 30 MB in total. Doubling to 320 × 180 makes that 118 MB — for a sharpening along the two axes that were already the fine ones.

⚠ What reprojection did introduce is a lateral cost that did not exist before: every frame the history is resampled at a non-integer grid offset, so fast lateral camera motion smears the volume along the screen axes at roughly one froxel per frame of accumulation. It is bounded — the blend weight decays it — and it is worst exactly when the eye's own acuity is lowest. If you find a scene where it reads, feedback is the lever, not the grid width.

⚠ A froxel has no surface normal, so the slope bias and the normal offset that a wall's shadow needs both vanish here — the air is biased by the constant term alone. This is not a simplification: a volume of air has no face to be oblique to and nothing to lift a sample off. The constant bias is the scene pass's own, deliberately, so the shaft and the shadow it belongs to meet.

The sun is not the only thing a medium scatters. A street light in mist has a cone under it, and that cone is the same in-scatter the sun's beam is — a phase function applied to whatever radiance reaches the froxel. So the lighting pass also walks the frame's punctual lights, per froxel.

It turns itself on too, on exactly the shadowing rule. The pass looks for two buffers and uses them if the frame published both:

⚠ The standard frame publishes neither, so this is off unless you build a frame that culls. The standard frame lights its lamps per object; the cluster cull is a node a hand-authored document adds. That is not a defect to work around — an object-sized light list is chosen per object, and a fog volume covers the whole frustum, which is the shape that choice serves worst. A frame that wants lamps in its air runs the cull, and then the fog reads the same lists the walls do.

Two things are worth knowing before you turn it on:

• The lamps are unshadowed in the medium. The punctual shadow atlas is a composed feature of the scene pass, and this is a compute dispatch that composes nothing — so a lamp behind a wall lights the air on both sides of it. The ground's lit path carries the same debt. The sun, which is what a shaft is made of, is the shadowed one.
• They arrive one frame late. The buffers are published while the scene pass executes, and the dispatches are built before it runs. So the first frame after a cut has fog the sun lights alone.

There is no cosine here. A surface receives light across a tilted face; a froxel has no face, so what weights a lamp's radiance is the phase function — the same one the sun's term uses, with the same phaseG. A lamp is therefore brightest in the air when you are looking almost into it.

The three dispatches sit between the shadow passes and the scene, and the composite sits after the temporal accumulation. Those pull in opposite directions and both are deliberate:

• The dispatches need shadows and lights and not scene colour, so they belong before anything draws. Saying that as a declared edge is what puts the barriers in.
• The composite must be after TAA. The accumulator averages radiance and must only ever see the finished frame — fog added before it would be smeared by a reprojection meant for surfaces.

⚠ Splitting the feature across the accumulator is only safe because the volume does its temporal work in its own space rather than the screen's. A volume inside TAA would be reprojected as though it were a surface, which it is not.

The grid is coarse — 160 × 90 by a few dozen slices — and a shadow edge crossing it lands on a different sample the moment the camera moves a fraction of a froxel. What the eye reads from that is not softness; it is crawl. So the lighting pass keeps a history:

• Each frame the sample point inside every froxel is offset by a Halton (2, 3, 5) step, and the injection and the lighting pass are handed the same offset. Three axes, not two: the depth axis is the one a shaft's edge lives on.
• The froxel's world point is pushed through the previous frame's view-projection and read back as a grid position — tile from xy/w, slice from the grid's own logarithmic distribution applied to w. Only w and xy/w are used, never z, so none of this depends on reverse-Z or on where the far plane is.
• The two are blended at feedback, default 0.9 — about ten frames to answer a light that moved.

Only the in-scatter is averaged. The extinction is what the medium is, written each frame from a closed form, so blending it would make a cellar's cleared air bleed out into the valley over several frames. And the march does not jitter: it runs after the blend, so an offset there would be a wobble in the finished picture rather than detail paid for over several frames.

⚠ A document that names scattered: itself switches this off, and switches the jitter off with it. The pair has to be memory the node owns — a resource the graph declares lives for one frame, and next frame's history would be aliased over something in this one. Jitter without a history is the same coarse grid asking a different wrong question every frame, which is the crawl rather than a cheaper way out of it, so the two go together.

Ground mist in a valley — thick at the bottom, thinning with altitude:

A back-lit haze, where the anisotropy is what makes looking toward the sun bright:

⚠ ambientColour of zero is not "no ambient", it is a valley that goes black whenever the sun is behind the viewer. A phase function is normalised over the sphere, so one directional light contributes almost nothing outside its forward peak; what makes air visible from every angle is what arrives from the whole sky. The same term is why deep water is blue rather than black.

Three fields on a post-process volume's settings override the node where the volume applies:

⚠ 0 and "unset" are different answers, and this is the field where the difference bites. Leaving volumetricDensity unset means this volume has no opinion about fog — whatever is underneath it stands, so a volume that only darkens the grade leaves the mist alone. Setting it to 0 means *there is no fog here*, which is an opinion and a strong one. If the two were flattened, every volume in the level that cared about anything else would silently delete the fog.

See Making a room look different for how volumes overlap, blend and resolve by priority — the volumetric fields follow exactly those rules and need no shape, weight or priority of their own.

1. density — how much there is. Everything else is shape.
2. far — where the resolution is spent. See above.
3. phaseG — 0 is an even glow, 0.7 is air, 0.9 is a searchlight beam.
4. ambientColour — the floor the fog never goes below. ⚠ Zero is not "no ambient", it is a valley that goes black whenever the sun is behind the viewer: a phase function is normalised over the sphere, so one directional light contributes almost nothing outside its forward peak. The shadowing multiplies the sun's term only, for the same reason — shaded air still sees the sky.
5. slices — raise it if a shadow edge crossing the volume bands, and lower it first if the feature costs too much. It is the one knob that moves the shadow work.

• Working the frame after it is drawn — the analytic fog this stands in front of, and the rest of the chain.
• The frame, as one node — where the two nodes land and why.
• Choosing how hard the frame works — the tier knobs.
• Light that behaves like light — the units the sun colour is in.