The plume as a distance field

This section is the design notes behind @aeronautic/afterburner: why the plume is a signed distance field, what each dial means physically, and why the tiers exist. You do not need it to use the library, but it explains why the dials are the ones they are.

What the plume isCopy Anchor Link

A cone of light has a closed form: one quadratic per fragment. A plume does not. It is a shape with cells and eddies cut into it, so the fragment shader traces it. For every pixel the plume covers:

1. A proxy hull, one instance per nozzle, gets the fragment onto the screen.
2. The ray is taken into the nozzle's own frame.
3. A bounding cylinder says where the ray enters and leaves the plume.
4. plume_bound asks whether anything burns near the current point. If nothing does for d metres, the march skips d metres in one iteration and asks again.
5. Where something does, plume_field gives its density and colour, split into the flame channel and the shock channel (the Mach disks).
6. Both are composited by Beer–Lambert, 1 − exp(−density × depth), and added over the scene.

Every nozzle is one instance, written once when it is placed. Only the throttle and the nozzle frame are uploaded every frame.

Why a distance fieldCopy Anchor Link

The shape is written as a distance rather than as a radius profile, and two things fall out of that.

The empty space is free. plume_bound answers "how far is the nearest burning thing, at least" — the smooth body with the wander applied and the shell and the chew allowed for, and nothing about the cells or the eddies, both of which only ever pull the surface inside what it allows. So a positive answer is a promise that the full field is empty for at least that far, and the ray crosses the gap between the hull and the fire in one iteration instead of twenty. Every iteration of the march either skips or samples, never both:

float gap = plume_bound(point);

if (gap > 0.0) {
  t += max(gap * uStepSafety, stride);

  continue;
}

Because a skip advances by at least a stride, the loop still covers the whole span in no more iterations than an even march would have taken. It is a strict improvement, not a trade: the worst case is what the old fixed march cost every time, and the common case — a ray clipping the edge of a plume, or crossing the tapered back half of the hull — is a fraction of it.

uStepSafety is why it is a bound and not a distance. The radius profile shears along the axis and the wander shears it again, so length(lateral) - radius runs long where the profile is steep. Two thirds covers the worst of it.

The shape is one thing. A Mach disk is a waist in the flow, so it is a waist in the surface — surface = radius * (1 - pinch * cell * lit) — and not a brightness painted over a smooth cone. The eddies displace that same surface, in metres. There is one shape, and everything that happens to a plume happens to it.

The trap: a distance field is uniform inside itselfCopy Anchor Link

This is the one thing that does not come for free, and the first version had it wrong. A ray crossing three metres of plume interior accumulates three metres of exactly the same number, so every cell and every eddy averages straight back out of the integral and what comes back is a smooth cone with all its detail confined to a silhouette one pixel wide. It looked airbrushed, and no amount of turning the shock dials up changed it.

So the same lookups feed an interior term as well as the surface:

// A pinch conserves the mass it pinches, so the narrow part is the dense part
float interior = 1.0 + uShockPinch * cell * lit;

// And the eddies thin the medium as well as chewing its edge
interior += v_flow.z * uEddyBite * (eddies - 0.5) * 2.0 * agitation;

Both are free — the noise was already fetched and the cell was already computed. eddy_bite is the dial, and at zero the plume goes back to being a lit cone with a ragged rim.