Plexus

Each frame, Plexus takes the point cloud from the Points input (capped at 1500 points on the CPU path; if nothing is connected, a small 120-point fallback cloud is generated so the node is never empty). It then finds candidate pairs whose true 3D distance lies inside the [Min Distance, Max Distance] band, using a spatial grid so the search costs O(n·k) instead of O(n²).

Candidates are ordered by the Order mode, then selected greedily under two caps — Max Connections per point and Max Segments total (hard safety cap 50 000 segments):

Selected segments are shortened symmetrically toward their midpoint by Line Length, with per-pair deterministic jitter from Line Variability (seeded — the pattern never flickers between frames). Distance Fade gives every line a tolerance-centered soft edge: a connection smoothly fades out as its length approaches Max Distance (and fades in near Min Distance when a minimum gate is set) instead of popping when points drift in or out of range. The fade width is a fraction of the [Min, Max] window.

The cloud is projected 3D → 2D with an automatic perspective camera centered on the cloud's centroid: the camera distance adapts to the cloud's depth extent (camDist = zHalf × 2.5 + 2.0), so genuinely 3D clouds get gentle foreshortening while a flat cloud (z ≈ 0) projects exactly to itself — no toggle, no tuning.

Color Mode selects how lines and points are colored. Mode 0 uses the solid Line Color / Point Color; modes 1–7 sample a connected Gradient node along a chosen axis of the cloud's bounding box:

GPU mode (the GPU toggle) moves neighbor search and line building into compute shaders with an indirect instanced draw — the CPU never touches a single line. Use it for big or heavily animated clouds. Two documented differences from the CPU path: Order is not applied (pairs land in grid order), and Max Connections becomes a deterministic mutual k-nearest-neighbors rule — each point links to its K nearest points that also chose it back, with K capped at 16.

Depth of field (DoF > 0) is a fast per-element effect, not a blur pass: an element whose depth lies outside the [Focus Distance ± Focus Range] band gets dimmer (alpha drops by up to 92 %) and larger (bokeh-style radius growth up to 4×), with a smoothstep transition over Falloff. A band of the cloud stays crisp while the rest defocuses.

A rebuild cache hashes the point cloud plus every input; when nothing changed, the previous frame's geometry is returned in ~0 ms, so static patches cost almost nothing.

Plexus is a generator: it renders to an offscreen texture exposed on the Image output (with Width/Height reporting 2 × 2 Grafit units) and needs a Canvas node to appear on screen. Output is premultiplied alpha; coordinates follow the Grafit convention (origin at viewport center, shorter edge maps to [-1, 1]). Setting Enable to false clears the texture.

Parameter interactions worth knowing: Min/Max Distance define the connection band and the fade window; Line Length and Line Variability act on the already-selected segments; DoF works on raw point z, so it needs actual depth variation in the cloud to show.

• Classic animated plexus: Geometry to Points (Sphere or Cube, Random distribution) → Point Animation (drift) → Plexus — the standard floating-network motion graphic
• Audio-reactive web: Audio FFT → Remap → Max Distance, so connections grow and collapse with the music; add Glow for a club look
• Growing-tree constellation: L-System Live Tips output → Plexus Points — the advancing growth front becomes a slot-stable web that never flickers
• Face/pose mesh: a Vision landmark structure → Plexus with Max Connections 3 and Nearest order for a wireframe-scan look over a person
• Depth-graded network: a 3D cloud with Color Mode Z-Depth + Gradient node, plus DoF 0.8 with a narrow Focus Range for a cinematic rack-focus feel
• Particle flocks: for thousands of moving points, enable GPU and let the compute path build the k-NN web every frame

• Like every generator, Plexus needs a Canvas node to be visible — the Image output is an offscreen texture until a Canvas composites it
• Max Distance is the main cost driver: a large band over a dense cloud explodes the candidate count. Cap it with Max Connections / Max Segments before reaching for GPU mode
• The CPU path caps the cloud at 1500 points; anything beyond is truncated. Switch on GPU for big clouds — but remember Order is not applied there
• Gradient coloring only activates with a real Gradient cable and a non-Solid Color Mode; unconnected structure inputs are deliberately ignored
• Distance Fade is what makes motion look organic — at 0, lines pop in and out; at 0.5+, they breathe
• Line Length below 1.0 with some Line Variability turns the web into scattered midpoint dashes — a great "data noise" texture
• Anti-aliasing is global (the app-wide AA multiplier); there is no per-node AA quality control