SciGraphs · Lay out a graph
Choose where the nodes of an abstract graph go. The coordinates are not in your data, they are produced by an optimizer, so this walkthrough is as much about reading a layout as about running one.
Mirrors the Layout & Positioning panel.
Before you begin
Import a graph (see Import an abstract graph) and keep it active. If the header card says No graph loaded, select the graph object in the viewport.
1. Read the status card first
Above 1,000 nodes the card flags Spring and Spring 3D as slow; above 10,000 it recommends DrL, Yifan Hu or LGL instead. Take it seriously: the NetworkX spring layouts are the slowest entries in the list and the ones most people reach for first.
2. Pick an algorithm
The Algorithm Info card under the dropdown reports dimensionality, an indicative speed and what the algorithm suits. Rather than working down the list, decide what the drawing has to answer:
- Is there community structure? ForceAtlas2 (the default), Yifan Hu (sfdp) or Spectral 3D. Densely connected groups come out as visible clumps.
- Is this a hierarchy or a process? Sugiyama for a DAG, Hierarchical 3D for a tree, Graphviz Dot for the canonical layered drawing.
- Is it too big to be drawn at all? igraph DrL or igraph LGL.
- Do I just need a reference frame? Grid, Sphere, Cube or Random. These ignore the edges entirely, which makes them useful as a before-and-after baseline.
3. Apply it once
In Quick Settings, set Scale (5.0 is the default and a good size for a few hundred nodes) and press Apply Layout Now.
Every layout operator calls rebuild_edges, which deletes and recreates the mesh edges. Any attribute on the EDGE domain survives as a correctly named, correctly sized layer of zeros. Node attributes are untouched.
Lay the graph out before you compute or import anything edge-valued, and recompute anything edge-valued that predates the layout. Notebook 01 demonstrates it on a real column.
4. Tune the parameters
Algorithm Parameters appears only for algorithms that expose any: Spring, ForceAtlas2, the igraph family, Yifan Hu and the Graphviz engines. These change what the algorithm is optimizing, not how the result is displayed.
Some parameter groups, Fruchterman-Reingold’s among them, say so in the panel itself: they are honored in Interactive Mode only, and Apply Layout Now uses defaults for those. If a slider appears to do nothing, check for that note above it.
5. Watch it converge
For a force-directed layout, Interactive Mode computes positions iteratively across the timeline so you can stop when the picture is right rather than after a fixed iteration count. Press Start Execution; the Current Status box reports iteration and energy, and says Converging once the energy settles below Auto-Stop Energy. Bake to Animation keeps the motion as keyframes; Reset restores the positions the layout started from.
6. Separate the layers
Network Splitter 3D takes a finished layout and pulls it apart along Z by Community, Attribute, Degree, Centrality or connected component. Keep Preserve XY Positions on so the layout you just computed still reads inside each layer, then press Split Network. Reset flattens it again.
What the panel will not tell you
Run the same algorithm on the same graph with the same parameters twice and the nodes are usually somewhere else. The add-on prepares a seeded generator on every layout call and then the algorithms that need it read a different one: the NetworkX spring layouts call spring_layout with no seed, igraph keeps its own RNG that is never seeded, and the cube layout draws from NumPy’s global state. Measured on the karate club, two identical imports laid out with Spring (2D, NetworkX) came out 9.0 units apart on a drawing about 10 units across. There is no seed field in this panel, so it cannot be fixed from the sidebar.
What does repeat, measured over three runs each in Layouts: Grid, Sphere, Random, igraph Kamada-Kawai, Spectral 3D, MDS 3D, Graphviz Dot, and ForceAtlas2, the only force layout in the set that consults the seeded generator. Good reason to leave the default alone.
Yifan Hu / Graphviz sfdp adds two of its own. The first sfdp call in a Blender session differs from every later one, so a figure made once and never remade is not the figure you get on a rerun. And its default 2D + Z depth mode derives Z from an eigenvector whose sign is arbitrary, so the depth axis flips between runs; setting Dimensions to 2D (flat) removes that one entirely.
The dispatcher marks ForceAtlas2 as a SPRING (2D fallback) whenever the legacy fa2 package is missing, but modern NetworkX has a real ForceAtlas2 and the layout uses it: the fallback is reported and did not happen. Separately, the graph object records the algorithm you requested on import, not the one that executed, so when a genuine fallback does occur the object still says what you asked for. If it matters which algorithm produced a figure, note it yourself.
Circular Hierarchy builds its concentric rings in the XZ plane with y = 0, while every other 2D layout here uses XY. From Blender’s default top view it is a straight line. Look at it from the front (Numpad 1), or rotate the object 90 degrees about X.
Graphviz sfdp, the engine behind both Yifan Hu and Graphviz SFDP, aborts the whole Blender process on a graph containing an isolated node, with no traceback and no chance to save: the bundled Graphviz asserts on a node whose only neighbor is itself, and an abort() inside a compiled library takes Blender with it. SciGraphs guards the case by downgrading Smoothing to None, so the layout completes but with a smoothing you did not choose. This tab has no drop-the-fragments button outside the SuiteSparse importer’s Giant Component Only; on an existing graph object, City2Graph’s Graph Tools → Remove Isolated does the job. An object whose stored edge list is empty reaches the layout as all isolated points, however many mesh edges you can see.
Next steps
Analyze a graph computes the quantities you will color by, or go straight to Style and render.