#gamedev #wfc #mesh-generation #3d-model #2d-grid #model-synthesis #procedural-gen


2D & 3D procedural generation with WFC/Model synthesis

1 unstable release

0.1.0 Jan 19, 2024

#283 in Game dev

Used in bevy_ghx_proc_gen



Bevy tracking ghx_proc_gen on crates.io ghx_proc_gen on doc.rs bevy_ghx_proc_gen on crates.io bevy_ghx_proc_gen on docs.io

Ghx Proc(edural) Gen(eneration)

A Rust library for 2D & 3D procedural generation with Model synthesis/Wave function Collapse, also available for the Bevy engine.

With Model synthesis/Wave function Collapse, you provide adjacency constraints as an input to the algorithm, and internally, a solver (AC-4 in this case), will try to generate a solution with satisfies those constraints, very much like a sudoku solver.

Altough it can be applied to do texture synthesis (mainly with bitmaps), ghx_proc_gen focuses more on grid-based use-cases such as terrain or structures generation.



cargo add ghx_proc_gen

In ghx_proc_gen, the building pieces of a generation are called Models, and adjacency constraints are defined with Socket. Every Model has one or more Socket on each of his sides.

Connections are then given between some of those Sockets, which allows Models with matching Sockets on opposite sides to be neighbours.

Let's build a checker board pattern:

  1. Start by creating the Rules for the algorithm:
  // A SocketCollection is what we use to create sockets and define their connections
  let mut sockets = SocketCollection::new();
  // For this example, we will only need two sockets
  let (white, black) = (sockets.create(), sockets.create());

  // With the following, a `white` socket can connect to a `black` socket and vice-versa
  sockets.add_connection(white, vec![black]);

  // We define 2 very simple models: a white tile model with the `white` socket on each side
  // and a black tile model with the `black` socket on each side
  let models = vec![

  // We give the models and socket collection to a RulesBuilder and get our Rules
  let rules = RulesBuilder::new_cartesian_2d(models, sockets).build().unwrap();
  1. Create a GridDefinition
  // Like a chess board, let's do an 8x8 2d grid
  let grid = GridDefinition::new_cartesian_2d(8, 8, false, false);
  1. Create a Generator
  // There many more parameters you can tweak on a Generator before building it, explore the API.
  let mut generator = GeneratorBuilder::new()
  1. Get a result
  // Here we directly generate the whole grid, and ask for the result to be returned.
  // The generation could also be done iteratively via `generator.select_and_propagate()`, or the results could be obtained through an `Observer`
  let checker_pattern = generator.generate_collected().unwrap();

If we simply print the result in the terminal we should obtain:

  let icons = vec!["◻️ ", ""];
  for y in 0..checker_pattern.grid().size_y() {
      for x in 0..checker_pattern.grid().size_x() {
        print!("{}", icons[checker_pattern.get_2d(x, y).model_index]);


For more information, check out the ghx_proc_gen crate documentation or all the examples.

More information

Model variations

In order to facilitate the rules-definition step, ghx_proc_gen can create some models variations for you automatically. This will take care of rotating all the model sockets properly.

We will take this rope-bridge model as an example:


  let bridge_model = SocketsCartesian3D::Simple {
    x_pos: bridge_side,
    x_neg: bridge_side,
    z_pos: bridge,
    z_neg: bridge,
    y_pos: bridge_top,
    y_neg: bridge_bottom,

With the above declaration, we declared our base model (with Rot0 allowed by default), and allowed an extra rotation of Rot90 degrees. Internally, when building the Rules, two models variations will be created.

When retrieving generated results, you get ModelInstances which reference the original model index as well as the ModelRotation applied to it.

You can also manually create rotated variations of a model: bridge_model.rotated(ModelRotation::Rot180) and use a different asset for it, change its weight, etc.

Coordinate systems & axis

ghx_proc_gen uses a right-handed coordinate system. However, the rotation axis used to create model variations is up to you. When using Cartesian3D, it defaults to Y+ and can be customized on the Rules. with Cartesian2D, it is fixed to Z+.

For Bevy, see the Unofficial bevy Cheatbook.


As seen in the quickstart, socket connections are declared through a SocketCollection.

Do note that sockets connections situated on your rotation axis should be handled differently if they are to be used on a model that can have rotations variations.


Rotating Model 2 in the above figures causes its top socket(s) (here B) to be different(s). For this example, we could use:

  // a socket `B` can only be connected to another `B` if their **relative** rotation is 0°
  sockets.add_constrained_rotated_connection(B, vec![ModelRotation::Rot0], vec![B]);

Let's imagine that Model 1 and 2 had different sockets declarations on their top and bottom respectively, and that these sockets were only compatible when their relative rotation is 0° or 180°:

  // a socket `model_2_top` can only be connected to another `model_1_bottom`
  // if their **relative** rotation is 0° or 180°
    vec![ModelRotation::Rot0, ModelRotation::Rot180],

See for axample the bridge_start_bottom socket in the canyon example, which can only face outwards from a rock.


Instead of collecting the results of a generator call direclty, you can retrieve them via an Observer connected to a Generator. This is what the ProcGenDebugPlugin does.

Grid loop

Grids can be configured to loop on any axis, this is set from their GridDefinition.


Cargo features

Find the list and description in ghx_proc_gen/cargo.toml


Disabled by default, the debug-traces feature will add many debug traces (using the tracing crate) to the core algorithm of the crate. Since some of those logs are on the hot path, the feature should only be enabled in debug.

When creating models, you can register a name for them with the with_name function. With the feature disabled, the function does nothing. But when enabled, the name of your models will be visible in the debug traces of the core algorithm, providing useful information about the current generation state.

The log level can be configured by the user crates (tracing::level, the LogPlugin for Bevy, ...).



Disabled by default, the bevy feature simply add some Component derive to common struct of ghx_proc_gen.

For Bevy users

Instead of using the ghx_proc_gen crate directly, you can use the bevy_ghx_proc_gen crate which depends on and exports ghx_proc_gen (with the bevyfeature enabled) as well as additional plugins & utilities dedicated to Bevy.

cargo add bevy_ghx_proc_gen

Bevy quickstart

Steps 1 to 3 are the same as in the above quickstart.

  1. To automatically spawn our assets for us, we use the ProcGenSimplePlugin
  app.add_plugins(ProcGenSimplePlugin::<Cartesian2D, PbrMesh>::new());
  1. To see something in the Bevy viewport, we setup some assets:
fn setup_generator(
    mut commands: Commands,
    mut meshes: ResMut<Assets<Mesh>>,
    mut materials: ResMut<Assets<StandardMaterial>>,
) {
  // ... Steps 1 to 3: Generator setup ...

  // Simple procedural cube mesh and materials.
  let cube_mesh = meshes.add(Mesh::from(shape::Cube { size: CUBE_SIZE }));
  let white_mat = materials.add(Color::WHITE.into());
  let black_mat = materials.add(Color::BLACK.into());*
  // We create our models asset here, in a separate collection for the sake of simplicity.
  // (We could also declare them with our models)
  let mut models_assets = RulesModelsAssets::<PbrMesh>::new();
  models_assets.add_asset(0, PbrMesh {
          mesh: cube_mesh.clone(),
          material: white_mat,
  models_assets.add_asset(1, PbrMesh {
          mesh: cube_mesh.clone(),
          material: black_mat,

  // ...
  1. Spawn an Entity with a GeneratorBundle
  // The ProcGenSimplePlugin will detect this, generate and spawn the nodes. 
  commands.spawn(GeneratorBundle {
      spatial: SpatialBundle::from_transform(Transform::from_translation(Vec3::new(
          -grid.size_x() as f32 / 2., -grid.size_y() as f32 / 2., 0.,
      asset_spawner: AssetSpawner::new(models_assets, NODE_SIZE, Vec3::ONE),


For more information, check out the bevy_ghx_proc_gen crate documentation or all the examples.

Bevy plugins

Grid plugin

GridDebugPlugin provides debug utilities for the grid-types bundlded within ghx_proc_gen:

  • Can draw a debug view of any 2d/3d grid
  • Can draw debug markers on any cells of a grid (controlled via bevy events)

Use it by inserting a DebugGridView3d bundle on your Grid entity (or DebugGridView2d, depending on your Bevy Camera).


ProcGen plugins

ghx_proc_gen does not need a plugin to work, but if you want a really quick way to get started, or are in need of some debug utilities for your generations, there are some ready-made plugins for this:

  • ProcGenSimplePlugin: Really simple, just here to generate and spawn the nodes assets. See its sources.

  • ProcGenDebugPlugin [Depends on GridDebugPlugin]: Just a bit more complex, and not focused on performance but rather on demos & debugging use-cases. You can view the generation one step at a time, see where the contradiction occurs and more. See its sources.

Both of those plugins start their work when you insert the components from a GeneratorBundle on an Entity.

Cargo features of bevy_ghx_proc_gen

Find the list and description in bevy_ghx_proc_gen/cargo.toml

  • default-assets-bundle-spawners: This feature enables some simple AssetBundleSpawner impl for basic types: (Handle<Image>, Handle<Scene>, MaterialMesh and PbrMesh). Disable the feature if you don't need them, or want to customize their implementation.
  • grid-debug-plugin compiles the grid debug plugin and its systems.
  • simple-plugin compiles the simple plugin and its systems.
  • debug-plugin compiles the debug plugin and its systems.

Compatible Bevy versions

Compatibility with Bevy versions:

bevy_ghx_proc_gen bevy
0.1 0.12


Checkerboard Unicode terrain Bevy-checkerboard Pillars Tile-layers Canyon
Grid coordinate system Cartesian2D Cartesian2D Cartesian2D Cartesian3D Cartesian3D Cartesian3D
Assets Unicode Unicode Procedural meshes .glb .png .glb
Engine None None Bevy Bevy Bevy Bevy
Camera N/A N/A 3D 3D 2D 3D

Examples videos for unicode-terrain, pillars, tile-layers & canyon are slowed down (with the ProcGenDebugPlugin for Bevy) in order to see the generation happen

[Command-line] Checkerboard example
cargo run --example checkerboard

Simple standalone example, the same as in the quickstart section.

[Command-line] Unicode terrain example
cargo run --example unicode-terrain

Simple standalone example which generates a top-down 2d terrain and displays it in the terminal with unicode characters.


[Bevy + ProcGenSimplePlugin] Bevy checkerboard example
cargo run --example bevy-checkerboard

Simplest Bevy example, the same as in the bevy quickstart section.

[Bevy + ProcGenDebugPlugin] Pillars example
cargo run --example pillars

This example generates multiple pillars of varying sizes in an empty room. Its rules are really simple with only 4 models: a void block, a pillar base, a pillar core and a pillar top.


See the keybindings

[Bevy + ProcGenDebugPlugin] Tile-layers example
cargo run --example tile-layers

This example uses Bevy with a 2d Camera but generates a top-down tilemap by combining multiple z-layers, so the grid and rules used are still 3d.


See the keybindings

[Bevy + ProcGenDebugPlugin] Canyon example
cargo run --example canyon

This example generates a canyon-like terrain with some animated windmills.


See the keybindings


Keybindings for the Pillars, Tile-layers and Canyon examples:

  • F1: toggles the debug grid view
  • F2: toggles the FPS display
  • Space unpauses the current generation
  • Right used only with GenerationViewMode::StepByStepPaused to step once per press
  • Up used only with GenerationViewMode::StepByStepPaused to step continuously as long as pressed


Rules-writing tips:

  • Start simple, then add complexity (new models, sockets and connections) iteratively. Adding one model can have a huge influence on the generation results, and may require weights tweaks.
  • Changing the Node selection heuristic may drastically change the generated results.
  • On rectangle grids, diagonals constraints are harder and need intermediary models.
  • There are often more than one way to achieve a particular result, and WFC/Model Synthesis shines when combined with other tools & effects. In particular you might find it useful to do some post-processing on the generated results (adding supports, combining models, ...).


Why "ghx" ?

  • It serves as a namespace to avoid picking cargo names such as proc_gen or bevy_proc_gen


Thanks to:



ghx-proc-gen is free and open source. All code in this repository is dual-licensed under either:

at your option.

Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in the work by you, as defined in the Apache-2.0 license, shall be dual licensed as above, without any additional terms or conditions.



~164K SLoC