Skip to main content

3D Autotiling: Expanding and Optimizing

In the last post we were able to randomly generate “correct” looking 3D shapes. However, my goal was something that a user or game developer could create levels with. Let’s look at our requirements:

  • The user should have direct control of the generation. BorisTheBrave describes calls this “Driven WFC”.
  • We need a lot of tiles. That would take a lot of time to model for a game developer. This isn’t totally avoidable, but we can optimize our workflow.
  • If the user is going to interact with this, the algorithm for solving the generation needs to be fast. Our naive implementation needs a major upgrade.

Driving WFC

Initial Constraint

Our method of “driving” WFC will be specifying which cells aren’t empty as an initial contraint.

In the first post I described a dual-grid. Because our tiles describe transitions, we can’t say a single cell is filled. We must use at least 4 non-empty tiles in 2D, or 8 tiles in 3D to create a manifold shape.

cross grid

If a user fills the coordinates: [(0, 0, 0), (0, 1, 0), (0, 2, 0)] we will expand that to 16 coordinates. It’s not 24 because the affected cells overlap.

For each those expanded coordinates, we remove empty from the possibility list. The cells not included in that list have their possibility list reduced to only the empty prototype. The full implementation is here.

This is the result:

Ambiguous Cases

The example above should be a plain rectangle. The problem is our constraint says “every cell must have a non-empty tile”. That is satisfied.

Special “empty” Tile

We could expand the constraint to say “all our tiles must be connected on a non-empty face”.

    "name": "empty",
    "mesh": "",
    "rotation": 0,
    "north": "empty;empty_only",
    "east": "empty;empty_only",
    "south": "empty;empty_only",
    "west": "empty;empty_only",
    "top": "v_empty;v_empty_only",
    "bottom": "v_empty;v_empty_only",

We can give the empty prototype a special socket empty_only.

  • An empty tile can allow the empty socket from any tile.
  • A tile with a mesh that has an empty socket shouldn’t connect to other other meshes outside faces.

In the “constrain to neighbor” step, we modify our rules a bit:

if not cell.possibly_empty():
    if "empty" or "v_empty" in allowed_sockets:
        allowed_sockets -= {"empty", "v_empty"}
        allowed_sockets += {"v_empty_only", "empty_only"}

filled cells can’t use their empty side to connect to other filled cells now.

The Donut Case

Stopping empty sockets from connecting to each other restricts us from getting some valid results. It is also insufficient. Half of the connecting faces of the L tiles are not empty, so our special rule isn’t triggered. How can we fix this?

Simplified Marching Cubes

marching cubes

Again, we go back to looking at Marching Squares, except this time Marching Cubes. In that algorithm, we set a bit on the cell’s tile_index that corresponds to each “filled” corner from the dual grid. We can expand it to 3D, generating a cube_id based on which octants contain the mesh.

The cognitive load is too much for manually labeling the tiles. We can simplify this to just count the number of filled octants. We’ll call that the volume of a tile. This reduces the manual labeling workload.

When we’re setting our initial conditions, we find the volume for each cell, then constrain it to all the tiles with that volume:

corner_counts = {}
for coord in crossgrid:
    for coord in expand_coord(coord):
        if coord not in corner_counts: 
            corner_counts[coord] = 0
        corner_counts[coord] += 1

protos_by_voume = itertools.groupby(prototypes, lambda p: p.volume)

for coord, cell in grid:
    volume = corner_counts[coord]
    cell.possibilities = copy(protos_by_voume[volume])

This ends up being much cleaner. No special casing empty or looking for fully surrounded cells. And it gets us the result we want.

Impossible Configurations

You may have noticed a while ago, or maybe you noticed when we brough up marhcing squares. If a square has 4 corners, or 4 bits, there are 2⁴ tiles to cover every case. In 3D this explodes to 2⁸ (256) tiles. Our manually created tileset cannot handle all possible user input.

Generating Tiles

While we can rotate and flip our tiles to increase the number of tiles, we can’t be sure we covered all the cases. Instead it’s easier to generate most of the tiles from smaller parts. At the same time, we can annotate the tiles with the “volume” data instead of doing that manually for 256+ tiles.

Enumerating Tiles

To ensure we don’t miss any possibilities, let’s enumerate all the cases in code. Intuitively, there can’t be 256 unique cases. Let’s figure out how many are left if we ignore duplicates.

bits per octant

We can take an integer between 0 and 255 and convert it to or from a 2x2x2 array of “filled” or “empty”.

def int_to_cube(c_int: int) -> np.ndarray:
    s = format(c_int, "#010b")
    o = np.array([int(s[i]) for i in range(2, len(s))]).reshape((2, 2, 2))
    return o

def cube_to_int(cube: np.ndarray) -> int:
    return int("".join(cube.reshape(8).astype(str)), 2) ```

We're using Python 3 with NumPy because `ndarray` is very convenient.
This code doesn't need to run in our game engine and will end up
being more useful in Blender scripts, which must be Python.

Similar to 2D, many of cases are transformations of other cases. We can generate
the unique cases pretty easily:

def possible_tiles():
    seen = set()
    unique_cubes = set()

    def add(cube, unique=False):
        v = cube_to_int(cube)
        if unique and v not in seen:

    # skip 0 and 255 because they don't need a model
    # they're empty and interior tiles
    for i in range(1, 255):
        if i not in seen:

        # register the transformations as "seen"
        # so that we will ignore them in future iterations
        rc = int_to_cube(i)
        for rot in range(4):
            rc = np.rot90(rc, axes=(0, 1), k=rot)
            seen.add(cube_to_int(np.flip(rc, axis=1)))

    return unique_cubes

if __name__ == "__main__":
    tiles = possible_tiles()

53 unique cases. Not bad! We can safely ignore the 0 and 255 cases. Because 0 is empty and 255 is completely on the interior, so we can’t actually see it, there is no need to provide a model for them.

Less Modeling

If I were to manually build the 53 models, there would certainly be repetitive substructures. While iterating on art, it would be very annoying to update every variation to keep the look consistent.

What if we imagine each octant as a tile? It can’t be that hard to auto-tile a 2x2x2 grid. In the talk Beyond Townscapers, and on twitter Oskar uses this image to show tiles for a dual-grid.

oskar’s beautiful dual tiles diagram

I tried to classify the square cases.

  • Empty
  • Corner
  • Edge
  • Diagonal
  • Bend
  • Full

Those are the cases I want to end up with at the end. The combinatorial explosion into 3D is big. Let’s cut those into quadrants to make the dual-dual tiles. We end up eliminating the Diagonal case. Only four models are needed to start:

basic tiles

Expanding to 3D isn’t that bad. We just need walls, edges and bends along the Up axis. That can be accomplished with some simple rotations.

It doesn’t fully cover it though. We have to be able to transition vertically between flat tiles or edge tiles. We’ll call these “lip” tiles since they’re kind of skinny. There are 5. 4 for to match up to the base cases and a 5th that is a corner between two lips. It’s basically the “outward” part of an “outward blob cut” approach.

lip tiles

Finally, we duplicate all of these and flip them upside down so we have some dedicated for the top layer of our 2x2x2 grid that transition from the ceiling downward instead of from the floor upward.

In total There are 26 tiles to model. Very manageable!


Assembling WFC Tiles

To combine the tiles, we rely heavily on some naming/classification. The names of the Blender object concatenate:

  • Their base type - Lip, Corner, Bend, Edge
  • Their Layer - Top, Bottom
  • Their Direction - Horizontal, Vertical, Horizontal + Vertical, Corner

We can then determine which micro-tile belongs in an octant based on which neighboring octants are also filled. The implementation is here. I wrote it on a long flight and jetlag affected the readability. The rules are:

  • If 4 octants are filled in the same “plane”, this is either flat or a “lip”.
  • If 3 adjacent octants are filled, this is a 3-way bend.
  • If 2 adjacent octants are filled, this is a 2-way bend.
  • If 1 adjacent octant is filled, this is an edge.
  • If 0 adjacent octants are filled: this is a corner.

Orienting the Tiles

We know which category of tiles go where but the simple rules above don’t explain how to rotate them into alignment. I was far too lazy to figure this out. Turns out I wrote an algorithm for this already. WFC!

If we run our adjacency script, then setup a 2x2x2 WFC grid and set the initial conditions of each cell to have all the rotations of the mesh that we’re sure we want… it will spit out exactly what we need!

Additional Tiles

We’ve discussed how we could use 8 bit integers to represent a cube divided into 8 octants. This idea inspired me to abandon mesh data as a means for adjacency data in favor of 2D images.

A “socket” in our WFC we only have a few requirements:

  • Capture the contour of a single face of a tile
  • We can easily check if it is flipped or symmetrical.
  • Easy to determine if it’s a rotation of another tile.

What are the drawbacks of the current approach?

  • An opaque string does not preserve the visual representation.
  • Suffix markers: s (symmetrical) r (rotations), f (flipped) have to be trusted. We can’t reproduce them from their base case without the mesh.
  • Meshes must be perfectly constructed to have verts sitting on bounding box faces.
  • Small errors in the model can cause incompatible tiles.

The current model makes for a nice workflow in the happy path, but failure cases are a pain to debug or notice. Instead, we can use this cube data to create simple 2D images:

sockets from cubes

We can auto-generate these for our standard 256 prototypes. This does place a new burden on the artist: they must manually draw images for additional tiles outside of the base set. Manually drawing low-resolution images turns out to be easier than makinag sure 3D meshes perfectly align. It is often easier to hide seams with details like stones or other decorations. The image approach enables that sort of artistic liberty.


Base64 encoding and decoding images could work. Instead I decided to over-engineer this a bit. I don’t want to rely on using any library’s Image abstraction. My images don’t need much resolution at all.

The base tiles could be captured with a 2x2 image. I landed on using 7x7 monochrome images encoded into 64-bit integers. 2x2 isn’t enough if we create detailed tiles down the road. An 8x8 image would use 64 bits, and we want leftovers.

Using some leftover bits, we can store information like whether the image is the flipped version of another image. Now when we compare a horizontal socket, we don’t need to loop over the image and actually flip it. We can simply figure out the unique set of images in our tileset before hand, and mark some as flipped. The check becomes a == b ^ (1 << 63). We might use another bit to store whether the image is symmetrical so we know NOT to flip the bit.

For example, 0b00000111 would look like:

socket 7

The advantages we get:

  • Dense serialization
  • Fast compatibility checks
  • Canonical unflipped socket.
  • Canonical 0 rotation socket.

Those last two are very useful properties. With the old approach, the first hash we saw would get a name, and the transformations would be derived from that name. With this approach, the “canonical” untransformed image can be determined by performing the transformations to the image and sorting by the integer value.

Images from Octants

3 oct sock

Converting from 2x2 to an arbitrary size isn’t too hard. Just loop through and fill up the corners. For odd numbered resolutions, like 7, a corner would be 3x3. There is additional logic to fill these gaps where it makes sense.

def gen_sock(cube_id: int, face_name: str, seen_socks: Sockets, bitmap_size=7) -> int:
    cube = int_to_cube(cube_id)
    face = get_face(cube, face_name)

    assert bitmap_size % 2 == 1  # must be odd for 1 pixel gutter in center
    fill_size = int(bitmap_size / 2)
    socket_bitmap = np.zeros((bitmap_size, bitmap_size), dtype=int)
    for x in range(len(face)):
        for y in range(len(face[x])):
            if face[x, y] == 0:
            sx = x * (bitmap_size - fill_size)
            sy = y * (bitmap_size - fill_size)
            ex = sx + fill_size
            ey = sy + fill_size
            if x == 0 and face[x + 1, y] == 1:
                ex += 1
            if y == 0 and face[x, y + 1] == 1:
                ey += 1
            socket_bitmap[sx:ex, sy:ey] = 1

    return encode_socket(socket_bitmap)

Manually Authored Tiles

socket painting

The whole reason we’re using WFC and not Marching Cubes is to allow for interesting shapes to emerge under certain conditions. We can texture some cubes parented to manually modeled tiles that extend our base set of 53. The sockets for the base set are simple and easy to reason about. We can be a bit clever designing socket images for our custom tiles.

Luckily the Blender Python API (bpy) gives us a representation that is compatible with NumPy. We can keep our authoring workflow fully inside of Blender.

imgs = [node for s.material.node_tree.nodes if node.type == 'TEX_IMAGE']
img = np.asarray(imgs[0].pixels).reshape((21, 28, 4))

Optimizing the Algorithm

The workflow for creating tiles is optimized, and we may have squeezed out a bit of performance by comparing integers instead of strings. It’s still nowhere near fast enough.

Current State

For each cell, we might make a random choice. For each random choice, we do a propagation for each face of a cell. There’s a chance a single cell with can be affected in circular paths before the propagation is complete. This means we’re worse than O(N^2) propagations.

At the deepest part of this nested loop/recursive iteration we have an expensive computation:

allowed_sockets = [allowed_socket(other[opposing_face]) for other in other_cell.possibilities] 
self.possibilities = [p for p in self.possibilities if
                      len(p.sockets_for_face(my_face).intersection(allowed_sockets)) > 0]

Adjancy Table

Instead of doing all the work to check socket compatibility during generation, we can do it upfront. Consider each Prototype as having an id which is its index in the initial list.

adjacency = []
for proto in prototypes:
    allowed = {} 
    for face in faces:
        allowed[face] = allowed_for_face(proto, face, prototypes)    

def allowed_for_face(proto, face, prototypes):
        for i, other in prootypes 
        if compatible(proto, face, other)

We can reduce the work in the loop to be an intersection of sets:

allowed_protos = set()
for proto in other_cell.possibilities:
self.possibilities = allowed_protos

Bit Set

Bit representations of cubes and images got me thinking: we don’t have a crazy number of protos. What if the possibilities list was just a set. Each bit represents one prototype. In Python, a number can have more than 64 bits. In other languages, you’ll need to roll an abstraction over a list of integers. We have 256 base tiles, plus some manually created ones. Let’s say our max tilset size is 512.

If possibilities is now something we can use the binary & operator on we get some micro-optimzations:

  • We don’t necessarily need the heap.
  • Single instruction intersections.
  • Maybe it’s cache efficient? I don’t do this kind of optimization often.

This could be called a micro-optimzation, but for large tilesets it is probably worth it. Instead of doing O(N) union of hash tables, we can do O(N / sizeof(int)) bitwise & operations. For a large tileset of 2048 tiles this is only 32 & operations compared to 2048 hash table lookups.


I know this was a long post and could have been split into multiple articles, but I just wanted to get this out. The game I was planning to make with this just isn’t something I have time to work on. This is the first time I’ve written a tech blog and I’ve done it just so I can share this work in some form.