Wave Function Collapse

This is a free online Wave Function Collapse generator. Give it a small sample image, like a piece of a dungeon, a map, a maze or a texture, and it creates a new, bigger image that looks the same everywhere you look closely: every small square of the result (3 x 3 pixels, for example) is a square that appears somewhere in the sample. Try one of the samples, or use your own pixel art or tile map, and download the result as a PNG image or as a JSON map of colors or tiles.

Wave Function Collapse is a popular procedural generation algorithm for game developers: it creates endless variations of levels, dungeons, cities, terrains and textures that follow the style of a hand-made example, without writing any rules. Watch it work step by step: the cells that are not decided yet show the average of the colors they can still become, and the image "collapses" into a final result.

Everything runs directly in your browser. No image is uploaded to a server, and nothing needs to be installed.

How to use Wave Function Collapse?

To create a new map, level or texture from a sample, follow these steps:

1. Choose a sample - Click on one of the samples: a dungeon, islands, a maze, a cave, a platform game level, a tile map, bricks or a town. Each one comes with the settings that suit it best.
2. Or use your own image - Drop a small image in the box, or paste it with Ctrl+V. The algorithm works best with pixel art of a few colors, up to about 64 x 64 pixels. If your image is a tile map, type the size of its tiles in "Tile size": the result is then made of the same tiles.
3. Watch it generate - The output starts as a blur of all the possible colors. Cell after cell, the algorithm makes a choice and the image becomes sharp. Choose the speed of the animation, or "Instant" to see only the result.
4. Click on the image for a new output - Every click picks a new random seed. Write down the seed of an output you like: typing it again recreates exactly the same output.
5. Adjust the patterns - A bigger pattern size copies bigger pieces of the sample, a smaller one mixes them more freely. Turn on the rotations and mirrors for top-down maps and textures, to get more variety from a small sample.
6. Set up the output - Choose its width and height, make it seamless to use it as a repeating texture, or turn on the ground for a side view with a floor at the bottom.
7. Download the result - Save it as a PNG image, either with big pixels as shown or at its real size (one pixel per pixel of the sample, or the real tiles), or as a JSON map with the color or tile number of every cell, to use it in a game or a level editor.

Settings

Below are the parameters you can configure to generate the output.

What is Wave Function Collapse?

Wave Function Collapse (WFC) is an algorithm published by Maxim Gumin in 2016 that generates images similar to a sample. Its name is borrowed from quantum mechanics: before a measurement, a particle is in a superposition of several states at once, and the measurement "collapses" it into one of them. In the same way, at the start every cell of the output can be any pattern of the sample, and the algorithm collapses the cells one by one.

This tool uses the overlapping model of the algorithm:

1. Read the patterns - Every N x N square of the sample is a pattern (with its rotations and mirrors, if the symmetry is on). The algorithm counts how often each one appears, and checks which patterns can overlap each other when they are shifted by one cell in each direction.
2. Observe - The cell with the lowest entropy, the one with the fewest possibilities left, is collapsed to one of its patterns, chosen at random. The patterns that appear more often in the sample are chosen more often.
3. Propagate - Neighbor cells overlap, so the choice removes the patterns of the neighbors that do not agree with it. These removals remove other patterns further away, and so on, like a wave, until everything agrees again.
4. Repeat - The algorithm observes and propagates again until every cell has a single pattern. If a cell loses all its patterns, there is a contradiction: the algorithm backtracks (undoes its last choices) or starts over.

The algorithm never needs rules written by hand: the sample is the rule. This made it very popular in games, which use it to build cities, dungeons, terrains and whole worlds, for example in the games Caves of Qud, Bad North and Townscaper. It is related to the texture synthesis methods of computer graphics, and to constraint solving: a WFC output is a solution of a constraint satisfaction problem, found with the "minimum remaining values" heuristic of these solvers.

The original algorithm also has a simple tiled model, where the sample is not an image but a set of tiles with hand-written rules of which tiles can be next to each other. This tool takes another way to work with tiles: with the "Tile size" option, a tile map is read like a pixel art image where every tile is one "pixel", so the overlapping model learns the rules from the map itself, and you don't have to write them.

Tips for good samples

• Keep it small and simple. 16 x 16 to 48 x 48 pixels and 2 to 6 colors are enough. Every different N x N square is one more pattern, so photos, gradients and antialiased images have too many patterns to work.
• Show every case you want. The output can only contain squares that are in the sample: if a corner, a junction or a crossing should appear, draw it at least once.
• Make lines and walls 1 or 2 pixels wide. With a pattern size of 3, a structure wider than 2 pixels has an inside that the patterns cannot see past, and its size becomes random.
• Make the sample tileable if its edges continue on each other, and keep "The sample wraps around" on: it adds the patterns that cross the edges.
• Use symmetry to get variety from very small top-down samples.

Frequently Asked Questions (FAQ)

Yes, this tool is 100% free and works directly in your browser, with no installation and no sign-up required.

No. Your sample image is read and the output is calculated by your own browser, so the image never leaves your computer.

Yes. The outputs made from the built-in samples are yours to use in any project, including commercial games. For outputs made from your own images, the rights are the same as the rights you have on the sample.

Some samples and settings can lead to a contradiction: a cell where no pattern of the sample fits. Backtracking solves most of them. If it still fails, try another seed, a smaller output, a smaller pattern size, or more symmetry, which gives the algorithm more patterns to choose from.

Every different N x N square of the sample is a pattern, and the algorithm compares every pattern with every other one. Photos and smooth images have thousands of different squares. Use pixel art with few colors, a smaller image, a smaller pattern size, or less symmetry.

Lower the pattern size (from 4 to 3, or from 3 to 2), turn on the symmetry, or draw a sample with more variations. With a big pattern size and a small sample, there are only a few ways to fit the patterns together.

Download the JSON map: it has the color (or tile) number of every cell, row by row. A small script can then place a tile, a wall or an object for each number. For a tile map, the "tiles" list tells where each tile is in your sample image, so the number can be matched with your tileset. The PNG at 1 pixel per pixel can also be read as a map, one pixel per cell.

Noise like Perlin noise and the plasma fractal creates smooth random terrain, and a maze generator follows one fixed set of rules. Wave Function Collapse follows the style of any sample you give it: the same algorithm makes mazes, dungeons, cities, carpets or brick walls, depending on the sample.