Skip to content


Repository files navigation

Poisson disk sampling

Poisson disk sampling can be used to procedurally sample a region of space. It's not as messy as random placement or as uniform as grid-based methods. This solution is inspired by Herman Tulleken's article which I've extended to support sampling regions using varying sampling radii and sample separation distances (using layers - a layer defines a sample's min / max radius and separation distance). The approach uses brute force so could be optimised (see further work section below)

Some applications:

  • Sampling an image / regions of an image
  • Procedurally placing objects of varying sizes in a game with no overlap

How to use

To use in your project simply add poisson.h and poisson.cpp to your project and then define your layers (see the example in main.cpp).

To build the demo:

git clone
cd poisson_disk_sampler
cmake -B build -G "Visual Studio 15 2017" -A x64 .
cd build
cmake --build . --config Release

The output is an array (an std::vector) of layers where each layer contains an array (an std::vector) of circles:

PoissonDiskMultiSampler sampler(/*refer to main.cpp*/);
PoissonDiskMultiSampler::PointListArray layers;
for(const auto& layer : layers)
    for(const auto& circle : layer)
        // the circle represent's the bounds of an object in the current layer


Discard points by implementing the placeObject method to perform your filtering (i.e. discard points in scene meshes or using simle formulae):

struct Distribution : public PoissonDiskMultiSampler::RealFunction2D
virtual bool placeObject(int layerIndex, float x, float y) {
    return int(x) % 2 == 0;
Distribution distribution;
PoissonDiskMultiSampler sampler(-(SIZE / 2.f),
                            -(SIZE / 2.f),
                             (SIZE / 2.f),
                             (SIZE / 2.f),
                             200*SIZE, // nb. set to zero to fill until full
                             minDist.size() > 1,

The above will discard items placed on uneven columns to create a furrowed field effect for instance:

Similarly you can implement a filter in terms of noise to selectively place objects in coherent clusters:

Performance and further work

Here are some performance statistics for tile sizes 8,16,32,64,128,256,512 for a distribution containing 3 layers. NB. The culling of overlapping objects accross layers uses a brute force approach and so could be improved upon (see checkPoint()):

Tip Jar / Patreon

If you find this project useful and want to buy me a coffee then you can do so via my page by downloading my free software and making a donation as part of that process here. Alternatively if you want to keep this code monkey in bananas then you can support me over on Patreon, thanks!