Blue noise point clouds are sets of points that look random but never clump and never leave big gaps. A new open database from Rémi Bardenet, Raphaël Lachièze-Rey and Armand de Cacqueray, shared on Hacker News on 24 October 2026, ships ready-made ones from 1,024 up to 131,072 points, in 2 to 32 dimensions.
Instead of regenerating an expensive sampler every time, you download a file, shift it randomly and reuse it. Here is what that means and why graphics and simulation people care.
Key takeaways
- The Hyperuniform Point Clouds Database holds precomputed point sets on the unit hypercube, in dimensions 2 to 16, plus 20, 24, 28 and 32.
- Sizes are powers of two, from 2^10 to 2^17 points. Tiling reaches 2^24 points, and projection lowers the dimension, both at the cost of some of the “stealthiness” property.
- Code and data are MIT-licensed and stored as HDF5, readable from Python, C++, MATLAB, Julia and R.
- Two variants exist: an isotropic one for blue-noise uses and one tuned for quasi-Monte Carlo integration.
What is blue noise, and why not just use random numbers?
A raw random generator gives you natural clumps and empty patches. Drop 1,000 grains of rice on a table at random and you will see both.
Blue noise fixes that without going to the other extreme. A grid covers space evenly but creates repeating patterns, which cause aliasing (false patterns such as moiré). Blue noise sits in between: no ordered structure, yet no clusters or voids.
In spectral terms, you take the Fourier transform of the point cloud, which shows how much energy sits at each spatial frequency. Blue noise has little energy at low frequencies and no sharp spikes. White noise, by contrast, is flat. A paper by Heck, Schlömer and Deussen, catalogued by SIGGRAPH History, argues that oscillations in a sampling pattern’s power spectrum are what cause structured aliasing.
What does hyperuniform mean here?
Hyperuniformity is the physics name for the same low-frequency idea. A pattern is hyperuniform when its structure factor, a measure of density fluctuation at each wavelength, goes to zero as the wavenumber goes to zero. The concept was introduced by Torquato and Stillinger in 2003, as summarised on Wikipedia.
The database goes further with “stealthy” clouds. Per the repository, stealthiness means the structure factor S(k) is at most 10⁻¹⁰ across a chosen exclusion set of frequencies. In plain words, those frequencies are effectively empty.
How are the point clouds made?
According to the repository, the method is spectral optimisation. You pick the set of frequencies to suppress, start from random independent points, run gradient descent on a Fourier loss, then finish with a quasi-Newton step. A non-uniform Fourier transform keeps the computation tractable, using the authors’ qmc-sampler package.
The optimisation enforces periodic boundaries. That matters for reuse: you can apply a uniform random shift to a stored cloud and get an independent replica, which gives you error bars. The README also notes that spectral methods can support sampling toward a target distribution.
Which variant should you pick?
| Variant | Exclusion set | Best for |
|---|---|---|
| l2 | Isotropic Euclidean ball | Blue-noise behaviour: rendering, anti-aliasing, particle initialisation |
| wd | Wrap-around discrepancy shape, axis-oriented | QMC integration of functions with low effective dimension |
The README lists use cases including rendering, periodic particle setups, Monte Carlo and QMC integration, and experimental design.
How do you use it?
Install numpy and h5py, clone the repo and run python src/load_data.py --dim all --ball all. To generate a fresh randomised cloud, run python src/new_data.py --dim 2 --npoints 1024 --ball l2 --seed 42. A quick_start.ipynb notebook is included. Larger 2D to 4D clouds are a separate 2.7 GB download.
What is the catch?
Precomputed means fixed sizes. Tiling and projection give you other counts and dimensions, but the authors state they lose some stealthiness. If your pipeline needs exactly 5,000 points in 7 dimensions with full guarantees, you will compromise or optimise your own.
Opinion: the real value is reuse. Optimising a good point set can take far longer than the job that uses it, so publishing the output once is a sensible trade.
What happens next
The repository lists version v5 of the dataset with a 2026 date, and at the time of writing it had few stars, so adoption is early. Watch for benchmarks against standard low-discrepancy sequences in real rendering or integration tasks. The theory behind it is well studied, including work on hyperuniform point sets on flat tori, so the open question is practical gains, not mathematical validity.
FAQ
Is blue noise the same as hyperuniform?
They describe closely related ideas. Both mean suppressed low-frequency fluctuations. Hyperuniform is the physics term, blue noise the graphics one.
Can I use the database commercially?
The repository states an MIT license for both the data and the code, which permits commercial use with attribution.
Do I get independent samples from one stored cloud?
Yes. Because the clouds use periodic boundaries, a uniform random shift produces independent replicas you can use for error estimates.
Next step: clone the repo, run the quick-start notebook in 2D, and plot a cloud next to plain random points to see the difference yourself.
Leave a comment