Gaussian Blue Noise (2022)
https://arxiv.org/abs/2206.07798
Automated Summary
This article introduces Gaussian Blue Noise, a project developed through arXivLabs, a framework that allows collaborators to create and share new features on the arXiv website. The project's focus is on generating blue noise masks that are compatible with Gaussian mixtures. Gaussian Blue Noise aims to improve the user experience of arXiv's community, adhering to the platform's values of openness, community, excellence, and user data privacy. The project encourages users to engage with arXivLabs if they have ideas for valuable features or improvements.
Usually if i don't understand the relevancy of a trending post on HN or never heard an abbreviation/ product name the answer is always that its something to do with ai. If anyone else wonders why blue noise is needed https://xchhuang.github.io/bndm/
Blue noise is broadly useful in imaging, you can use it for high quality dithering, or for perceptually smoother sampling in raytracers for example.
Here you can see the same rendering using white noise (true randomness) or blue noise, with the same number of samples:
https://www.shadertoy.com/view/3sfBWs / https://i.imgur.com/uDybnPm.png
Not only in imaging. More generally, blue noise is often used with all kind of Monte Carlo simulations, since the more evenly spaced sampling can speed up convergence. Raytracing can be seen as one such MC simulation
In path tracing blue noise does not increase convergence, but does reduce the perceptual error.
The upper bound of the approximation error of random Monte Carlo methods decreases with O(1/sqrt(N)), whereas the error for quasirandom Monte Carlo methods (omitting some details) decreases with O(1/N) (where N is the number of samples) [0]. I don't think path tracing is an exception here.
I would even argue the opposite: For the same variance I would expect the error to be more perceptible for (blue noise) quasirandom MC, because it can lead to regular patterns in the noise.
[0]: https://en.m.wikipedia.org/wiki/Quasi-Monte_Carlo_method
Thank you. Now I understand the use case.
Blue noise is not specific to AI at all, it's often needed for wherever you need your random to look random to the human eye: true random/white noise, doesn't look random to us, because it has elements that look too clumped up to each other, blue noise has nicely separated elements in places that look random.
If you're willing to watch a developer go through some concrete reasons why blue noise is useful, Casey Muratori has some episodes dedicated to exploring ways of calculating light propagation in a 3D environment in his HandMade Hero series. (The first mention I could find for "blue noise" is in ep. 428: https://guide.handmadehero.org/code/, so I would start from there)
I notice that the actual noise has more moiré patterns than traditional blue noise. But it does probably sample better over a time series.
Yeah, this annoyed me about the paper. Clearly there are significant visual issues with patterns forming, but they just hand wave it away and basically say "according to all these common quality metrics, this method is superior". These patterns make it a non-starter for a ton of the use cases for blue noise, the authors should have been far more humble in making their case and addressing counterpoints.
There are definitely cases where you want the frequency domain power spectrum to be as close to perfect as possible, and this paper (GBN) is a step above the previous SoTA (Blue Noise Through Optimal Transport).
However I agree about the visual result… Qualitatively, the GBN results look almost like WSJ illustrations, with tight, flowing patterns. In contrast, GBN looks noisier but much more organic, like film grain.
Whether the visual difference will matter depends on the application. If you're reducing grayscale to 1 bit, as in the paper's examples, you're very aware of any patterns. However if you're reducing a floating point color image to 8 bits per color channel, you may never be able to spot the difference.
The published version of this paper is open access and has additional clarifications, as well as a supplementary video by the lead author Abdalla G. M. Ahmed: https://dl.acm.org/doi/10.1145/3550454.3555519
Check out all these other super cool papers too: https://abdallagafar.com/
[dead]