Question

I’m trying to implement 2D Perlin noise generation in C++, and some implementations I found use no seed at all (here, here or here). Other implementations take a seed value to get different noise depending on the noise value.

However I found example code where one added the seed value to the function parameters calculating the noise value for each octave (see PerlinNoise::Total() in the linked code). Another one uses a 3D seed function and uses the fixed seed value as the z value (couldn’t find the example just now). Other articles suggest using other noise functions.

So my question would be, what the best way would be to add a seed value to Perlin noise generation is. Given the same seed value, the same noise values should be generated. If the solution would be to have a custom noise function, I would be interested if it could be implemented using Boost.Random (or C++11’s Standard C++ Library classes).

Edit: To answer what I mean with “best” way: What’s the best way that gives me Perlin noise like it was supposed to work, e.g. a gradient noise function.

Answer 1

Since no one is going to write up an answer from the comments, I’m trying myself. Please upvote when I’m correct, comment when not 🙂

There are several implementations and example code that (try to) implement Perlin noise. First, there is the Improved Noise reference implementation from Ken Perlin himself.

Case 1: Improved Noise reference implementation

The noise function takes three double values and outputs a value. When generating a 2D bitmap using x and y, and keeping z constant, one gets the well known Perlin noise pattern. When z is varied between 0.0 and 1.0, the noise clouds seem to “change” slowly. So a seeding method that sets z, e.g. z = 10.0 * seed, could work for “seeding”.

Another way to seed the noise function would be this: If you always just get noise in a range of [0.0; 64.0[ for x and y, one could seed the noise by adding an offset to x, y or both when calling the noise function: noise(x + 64.0*seed, y + 64.0*seed).

Case 2: Tutorial style Perlin noise code

Then there is an implementation of Perlin noise (adapted and used in many other Perlin noise tutorials) that have a base noise function like this (pseudocode):

function Noise2(integer x, integer y)
    n = x + y * 57
    n = (n<<13) ^ n;
    return ( 1.0 - ( (n * (n * n * 15731 + 789221) + 1376312589)
       & 7fffffff) / 1073741824.0);    
end function

My main skepticism came from the magic numbers and the trust of the authors of these pages that the formula leads to uniformly distributed noise. Other authors added the seed value somewhere in this formula.

The solution to add a seed to this type of Perlin noise implementation is to write a function that uniformly distributes output values for given x and y values (and by returning the same value for the same x and y values, of course). This function can be written using Boost.Random (code not tested):

double Noise2(int x, int y)
{
   uint32_t seeds[3] = { uint32_t(x), uint32_t(y), seed };
   boost::mt19937 rng(seeds, seeds+3);
   boost::uniform_real<> dist(0.0, 1.0);
   boost::variate_generator<boost::mt19937&, boost::uniform_real<> >
      die(rng, dist);
   return die();
}

The random number generator has some ctors, among them one that takes a range of uint32_t’s that determine the initial state of the RNG.

There also are libraries that generate coherent noise, such as libnoise, that may be of help here.

Simplex Noise

I didn’t ask of Simplex noise, but the one implementation (from Stefan Gustavson) I found uses a similar technique (some precomputed tables) like Ken Perlin’s reference implementation, and could be seeded just like case 1 above. Commenter Robinson mentioned seeding when generating the look-up table, but I don’t know how that would work.