Skip to content

Instantly share code, notes, and snippets.

@eevee
Last active March 2, 2024 08:48
Show Gist options
  • Save eevee/26f547457522755cb1fb8739d0ea89a1 to your computer and use it in GitHub Desktop.
Save eevee/26f547457522755cb1fb8739d0ea89a1 to your computer and use it in GitHub Desktop.
Perlin noise in Python
"""Perlin noise implementation."""
# Licensed under ISC
from itertools import product
import math
import random
def smoothstep(t):
"""Smooth curve with a zero derivative at 0 and 1, making it useful for
interpolating.
"""
return t * t * (3. - 2. * t)
def lerp(t, a, b):
"""Linear interpolation between a and b, given a fraction t."""
return a + t * (b - a)
class PerlinNoiseFactory(object):
"""Callable that produces Perlin noise for an arbitrary point in an
arbitrary number of dimensions. The underlying grid is aligned with the
integers.
There is no limit to the coordinates used; new gradients are generated on
the fly as necessary.
"""
def __init__(self, dimension, octaves=1, tile=(), unbias=False):
"""Create a new Perlin noise factory in the given number of dimensions,
which should be an integer and at least 1.
More octaves create a foggier and more-detailed noise pattern. More
than 4 octaves is rather excessive.
``tile`` can be used to make a seamlessly tiling pattern. For example:
pnf = PerlinNoiseFactory(2, tile=(0, 3))
This will produce noise that tiles every 3 units vertically, but never
tiles horizontally.
If ``unbias`` is true, the smoothstep function will be applied to the
output before returning it, to counteract some of Perlin noise's
significant bias towards the center of its output range.
"""
self.dimension = dimension
self.octaves = octaves
self.tile = tile + (0,) * dimension
self.unbias = unbias
# For n dimensions, the range of Perlin noise is ±sqrt(n)/2; multiply
# by this to scale to ±1
self.scale_factor = 2 * dimension ** -0.5
self.gradient = {}
def _generate_gradient(self):
# Generate a random unit vector at each grid point -- this is the
# "gradient" vector, in that the grid tile slopes towards it
# 1 dimension is special, since the only unit vector is trivial;
# instead, use a slope between -1 and 1
if self.dimension == 1:
return (random.uniform(-1, 1),)
# Generate a random point on the surface of the unit n-hypersphere;
# this is the same as a random unit vector in n dimensions. Thanks
# to: http://mathworld.wolfram.com/SpherePointPicking.html
# Pick n normal random variables with stddev 1
random_point = [random.gauss(0, 1) for _ in range(self.dimension)]
# Then scale the result to a unit vector
scale = sum(n * n for n in random_point) ** -0.5
return tuple(coord * scale for coord in random_point)
def get_plain_noise(self, *point):
"""Get plain noise for a single point, without taking into account
either octaves or tiling.
"""
if len(point) != self.dimension:
raise ValueError("Expected {} values, got {}".format(
self.dimension, len(point)))
# Build a list of the (min, max) bounds in each dimension
grid_coords = []
for coord in point:
min_coord = math.floor(coord)
max_coord = min_coord + 1
grid_coords.append((min_coord, max_coord))
# Compute the dot product of each gradient vector and the point's
# distance from the corresponding grid point. This gives you each
# gradient's "influence" on the chosen point.
dots = []
for grid_point in product(*grid_coords):
if grid_point not in self.gradient:
self.gradient[grid_point] = self._generate_gradient()
gradient = self.gradient[grid_point]
dot = 0
for i in range(self.dimension):
dot += gradient[i] * (point[i] - grid_point[i])
dots.append(dot)
# Interpolate all those dot products together. The interpolation is
# done with smoothstep to smooth out the slope as you pass from one
# grid cell into the next.
# Due to the way product() works, dot products are ordered such that
# the last dimension alternates: (..., min), (..., max), etc. So we
# can interpolate adjacent pairs to "collapse" that last dimension. Then
# the results will alternate in their second-to-last dimension, and so
# forth, until we only have a single value left.
dim = self.dimension
while len(dots) > 1:
dim -= 1
s = smoothstep(point[dim] - grid_coords[dim][0])
next_dots = []
while dots:
next_dots.append(lerp(s, dots.pop(0), dots.pop(0)))
dots = next_dots
return dots[0] * self.scale_factor
def __call__(self, *point):
"""Get the value of this Perlin noise function at the given point. The
number of values given should match the number of dimensions.
"""
ret = 0
for o in range(self.octaves):
o2 = 1 << o
new_point = []
for i, coord in enumerate(point):
coord *= o2
if self.tile[i]:
coord %= self.tile[i] * o2
new_point.append(coord)
ret += self.get_plain_noise(*new_point) / o2
# Need to scale n back down since adding all those extra octaves has
# probably expanded it beyond ±1
# 1 octave: ±1
# 2 octaves: ±1½
# 3 octaves: ±1¾
ret /= 2 - 2 ** (1 - self.octaves)
if self.unbias:
# The output of the plain Perlin noise algorithm has a fairly
# strong bias towards the center due to the central limit theorem
# -- in fact the top and bottom 1/8 virtually never happen. That's
# a quarter of our entire output range! If only we had a function
# in [0..1] that could introduce a bias towards the endpoints...
r = (ret + 1) / 2
# Doing it this many times is a completely made-up heuristic.
for _ in range(int(self.octaves / 2 + 0.5)):
r = smoothstep(r)
ret = r * 2 - 1
return ret
@veluca93
Copy link

If anybody is interested, here is an implementation as a C++ python module (with a lot of speed hacks) that exposes (at least) the same interface (and also shares most of the algorithm).
On my computer, it seems to be more or less 10-20 times faster.

@levibessa1
Copy link

If anybody is interested, here is an implementation as a C++ python module (with a lot of speed hacks) that exposes (at least) the same interface (and also shares most of the algorithm).
On my computer, it seems to be more or less 10-20 times faster.

Hello Veluca93 , what sequence should i execute the files in C++ ?

@thorhunter1
Copy link

Usage example? Gives me constantly values of 0.0.

@nathan-sixnines
Copy link

@thorhunter1

Try calling the factory with values inside the interval [0-1].

For example

for i in range(frameSize):
    for j in range(frameSize):
        noise[i,j] = PNFactory(i/frameSize,j/frameSize)

instead of

for i in range(frameSize):
    for j in range(frameSize):
        noise[i,j] = PNFactory(i,j)

@Wouterr0
Copy link

Wouterr0 commented Mar 25, 2020

The improved version of Ken Perlin uses 6t**5 - 15t**4 + 10t**3 instead of 3t**2 - 2t**3 as smoothstep function.

@xavriley
Copy link

Original blog post with instructions and explanation is here https://eev.ee/blog/2016/05/29/perlin-noise/

@KaufNation
Copy link

Is this O(1) to calculate any random location on the continuous curve, or is this O(n)?

@Crowkk
Copy link

Crowkk commented May 15, 2021

Correct me if I'm wrong but there is an inconsistency in this snipet:

    for grid_point in product(*grid_coords):
        if grid_point not in self.gradient:
            self.gradient[grid_point] = self._generate_gradient()
        gradient = self.gradient[grid_point]

        dot = 0
        for i in range(self.dimension):
            dot += gradient[i] * (point[i] - grid_point[i])
        dots.append(dot)

When you call self.gradient[grid_point] (say grid_point = [2,3]) you are calling the gradient at row 2 and column 3; But in "position" space it is actually the reverse: the second row is on y = 2 and third col at x = 3 so when you make (point[i] - grid_point[i]) you're calculating the wrong difference because it is reversed. It should be grid_point[::-1][i] at least for 2D;

@N3RDIUM
Copy link

N3RDIUM commented Dec 4, 2021

This code is very useful for terrain generation. I'll try using it in my game!

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment