Skip to main content

Atlas

CI Wally Docs License: MIT

Atlas is a procedural generation toolkit for Roblox, written in fully-typed Luau.

  • Noise — seeded Perlin, simplex, and Voronoi (Worley) samplers in 2D and 3D, with analytical derivatives for slope-aware generation, binary island masks, plus fBm helpers for layering octaves.
  • Sampling — uniform random and Poisson disc (blue noise) point scattering, deterministic per seed.
  • Curves — monotone cubic response curves built from control points, for remapping noise (or any value) without overshoot.
  • Visualization — draggable on-screen EditableImage windows rendered through pixel buffers, with built-in presets for noise, curves, and point sets.

Installation​

[dependencies]
Atlas = "shmeatley/atlas@0.1.0"

Then run:

wally install

Quick start​

local Atlas = require(path.to.Atlas)

-- Seeded noise: same seed, same world.
local simplex = Atlas.Noise.Simplex.new(42)

-- Shape the raw noise with a response curve.
local falloff = Atlas.Curve.new({
	Vector2.new(0, 0),
	Vector2.new(0.4, 0.1),
	Vector2.new(0.7, 0.8),
	Vector2.new(1, 1),
})

local function terrainHeight(x: number, z: number): number
	local noise = Atlas.Noise.fbm2D(simplex, x * 0.01, z * 0.01, 4)
	return falloff:evaluate(noise * 0.5 + 0.5) * 100
end

-- Binary island mask: 1 for land, 0 for water.
local island = Atlas.Noise.Island.new({ seed = 42, breakage = 0.5, detail = 0.75 })
local function isIslandLand(x: number, z: number): boolean
	return island:sample2D(x * 0.002, z * 0.002) == 1
end

-- Slope-aware placement using analytical derivatives (no extra samples).
local _, dx, dz = simplex:sample2DWithDerivative(10, 20)
local steepness = math.sqrt(dx * dx + dz * dz)

-- Blue-noise scatter for trees, rocks, spawn points…
local trees = Atlas.Sampling.poissonDisc(512, 512, 8, 42)

-- See what you're generating (client-side).
local window = Atlas.Visualizer.new({ title = "Terrain preview" })
window:render(Atlas.Visualizer.Presets.noise(simplex, { scale = 0.02, curve = falloff }))

Full API reference: shmeatley.github.io/atlas

Performance notes​

  • Every sampler builds its seeded state once at construction; sampling methods are allocation-free.
  • In your own hot loops, localize the method once: local sample2D = noise.sample2D then call sample2D(noise, x, y).

The test place ships with benchmarks: build it with rojo build test.project.json -o AtlasTest.rbxl, open in Studio, and press Play — smoke tests plus per-call timings print to the output, and demo

Here is my benchmarks output from my machine:

Yes, it is faster than roblox when sampling pure 2d noise(seeded). It also supports derivatives which are extremely useful.

ROBLOX-PERLIN-2d                               4.036 ms      0.040 us/op  [n=100000]
ROBLOX-PERLIN-2d-cached                        4.074 ms      0.041 us/op  [n=100000]
Perlin.sample2D                                3.831 ms      0.038 us/op  [n=100000]
Perlin.sample2DWithDerivative                  4.751 ms      0.048 us/op  [n=100000]
ROBLOX-PERLIN-3d                               4.032 ms      0.040 us/op  [n=100000]
ROBLOX-PERLIN-3d-cached                        3.990 ms      0.040 us/op  [n=100000]
Perlin.sample3D                                5.575 ms      0.056 us/op  [n=100000]
Perlin.sample3DWithDerivative                  8.299 ms      0.083 us/op  [n=100000]
Simplex.sample2DWithDerivative                 0.062 ms      0.062 us/op  [n=1000]
Simplex.sample3DWithDerivative                 0.088 ms      0.088 us/op  [n=1000]
Voronoi.sample2D                               0.139 ms      0.139 us/op  [n=1000]
Voronoi.sample3D                               0.417 ms      0.417 us/op  [n=1000]
Voronoi.cell2D                                 0.171 ms      0.171 us/op  [n=1000]
Noise.fbm2D (4 octaves, simplex)               0.207 ms      0.207 us/op  [n=1000]
Curve.evaluate (8 points)                      0.076 ms      0.076 us/op  [n=1000]
Sampling.uniform (256 points)                  4.806 ms      4.806 us/op  [n=1000]
Sampling.poissonDisc (100x100, r=5)         1171.568 ms   1171.568 us/op  [n=1000]
Sampling.blueNoiseGrid (100x100, cellSize=3)  25.427 ms     25.427 us/op  [n=1000]
Visualizer render math (128x128)             112.179 ms   1121.792 us/op  [n=100]

License​

Atlas is licensed under the MIT License.