AI Reality Creator — Latent-Space World Generator

Latent Vector

Terrain detail 0.55

Water level 0.32

Latent distortion 0.30

Structure density 0.35

Generated Reality

seed = 1000

Structures spawned

Land coverage

Elevation range

0.0

Regenerations

How it works

A generative model doesn't paint pixels directly — it samples a point z in a low-dimensional latent space and a learned function maps that point to a full output (an image, a scene, a world). This simulator stands in a hand-built noise field for the learned function, so you can see the same idea with a formula you can read:

warp(x,y) = x + λ·(fbm_warp(x,y) − 0.5)
height(x,y) = fbm( warp(x,y), warp(y,x), seed, octaves, persistence )
fbm(x,y) = Σ amplitude_i · valueNoise(freq_i·x, freq_i·y) / Σ amplitude_i
freq_i = 2^i,  amplitude_i = persistence^i

The four sliders are the latent vector z = (roughness, waterLevel, λ, density) fed into that function:

Terrain detail — persistence of the fractal sum; higher stacks more fine octaves on top of the base shape.
Water level — the height threshold below which a cell renders as ocean instead of land.
Latent distortion (λ) — domain-warp strength: it feeds the noise field back into its own coordinates before sampling, the same trick diffusion/GAN latents use to fold structure into structure instead of tiling it.
Structure density — probability threshold for spawning a procedural structure on a given land cell, biased toward mid elevations.

New Seed re-rolls the hash offset — a different point in "model-space" that still obeys the same governing function. Auto-Drift slowly walks z along a smooth path (independent sine phases per dimension) and regenerates the world on the fly, the same way an autonomous generative agent might continuously hallucinate new realities instead of waiting for input.

Water Lowland Highland Peak Structure

⚠ Couldn't load the 3D engineThree.js failed to load from the CDN. Check your connection and reload.