AI Art Latent Space Explorer

Latent Vector z = (z₁, z₂)

z₁ 0.00

z₂ 0.00

Semantic Directions

Style warp 40%

Detail (octaves) 4

Readouts

z₁, z₂

0.00, 0.00

Decoder seed

Path length

0.00

Field energy

0.00

How it works

A generative model (GAN / VAE / diffusion decoder) learns a function G that maps a low-dimensional latent vector z to an image: G(z) → pixels. Nearby points in latent space decode to visually similar images, so a straight-line walk through z traces a smooth path through image space — the "latent manifold".

This simulator stands in a hand-built decoder for the neural one, using domain-warped fractal noise so the same continuity property holds and is fully inspectable:

base(p)   = Σᵢ 0.5ⁱ · noise2D(2ⁱ·p, seed)         (fractal Brownian motion, i = 0..octaves-1)
warp(p)   = p + w · (base(p + offset₁), base(p + offset₂))
G(z)      = base( warp( center + 1.5·z ) )          → pixel brightness → HSL color
flow(p)   = (∂G/∂y, -∂G/∂x)                         → curl-like field driving the brush particles

z₁, z₂ sliders — the latent vector; dragging them moves the marker on the 3D manifold on the right and redraws the artwork.
Style warp / Detail — extra "semantic direction" parameters (like directions found in a real model via PCA / InterFaceGAN) that bend and sharpen the decoded image without changing z itself.
Animate Latent Walk — drives z(t) along a Lissajous path, the real-time analogue of interpolating G(lerp(z₀,z₁,t)) between two latent codes.
Reseed Decoder — swaps the noise seed, simulating a different trained network (a different local optimum) with the same architecture.
The purple dots are a flow-field "brush": their velocity is the local curl of the same field G, so their swirl pattern is a direct visual read of the decoder's gradient structure.

Latent-space marker Decoded artwork plane

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