Latent Walk (Refik Anadol — Unsupervised / Machine Hallucinations)

Math essence

Latent chain L ∈ R^{N×D} is built as a smoothed random walk and rescaled to a hypersphere. A fixed random projector P ∈ R^{2×D} maps each row to 2D; the walker advances by spherical-linear interpolation slerp(L[i], L[i+1], t) with t advancing by 1/walk_steps per frame, then plots the projected point on a fading translucent canvas.

Parameters

Reference

First 60 lines of reference.py. Full source: library_external/anadol_webgpu/latentWalk/reference.py

"""latentWalk — Refik Anadol inspired latent-space walk.

Inspired by Refik Anadol Studio's *Unsupervised — Machine Hallucinations —
MoMA* (2022), in which a StyleGAN2 ADA model trained on MoMA's archive
is "walked" through its 1024-dimensional latent space.  The artist
describes the visualisation as a real-time 7D trajectory (x, y, z, r, g,
b, time) drawn from points in the latent embedding.

Here we collapse the idea to a tractable numpy analogue: a chain of N
points in a synthetic ``D=64`` dimensional space (a smoothed random
walk — the "latent walk"), projected to 2D via a fixed random matrix
``P : R^64 -> R^2``.  Each frame we integrate a step along the chain
(``slerp`` interpolation between the current pair of latent points),
drop a coloured point at the projected location, and recolour the
trail by timestamp so the path itself encodes time.

Reference:
- https://refikanadol.com/works/unsupervised/
  ("Unsupervised — Machine Hallucinations — MoMA", RAS, 2022)
- https://en.wikipedia.org/wiki/Refik_Anadol
"""
from __future__ import annotations

from typing import ClassVar

import numpy as np


class LatentWalk:
    """A synthetic latent-space walk projected to 2D.

    Parameters
    ----------
    dim : int
        Square canvas side length (px).
    chain_len : int
        Number N of latent points in the synthetic embedding chain.
    latent_dim : int
        Dimensionality of the synthetic latent space (default 64, in the
        spirit of small GANs; Anadol's StyleGAN used 1024).
    walk_steps : int
        Number of cumsum smoothing steps used to build the latent chain
        (controls curvature / smoothness of the projected path).
    projection_seed : int
        Seed for the fixed random projection ``P`` (so the same params
        always project the same chain to the same path).
    point_color_mode : str
        ``"time"`` colours each point by chain-index (``t/N`` along a
        blue→magenta gradient); ``"depth"`` colours by the latent
        walk step index (paler near idle, brighter when actively
        integrating); ``"spiral"`` modulates hue by slerp angle.
    """

    family: ClassVar[str] = "anadol_webgpu"
    sketch: ClassVar[str] = "latentWalk"

    def __init__(
        self,
        dim: int = 600,
        chain_len: int = 1200,