Beachcombing
Specimens 000 / 300

An online artwork in six movements

Beachcombing

When the tide goes out, what does it leave behind?

Follow the tide

I.

Gathering

Combing an archive is a kind of beachcombing. From a public collection of shell photographs (134 species), the search keeps three hundred: sixty kinds, five specimens each, chips and scars included.

Not all of them are from the sea; some are land and river snails.

II.

Drying

Laid out in the sun, colour fades and form remains. Each photograph is turned into a depth map. Pale means near, dark means far. The shell is now measured in light and shadow.

Depth Anything V2 Small · 320 × 320 · backgrounds cleaned for relief

III.

Learning

An autoencoder squeezes each relief into sixty-four numbers, then tries to rebuild it from those numbers alone.

128²64128²

240 shells teach it; the 60 in red are held back to test it. Past the line, each shell is shown as rebuilt: softer, keeping the form but not the chips.

80 rounds of training · held-out error down 94.5%

CNN · 128 × 128 input · 1.33 M parameters · best model at round 71

IV.

Sorting

Shells whose sixty-four numbers are alike are put together. k-means makes six piles, sorted by form and not by name. The shell at the heart of each pile is the one closest to its centre.

k-means, k = 6, on the standardised codes

V.

Laying

t-SNE unfolds those sixty-four dimensions onto a map, so shells that sit near each other in the code also sit near each other here. Depth becomes height: up to 18 mm for each shell.

Free: an archipelago at the original t-SNE positions. Where shells overlap, their heights add.

VI.

Finishing

Contour

Read as a chart of islands: a line for every 2 mm of rise, and a heavier one every 10 mm.

Drag to turn the work · scroll up to walk back along the shore

Beachcombing

Field Notes

This is a small piece about digital craft. An AI search stands in for walking the shore. An autoencoder studies the catch, a projection lays it out, and a few material studies finish the surface. Each step stands for a hand-skill, but it is done by computing.

Collection
300 photographs (60 kinds × 5 specimens) from 134 Shell Species Images, CC0. Species names follow the source data.
Depth
Depth Anything V2 Small on 320 × 320 crops. This is relative depth from one photograph, not a scan. Backgrounds were cleaned with masks before building the relief.
Autoencoder
CNN, 128 × 128 input, 64-dimensional code, 1.33 M parameters. 240 training and 60 held-out images, 80 epochs (about 4½ minutes on a CPU), best held-out model at epoch 71. Held-out error fell from 0.069 to 0.0038 (mean squared error), 94.5% lower. The rebuilt shells appear in chapter III only; the relief always uses the original depth.
Sorting
k-means (k = 6) on the standardised codes. The six groups are learned visual similarities, not taxonomy. In the finished work's free arrangement each group's largest island carries a chart name; the names are invented, the groups and their sizes are not.
Layout
t-SNE (perplexity 30, trustworthiness 0.96 at k = 10). Free: original positions within a 2100 mm field, 97 mm per shell, z = Σ 18·depth (mm), shown as islands in a sea; the Rhino model adds a 6 mm base. Grid: 18 × 18 tiles at 100 mm, each with a 6 mm base, assigned by linear sum assignment; 300 tiles and 24 empty cells. On screen, relief is drawn at 2× for legibility.
Finishes
Six real-time shaders on the same height field: contours, line drawing, plaster cast, fossil, nacre and ink rubbing. They are material studies, not simulations of fabrication.

Yidong He · Quantitative Aesthetics, Assignment 4 · 2026
Built with three.js (MIT) · Set in Gap Sans by Interstices (SIL OFL 1.1)

Beachcombing

the tide is coming in…