hr-studioAI · 01
A computational design primer · Embeddings & Latent Space

Where every point is a design.

Generative AI rests on one quiet idea: turn each thing — a word, an image, a shape — into a point in a space of numbers, arranged so that nearby points mean similar things. Once meaning becomes geometry, you can measure it, move through it, average it, and walk a smooth path from one design to another. This chapter builds such a space by hand, so the next ones can learn it.

Scroll to begin
i.

A shape becomes a vector

To compute with form, first measure it. Walk around a shape's outline and record the radius at a hundred angles, and the shape turns into a long list of numbers — a vector, a single point in a high-dimensional space. Two shapes that look alike give nearly the same list.

Morph the shape and watch its vector breathe. The picture and the numbers are the same thing, seen two ways.

Outline → vector (its fingerprint)
ii.

Geometry becomes meaning

Take many shapes as vectors and project that high-dimensional cloud down to a flat map, keeping near things near. Similar shapes land together; the map organizes itself by family. We never labelled anything — the structure was hidden in the numbers, and the projection simply found it. That is an embedding.

Hover a shape to light up its nearest neighbours — its lookalikes in the cloud.

Embedding map · near = similar
iii.

Directions carry attributes

The space isn't just clusters — its directions mean things. One axis makes a shape rounder or spikier; another makes it even or lopsided. Move that way and you add the attribute, like the famous word-vector trick where king − man + woman lands on queen. Meaning has a compass.

Slide along each direction and feel the attribute change, smoothly and on its own axis.

Directions = features
iv.

Walking between designs

Pick two shapes and draw the straight line between their points. Every spot on that line is itself a valid shape — a believable blend of the two ends. Sliding across it morphs one design into the other through a continuum of in-betweens that were never drawn by anyone.

This is the generative move: not picking from a list, but travelling through a space of possibilities.

Interpolation · A → B
v.

The latent design space

Here is the whole space at once: a map whose every location is a design, with families of shapes laid out as landmarks. Drag the cursor and the large shape is generated live from wherever you are — no gallery, no presets, just one continuous field of possibility you steer through. This is exactly the object that autoencoders, diffusion models, and GANs spend their training learning to build; here you can simply wander it.

Drag anywhere · the design is generated from your position
A point for every design, distance for similarity, directions for attributes, straight lines for blends. Build that space and generation stops being a lottery and becomes navigation — which is the whole trick the rest of generative AI is built on.