Latent Kiln

Latent Kiln asks: how does a machine perceive cultural form? Starting from 270 ceramic vases drawn from the Victoria & Albert Museum archive and spanning six geographic regions — East Asia, Africa, the Americas, Asia, Europe, and the Middle East — the project trains five parallel computer vision models, each encoding a distinct perceptual dimension of ceramic morphology.

Rather than treating classification as a sorting task, the project frames it as an act of curation. The resulting "machinic curator" places every artifact within a six-dimensional morphological universe, translating visual features into spatial coordinates that map cultural distance across time and geography.

The trained models power a dual-interface browser exhibition. On the left, visitors draw any vase silhouette freehand. On the right, the system operates in two modes: Universe mode, which surfaces the nearest morphological relatives and predicts the top three matching regions; and Remix mode, which enters an Assembly Station where the visitor's drawn vase is deconstructed and recombined with parts from across the archive — generating novel hybrid vessels in real time.

The archive focuses exclusively on vases as a typology — a form present across all six regions but differentiated by material culture, production method, and ornamental tradition. Each vessel is digitised as a high-resolution photograph against a neutral background, sourced from the V&A's open-access API.

Five parallel classifiers are trained on five distinct visual representations of the same vessel: the original photograph (color + texture), a grayscale rendering (light + shade), a thick-edge silhouette (2D linearity), a depth map (3D volume), and a binary mask (2D mass). Each model learns a different aspect of what makes a ceramic "look" like it belongs to a particular region.

Photo-based models achieve the highest test accuracy (73.3%), but the interface intentionally begins with user-drawn sketches — where ambiguity and abstraction are part of the experience. Edge and depth models, though lower in raw accuracy, are better at capturing form rather than surface.

Each vessel's classification probabilities across six regions act as directional forces on a radial map. The six axes represent the six cultural regions; the model's confidence in each region pulls the artifact toward that pole. The resultant vector — the sum of all six weighted forces — determines a precise X,Y coordinate in a two-dimensional morphological universe.

This spatial translation logic converts subjective cultural affinity into measurable geometry. A vessel deeply ambiguous between Africa and the Americas sits between those two poles; a canonical East Asian piece clusters tightly at the East Asian axis.

Beyond classification, the project physically deconstructs each vessel into three anatomical components — neck, body, and base — using a combination of AI-prompt segmentation and manual curation. Sixty vessels were fully processed, creating a combinatorial space of over 200,000 possible recombinations: a mathematical design space rather than a fixed archive.

A second Python script aligns components by interface width and relative proportion, producing volumetrically coherent hybrid vessels that merge aesthetic traditions across cultures and centuries.