A colophon, of sorts: what the site is set in, what reads the photographs, and whose openly-licensed work it stands on.
Type
Four typefaces, all under the SIL Open Font License, self-hosted as subset woff2 files:
- Fraunces — the serif: plate titles, the search field, the big count, and the scrolling vocabularies. By Undercase Type.
- Archivo — the masthead. By Omnibus-Type.
- Tourney — the [Open] and [Randomise selection] marks. By Tyler Finck, Etcetera Type Company.
- Roboto — small labels, wall numbers and metadata. By Christian Robertson.
Reading the photographs
Beyond Te Papa’s own cataloguing, the site adds a machine-read layer so 54,000 photographs can be browsed by subject, feeling and composition. It was built entirely with open-weight models running locally on one laptop — no image ever left the machine for processing:
- SigLIP 2 (Google, Apache 2.0) embeds every photograph and scores it against a thousand-term vocabulary — that’s the subject tags, and the same embedding space drives the feelings and composition rows. Tags ship in three tiers, labelled wherever they appear: human-calibrated (thresholds set by eye over a review sheet), audited (independent models agree and a blind precision check passed), and broad match (an uncalibrated floor, said so on the page).
- Qwen3-VL 8B (Apache 2.0), run through Ollama (MIT), wrote a structured caption for every photograph over about a week of background compute. Captions corroborate the tags, and where the model could read words photographed inside an image — shopfronts, signwriting, negative-plate imprints — the transcript is kept only when the catalogue record independently backs it up. Search matches those words too.
- E5 (MIT) provides the small text embeddings that tie captions to tag labels.
Deliberately not automated: subjects in te reo Māori / te ao Māori, and any term that classifies people by ethnicity, are never machine-published — they wait for human review.
A full retrospective of the pipeline — what ran locally, and what the whole thing cost in electricity (about one dryer load) — is on GitHub.
The feelings list
The “Or a feeling…” row browses 154 emotions drawn from Tiffany Watt Smith’s The Book of Human Emotions (Profile Books / Wellcome Collection, 2015) — from awumbuk to torschlusspanik. The composition & technique vocabulary was written for this site.
Other open source
- OpenSeadragon (BSD-3-Clause) — the deep-zoom viewer behind every photograph.
- Transformers.js (Apache 2.0) and ONNX Runtime (MIT) — run the embedding models in the build pipeline.
- The site itself is hand-written HTML, CSS and JavaScript — no framework — and its source is public.