Vistarium

Open-access landscape photography, filtered by location, time of day, and subject.

Site: https://joshkaspar.github.io/vistarium/

A curated selection of public domain and open-access landscape photographs, currently sourced from the US National Park Service. Images are pulled via API, filtered by predicted aesthetic quality, then classified with structured metadata (time of day, subject, people, etc.) by a local vision model. Additional sources (Library of Congress, Smithsonian, Met, Art Institute of Chicago, NYPL) are planned.

Data

Each record has two parts, populated separately, with one exception noted below.

Catalog metadata — pulled directly from the source API, unedited:

{
  "id": "string",
  "source": "nps",
  "source_url": "string",
  "image_url": "string",
  "title": "string",
  "photographer": "string | null",
  "date": "string | null",
  "park": "string",
  "license": "string"
}

Model judgment — structured, grammar-constrained fields from a local vision model (see Cohobate), which only ever receives pixels (no filename, caption, or EXIF data is included in its prompt):

{
  "is_photograph": true,
  "time_of_day": "morning | afternoon | evening | night",
  "time_of_day_evidence": "caption | exif_timestamp | visual_inference",
  "license_confidence": "confirmed | flagged_for_review",
  "license_evidence": "string",
  "primary_subject": "landscape | wildlife | structure | vehicle | human_activity | document | detail",
  "people_present": true,
  "people_prominence": "none | background | midground | foreground_focal",
  "crop_anchor": "center | top | bottom | left | right",
  "frame_type": "full_bleed | matted | multi_panel | stereograph",
  "color_mode": "color | monochrome",
  "dominant_color": "red | orange | yellow | green | blue | purple | white | gray | black",
  "tags": ["string"]
}

Exception: time_of_day and time_of_day_evidence The model's guess is only used in the absence of EXIF or caption data. time_of_day_evidence records which source was actually used.

Plus a thumbnail_crop_16x9 crop box, computed deterministically, and an optional aesthetic_score / aesthetic_method pair from a separate scoring stage. Full field-by-field definitions live in Cohobate's schema.json — the shape is documented there, alongside the pipeline that produces it, since this repo only holds the resulting data, not the schema's source of truth.

Curation overrides

hidden_ids.json and confirmed_ids.json are small, hand-curated id lists that this repo owns and Cohobate's build_site.py reads (via VISTARIUM_DATA_REPO_PATH — see Cohobate's SETUP.md) when it rebuilds docs/:

• hidden_ids.json — ids to exclude from the site regardless of what the catalog says (manual dedup-review hides, or license/content concerns). Denylist.
• confirmed_ids.json — ids the model flagged license_confidence: flagged_for_review that a human has since reviewed and confirmed as safe to publish. Allowlist, opposite polarity from hidden_ids.json.

Both are real curation decisions, not regeneratable scrape output — they're versioned here because they're this repo's data, even though Cohobate is what reads and acts on them.

Sorting

The site defaults to sorting by a predicted aesthetic score ("Aesthetic Rating (AI)"), with Date take and Date added as alternatives. The raw score itself isn't shown per photo — only used to order results.

License & Rights

• Vistarium's catalog (each curated, scored, classified record as a whole, not split field-by-field) is CC BY 4.0 — free to reuse, including commercially, with attribution.
• The images themselves are not covered by either license. Each image's rights status is recorded per-item in its own metadata (license, license_confidence, license_evidence), reflecting what the source institution states — not independently verified or guaranteed by this project.

My goal is to make this data useful, while making it easy to reuse. However, if you plan to use these images in your projects - especially if they are commercial projects - verify the rights of the image from the sources. Copyright-free status does not necessarily resolve every right that may apply — notably, a depicted person's right of privacy or publicity is separate from copyright and is not waived by an archive's public-domain designation. Users are responsible for verifying licenses on the original source and for their own lawful use of any downloaded material, including not infringing on the rights of third parties.

See TERMS_OF_USE.md for the full rights statement.

Repo layout

vistarium/
├── docs/                    # the static site (GitHub Pages): index.html, app.js, style.css, data.json, thumbs/
├── hidden_ids.json          # curation override: denylist
├── confirmed_ids.json       # curation override: allowlist (flagged-for-review overrides)
├── STATUS.md                # auto-generated curation progress
├── TERMS_OF_USE.md
└── LICENSE-DATA             # CC BY 4.0, covers the catalog

The pipeline

Cohobate is the pipeline that scrapes source archives, curates and deduplicates candidates, classifies images with a local vision model, and builds this repo's docs/ folder. If you want to know how this project was built, see the documentation there.