museum_art_1

Artworks from open-access museum collections (CC0 or public domain as designated by each institution), public-domain pulp magazine covers, posters, illustrations and WikiArt works, exported as aspect-ratio buckets and captioned twice with Gemma 4 31B.

Captions

Every image has two English captions, generated for this release with Gemma 4 31B from the image and a short catalogue line (artist, general period, medium):

• natural — a dense description (subject, composition, medium, technique, style, colour, text visible in the image quoted). Stored as the sample's .txt and as captions.natural in its .json.
• layout — a JSON layout (captions.layout, format: layout_json): {"summary": {"meta": [...], "canvas": "single|multi_panel|collage", "caption": ...}, "frame": {"caption", "tags", "palette"}, "items": [{"label": "character|object|text|panel", "caption", "text"?, "tags", "palette", "box_2d"}]} with box_2d = [y_min, x_min, y_max, x_max] normalized to 0–1000 over the stored image.

The manifest lists both in sample_captions. Captions do not name artwork titles or exact dates; text written in the image is quoted. Photographs that show reference material beside the work (colour charts, grey scales, rulers, labels) describe it in both captions.

Selection

Catalogue records were filtered to images of artworks, scored with SigLIP2-based aesthetic and junk probes, and checked on contact sheets. Text-dominated pages and calligraphy were removed. Near-duplicates (the same image under several records or in several sources) were removed after a geometric match check, keeping the largest copy.

The *_prints folders are a diversity sample of 10,000 prints and drawings from the eight museums' print and drawing collections: after the text-page, blank-sheet, size (at least 1024 px on the long side) and median-aesthetic filters, the works were clustered into 10,000 groups (spherical k-means on SigLIP2 embeddings) and the best-scored work of each group was kept; the folders hold the works left after near-duplicate removal. They are stored at the 1024 tier and below.

Sources

folder	source	rights	images
aic	Art Institute of Chicago	CC0	3,070
cma	The Cleveland Museum of Art	CC0	3,536
fng	Finnish National Gallery	CC0	4,583
getty	J. Paul Getty Museum	CC0	2,403
met	The Metropolitan Museum of Art	CC0	6,862
nga	National Gallery of Art, Washington	CC0	4,954
si	Smithsonian American Art Museum, Cooper Hewitt and National Portrait Gallery	CC0	6,392
smk	SMK – Statens Museum for Kunst	public domain	3,973
yale	Yale Center for British Art and Yale University Art Gallery	public domain	12,300
ia_pulp	Pulp magazine covers up to 1930, Internet Archive	public domain in the US	1,774
commons	Posters, prints and illustration plates, Wikimedia Commons	public domain in the US	5,934
wikiart	Paintings, drawings and prints, WikiArt	public domain (see Rights)	92,841
aic_prints	Prints and drawings, Art Institute of Chicago	CC0	2,090
cma_prints	Prints and drawings, The Cleveland Museum of Art	CC0	821
fng_prints	Prints and drawings, Finnish National Gallery	CC0	350
getty_prints	Prints and drawings, J. Paul Getty Museum	CC0	69
met_prints	Prints and drawings, The Metropolitan Museum of Art	CC0	1,464
si_prints	Prints and drawings, Smithsonian American Art Museum, Cooper Hewitt and National Portrait Gallery	CC0	1,391
smk_prints	Prints and drawings, SMK – Statens Museum for Kunst	public domain	1,362
yale_prints	Prints and drawings, Yale Center for British Art and Yale University Art Gallery	public domain	1,871
total			158,040

Each folder is a self-contained bucketed-shards export (<folder>/manifest.json, <folder>/buckets/<bucket_id>/shard-*.tar). Per-sample .json keeps the source record: source_metadata (title, artist, date, medium, classification, object_id, object_url, source, licence) and the image URL.

Rights

Images are CC0 or public domain as designated by the holding institution; pulp covers and Commons items are public domain in the US. WikiArt works are by artists who died in 1955 or earlier and are dated 1930 or earlier (undated works: the artist died by 1930), with death years from Wikidata; photographs of sculpture, installations and other 3D works are excluded. The per-sample licence and object_url fields record each museum's and pulp/Commons item's designation; WikiArt samples have no licence field.

Format

• Shards: uncompressed TAR, 3 files per sample — <key>.jpg, <key>.txt, <key>.json.
• Buckets: SDXL-style aspect buckets defined at a 1024×1024 base, exported at several base resolutions; bucket ids are r<base>_<proto> (e.g. r1024_p1216x832). Each work is stored once, at the highest tier its source resolution allows (never upsampled).
• Images: bicubic, antialiased cover-resize to the bucket, corner crop, JPEG quality 95 (source JPEGs that already match a bucket are stored as-is).

base resolution	images
2048	58,532
1024	42,116
768	33,442
512	23,950

Total size: 119.5 GB.

Loading

import json, tarfile
from pathlib import Path

for tar_path in sorted(Path("nga/buckets").rglob("shard-*.tar")):
    with tarfile.open(tar_path) as tf:
        for m in tf.getmembers():
            if m.name.endswith(".json"):
                key = m.name[:-5]
                meta = json.loads(tf.extractfile(m).read())
                natural = meta["captions"]["natural"]["text"]
                layout = json.loads(meta["captions"]["layout"]["text"])
                jpg = tf.extractfile(key + ".jpg").read()