jaddai/openart-mythic-creatures

OpenArt — Mythic Creatures

openart-mythic-creatures is the mythic creatures subject collection of the OpenArt family of open, public-domain art datasets: 12,933 works (4,053 paintings/illustrations · 8,828 photographed objects · 52 unclassified), each paired with a structured VLM caption plus medium, attribution and inscription metadata.

Mythological and fantastical beings — dragons, griffins, sphinxes, deities and hybrid creatures — rendered across media. Unusually metalwork-forward: much of this imagery survives on arms, armor, seals and decorative metal as well as in prints and sculpture.

Dataset Description

OpenArt draws from the open-access APIs of major museums and libraries (the Metropolitan Museum of Art, Rijksmuseum, Cleveland Museum of Art, Art Institute of Chicago, Smithsonian, Europeana, Wikimedia and others). Unlike the painting-only OpenBrush corpus, each OpenArt subject collection is mixed-medium: it contains both 2-D art (paintings, prints, drawings) and photographs of three-dimensional objects (sculpture, ceramics, metalwork, textiles). A brand column splits every row so you can take exactly the slice you want.

Key features

• 12,933 public-domain images, all passing a rights_status: safe gate
• Structured v2 captions — 9 semantic sections per image (same schema as OpenBrush)
• brand split — openbrush (4,053, 2-D art) vs openartifacts (8,828, photographed objects)
• Medium-tagged and attribution-aware: 2,016 rows carry a named artist; 1,453 of those are signature-verified against the pixels
• CC0-1.0 — public-domain dedication, no rights reserved (every row is museum CC0 / Smithsonian Open Access)
• Full provenance per row: source, landing_page, sha256

The OpenArt brand model

OpenArt is an umbrella with two medium-based brands plus subject collections:

• openbrush — 2-D art (paintings, prints, drawings). The standalone OpenBrush datasets are the painting-pure expression of this brand.
• openartifacts — photographs of real objects (sculpture, ceramics, metalwork, textiles).
• openart-<subject> — like this one: a subject collection that mixes both brands. Use the brand column to separate them.

Dataset structure

Field	Type	Description
image	Image	The artwork/object image (struct: bytes + path)
width, height	int	Pixel dimensions
brand	string	openbrush (2-D) / openartifacts (object) / ambiguous
theme	string	Subject collection (mythic-creatures)
tags	list[string]	Normalized descriptive tags
subject, action, setting, mood, style_description, lighting, color, composition	string	The 9 caption sections
caption_full	string	Full structured caption
medium	string	Medium/technique (e.g. oil painting, sculpture, metalwork)
artist	string	Artist/maker, or unknown
signatures	string	Signature text read from the image, if any
signature_verified	bool	Trust the artist only when this is true (a legible signature matched the name)
attribution_confidence	string	Model's attribution confidence
inscription_text	string	Transcribed inscription, if present
inscription_legible, inscription_verified	bool	Whether the inscription was legibly read / re-confirmed
title	string	Catalog title where available
rights, rights_status, license_class	string	Per-row license (authoritative)
landing_page, source, sha256	string	Provenance + integrity

Caption structure

Each image has a v2 structured caption (identical schema to OpenBrush). **STYLE:** maps to the style_description column; the other eight map to like-named columns:

**TAGS:** bronze, animal, lion, relief, patina, profile_view...

**SUBJECT:** A crouching lion modelled in bronze, head turned to the left...

**ACTION:** The animal is poised mid-stride, one forepaw lifted...

**SETTING:** Isolated object study against a neutral museum ground...

**MOOD:** Watchful, taut, restrained power...

**STYLE:** Cast bronze with chased detail and a dark green patina...

**LIGHTING:** Raking light from the upper left picks out the mane...

**COLOR:** Warm metallic browns and greens over a grey field...

**COMPOSITION:** Centered profile, the lifted paw breaking the silhouette...

Captioning methodology

Captions were generated with a two-model vision pipeline (distinct from OpenBrush's Qwen3-VL):

1. Caption + routing — Gemma 4 31B (google/gemma-4-31b-it): the full 9-section caption, identity-preserving subject description, inscription reading, and off-target routing (non-art / contaminated images were flagged and removed).
2. Verification — Gemini 3 Flash (google/gemini-3-flash-preview): a targeted second pass that re-grounded medium, attribution (artist / signature) and inscription claims against the actual pixels. signature_verified / inscription_verified come from this pass.

Tags were then normalized (lowercase, underscore, intensity/suffix stripping, *_lighting → *_light) to match the OpenBrush tag conventions.

Provenance & licensing

License: CC0 1.0 — public-domain dedication, no rights reserved (citation appreciated, not required).

Every image is public domain / CC0, pulled from museum & library open-access programs and verified at ingest (rights_status: safe):

• cc0 (the Met, Rijksmuseum, Cleveland, Art Institute of Chicago) — 11,771
• Smithsonian Open Access — itself a CC0 dedication — 1,162

Aggregator-sourced rows (Europeana, Wikimedia), whose upstream rights are less certain, were removed so the CC0 dedication holds for every row. Faithful photographic reproductions of 2-D public-domain works carry no new copyright (Bridgeman Art Library v. Corel Corp.); object photographs are released under the holding institution's CC0 / open-access program. Each row keeps source, landing_page and sha256 for traceability.

Statistics

Metric	Value
Total images	12,933
openbrush (2-D)	4,053
openartifacts (object)	8,828
ambiguous	52
Named artist/maker	2,016
Signature-verified	1,453
Inscription-verified	2,969

Top media: metalwork (4,383), sculpture (2,020), engraving (901), etching (862), drawing (747), ceramic (568), woodcut (383), photograph (276)

Top tags: metalwork, metal, weapon, steel, armor, antique, artifact, engraving, monochrome, sculpture, iron, japanese_art, ornate, ornament, gold

Top sources: met openaccess (9,302), rijksmuseum oai (1,364), smithsonian art (1,162), cleveland (752), met (353)

Usage

from datasets import load_dataset

ds = load_dataset("jaddai/openart-mythic-creatures", split="train")
print(len(ds))  # 12,933

# Just the 2-D paintings/prints (the openbrush brand)
paintings = ds.filter(lambda r: r["brand"] == "openbrush")

# Just photographed objects (the openartifacts brand)
objects = ds.filter(lambda r: r["brand"] == "openartifacts")

# Only trust attribution where the signature was verified
attributed = ds.filter(lambda r: r["signature_verified"])

# Filter by medium
sculpture = ds.filter(lambda r: r["medium"] == "sculpture")

Use cases

• Image-generation training — fine-tune diffusion models on a specific medium or subject
• VLM fine-tuning — object/art description grounded in museum provenance
• Classification — medium, brand, or subject recognition
• Cross-medium study — compare how a subject is treated in paint vs. in three dimensions

Limitations

• Captions are AI-generated best-effort metadata, not curatorial ground truth.
• Attribution is authoritative only where signature_verified = true. Other named artists are scholarly/unverified catalog attributions.
• Inscriptions: trust inscription_verified = true; some claimed inscriptions could not be re-confirmed legibly, and catalog labels/accession numbers can appear as inscription text.
• Source-institution and Western-canon collection biases are present.

Citation

@dataset{openart_mythic_creatures,
  title={OpenArt: Mythic Creatures},
  author={jaddai},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/jaddai/openart-mythic-creatures},
  note={Subject collection of the OpenArt family}
}

Related datasets

Part of the OpenArt family of open, public-domain art datasets — the whole family is CC0-1.0 (public-domain dedication, no rights reserved; citation appreciated, not required). License shown beside each name for clarity.

OpenBrush — 2-D fine art (paintings, prints, drawings), WikiArt-sourced (curated public domain), CC0:

• openbrush — flagship — 75,313 public-domain paintings · CC0
• openbrush-anonymous-masters — unattributed old masters · CC0
• openbrush-impressionism — Impressionist movement · CC0
• openbrush-landscapes — landscape genre · CC0
• openbrush-portraits — portrait genre · CC0
• openbrush-religious-art — religious / biblical · CC0
• openbrush-renaissance — Renaissance period · CC0
• openbrush-baroque — Baroque period · CC0
• openbrush-impressionist-landscapes — Impressionist landscapes (cross-cut) · CC0
• openbrush-van-gogh — Vincent van Gogh · CC0
• openbrush-ukiyo-e — Japanese woodblock prints · CC0
• openbrush-monet — Claude Monet · CC0
• openbrush-rembrandt — Rembrandt · CC0
• openbrush-renoir — Pierre-Auguste Renoir · CC0

OpenArt — mixed-medium subject collections (paintings + photographed objects), museum/library-sourced, CC0:

• openart-animals — animals & the natural world · CC0
• openart-items-artifacts — human-made objects & decorative arts · CC0
• openart-painterly-foundations — painting-forward fine-art core · CC0
• openart-portraits-classical — the human figure & portraiture · CC0

Feedback & collaboration

I'm always open to feedback and to suggestions for other datasets worth publishing — open a discussion on this repo or reach out on the Hub. I'm also open to collaborations with others in the community.

Acknowledgments

• Images sourced from the open-access programs of the Met, Rijksmuseum, Cleveland Museum of Art, Art Institute of Chicago, Smithsonian, Europeana, Wikimedia and others.
• Captions generated by Gemma 4 31B; attribution & inscription verification by Gemini 3 Flash.