Semantic Art Search 🎨🔍
Semantic Art Search is an open-source search engine that uses advanced machine learning to explore digitized art collections. With simple natural-language queries, you can uncover artworks by their visual content, themes, emotions, or styles across multiple museums.
Unlike traditional keyword searches, Semantic Art Search understands meaning directly from images and lets you discover artworks even when the metadata doesn’t contain exact keywords.
Currently, the system includes selected artworks from these renowned institutions committed to open data:
- Statens Museum for Kunst (The Danish National Gallery)
- The Cleveland Museum of Art
- Rijksmuseum Amsterdam
- The Metropolitan Museum of Art
- Art Institute of Chicago
🔎 How It Works
Just use the search bar — it's designed for natural language input. Try themes, moods, styles, or descriptions, for example:
- "Mourning"
- "Battle at sea"
- "Impressionism"
- "Bible scene"
You can also paste an inventory number (e.g., "KMS1") into the search bar to find visually or thematically related works. This is especially useful when clicking “Find similar” on a specific artwork.
When you search:
- The entire collection is ranked by relevance (best matches are shown first)
- You can filter by work type (e.g., painting, print, drawing)
- You can search across all museums or limit to one or more
Each artwork links to its page at the source museum, so you can dive deeper and explore the full context.
🔧 Search Mode
You can choose how artworks are matched to your query:
- Auto (default): Combines visual and title search for the best overall results
- Visual: Matches by visual appearance — finds artworks that look like your query
- By title: Matches by title — finds artworks that are named like your query
Jina supports queries in multiple languages, so you can search in your native language.
🤔 How Is This Search Different?
Traditional art search engines often rely on exact keyword matches in stored metadata such as titles, descriptions, and tags. If you search for "cat", such engines will return artworks that explicitly mention "cat" in their metadata. While this kind of search has its merits, it also has clear limitations. For example, a painting might depict a cat without mentioning it in the title or description, or the title might contain the word "cat" in another language. In these cases, you would get no results, even though the painting is relevant to your search.
Semantic Art Search helps you discover works that match the intent of your query. In Visual mode, it finds meaning directly in the image — if there's a cat somewhere in the painting, there's a good chance that this work will be found, even if the word "cat" does not appear in the metadata. In By title mode, it matches semantically against artwork titles — understanding that "feline" and "cat" are related concepts, not just exact words. The default Auto mode combines both signals.
Being natural language based, the semantic search engine also understands abstract concepts and phrases like "war", "ancient Rome", or "shipwreck in a storm", allowing you to search in a very flexible and intuitive way.
📝 Note: Filters such as work type and museum still use metadata — but search relevance is driven by visual and semantic similarity.
🧠 Behind The Scenes
The system uses Jina CLIP v2 (a multimodal neural network) to embed both images and search queries into a shared vector space. It then performs vector-based nearest neighbor search to rank artworks by semantic or visual relevance to the query. Users can search either by visual appearance (image embeddings) or by title (text embeddings).
🎨 Artwork Coverage
The system includes artworks that:
- Are in the public domain
- Have a photograph available in the museum dataset
Not all types of artworks are yet included - focus for now is on visual art like paintings, prints, and drawings. Coverage will grow over time.
🙌 Acknowledgments
This project is made possible by the excellent open data initiatives of:
📬 Get in Touch
Have questions, feedback, or ideas? Want to contribute?
Reach out via:
