Artographer: a Curatorial Interface for Art Space Exploration

TL;DR

Artographer presents a multimodal, embedding-based zoomable map of historical artworks to study how users explore art space and surface relationships. The authors build a curated dataset, develop a UMAP-based mapping pipeline with region-level clustering, and integrate AI-generated imagery as a navigational tool, then evaluate user behavior through a within-subjects study with 20 participants. They identify four exploration behaviors and four design values—Visibility, Agency, Friction, Serendipity—arguing for curatorial interfaces that balance exploration, control, and responsible AI curation. The work contributes open data, a novel spatial interface, and design guidance for pluralistic, user-driven art cartography that centers human agency in AI-assisted curation.

Abstract

Relating a piece to previously established works is crucial in creating and engaging with art, but AI interfaces tend to obscure such relationships, rather than helping users explore them. Embedding models present new opportunities to support discovering and relating artwork through spatial interaction. We built Artographer, an art exploration system featuring a zoomable 2-D map, constructed from the similarity-clustered embeddings of 15,000+ historical artworks. Using Artographer as a probe to investigate spatial artwork exploration, we analyzed how 20 participants (including 9 art history scholars) traversed the map, during a goal-driven task and when freely exploring. We observe divergent and convergent exploration behaviors (Jumping, Wandering, Fixation, Revisiting) and identify values enacted by spatial art-finding (Visibility, Agency, Serendipity, Friction.) We situate spatial maps within a space of Curatorial Interfaces, systems that select and present artworks, and discuss centering pluralism and agency in the design of more responsible AI systems for art curation.

Figures (11)

• Figure 1: Each artwork is represented by a multimodal 3456d feature vector that combines visual embeddings generated with ResNet50, text embeddings from MiniLM, and image+text embeddings from CLIP.
• Figure 2: Artographer (left) is a spatial map interface for historical artwork exploration. Our Baseline system (right) was a Query-Based Search Interface that mimics traditional interfaces for searching an artwork database, with support for browsing, text-search (including approximate matches), and search-by-image (supported by an image generator.) Both interfaces support exploration of the same database of 15,968 artworks.
• Figure 3: Comparing number of artworks collected in each system (Artographer vs. Baseline) for each of the two targeted task conditions ("Peaceful" and "Conflict"). The difference in collection size between systems is not statistically significant.
• Figure 4: While participants collected a similar number of artworks (around 8-13) regardless of system, they interacted with significantly more artworks (around 80 vs. around 21) when using Artographer to complete the task (p < 0.001 in both conditions).
• Figure 5: Comparing the number of images each participant generated with the number of unique artworks they interacted with. Participants who used wandering, rather than AI-generated jumping, explored more gradually, interacting with more images along the way.