PixPlot is a web-based visualization tool developed by Yale that leverages machine learning to organize and display large image collections. It enables users to enter and explore vast archives of digitized works as dynamic visual constellations.
For demonstration purposes, I have created two versions of PixPlot visualizations:
•a 2D constellation of approximately 23,000 post-1945 images
•a 3D constellation of the same collection
PixPlot groups images according to formal and aesthetic similarities, using a convolutional neural network to analyze visual features. The experience is akin to browsing a library shelf, and discovering adjacent works that share visual affinities. For example, one might find Color Field paintings naturally clustered together. This represents a kind of “extreme formalism,” offering the potential to uncover lesser-known artists who share stylistic qualities with canonical figures.
Within the visualization interface, a panel on the left lists a series of automatically generated “hotspots”—clusters of related images identified by HDBSCAN (Hierarchical Density-Based Spatial Clustering of Applications with Noise). These hotspots highlight areas of visual similarity, though they are not exhaustive; many additional clusters exist beyond those detected. The hotspot labels are provisional and may be refined over time. Users are encouraged to navigate freely and discover unexpected connections.
The overall constellation is computed using UMAP (Uniform Manifold Approximation and Projection for Dimension Reduction), which determines the spatial layout of images. UMAP parameters such as dimensionality and density can be adjusted to influence the structure and feel of the visualization.