DisParQ
Explore the concepts
Choose a dataset to browse its training images and learned concepts. Open an image to see its concept map, select a region to inspect its concept, or filter the image gallery by class and concept number. The five dataset cards below are the starting points.
DisParQ describes an image with discrete part concepts, learned from a frozen self-supervised backbone without class labels or language. Each image patch is assigned to exactly one concept from a learnable dictionary, only a small set of concepts is active in each image, and quantized attributes capture how a concept varies from image to image.
