A topographic video model predicts the spatio-functional organization of the human visual system

Abstract The human brain organizes visual processing into spatially structured functional clusters. Recent topographic neural networks model category-selective clusters but are limited to static images and ventral cortex. In this work, we model the topographic organization of dynamic visual representations across high-level visual cortex. We introduce a general method that spatially arranges and transforms units from a neural network onto a simulated cortical sheet. Applied to a pretrained video foundation model, the method produces spatially structured response correlations while preserving task performance and brain alignment. It captures ventral and dorsal clusters, including face-, body-, and motion-selective patches, and reflects system-level selectivity distributions for animacy, object size, and static versus dynamic content. Functional selectivity varies smoothly across both model and human tissue, suggesting predominantly general-purpose cortical processing with continuous transitions into local functional clusters. Taken together, we provide a computational framework for understanding the spatial and functional organization of high-level visual cortex.