Veronica bossio botero¹, Nikolaus Kriegeskorte; ¹Columbia University, Zuckerman Institute, ²Columbia University, Zuckerman Institute

Humans are experts at face recognition, yet the underlying perceptual space that drives subjective identity judgements remains to be characterized quantitatively. Deep image-generative models can synthesize photorealistic face images from latent embeddings in representational spaces. Such models offer a powerful new framework for studying face perception, operating as flexible stimulus generators and computationally explicit candidate models for the perceptual space itself. Here, we leverage Arc2Face—a diffusion-based generative model—to characterize the geometry of subjective face space and individual variability in face discrimination. We used Arc2Face to generate morph-lines radiating from 240 reference identities, each representing a trajectory of identity change in latent space. In a behavioral task, 40 participants judged whether pairs of faces along these trajectories depicted the same identity. We modeled these judgments using a hierarchy of probabilistic models to test hypotheses about the properties of human perceptual face space: its uniformity (is sensitivity invariant to location in face space?), isotropicity (is sensitivity invariant to direction of change?), and interindividual variability (do subjects exhibit idiosyncratic sensitivity profiles?). We established that a model based solely on the angular distance between embeddings is highly predictive, capturing 76% of the explainable variance in human judgments. Model comparison revealed that perceptual space is non-uniform and anisotropic, with subjective identity thresholds varying significantly depending on reference face and direction of change. Importantly, interindividual variability extended beyond differences in global sensitivity (larger or smaller thresholds for perceiving a match). An interaction model allowing unique observer-stimulus interactions significantly outperformed one that assumes observers differ only in their response criterion, demonstrating observers have stable, idiosyncratic sensitivity patterns to specific facial features unexplainable by global bias. These findings quantitatively characterize the geometry of perceptual face space. We introduce a validated tool for predicting human identity judgments and provide evidence that face discrimination relies on idiosyncratic, observer-specific facial cues.