Hyperface: a naturalistic fMRI dataset for investigating human face processing

Abstract

Faces convey information that guides social behavior, yet neuroimaging studies investigating human face processing typically use static images with small sets of identities under artificial conditions. Controlled designs limit our ability to characterize human face processing under naturalistic conditions or test whether computational models generalize beyond the laboratory. To address this gap, here we release hyperface, a naturalistic face viewing fMRI dataset designed to investigate human face processing in response to faces portrayed in videos mimicking more ecologically valid conditions. Twenty-one participants watched 707 unique face video clips that vary systematically in identity, gender, age, ethnicity, expression, and head orientation. Each clip was rated by independent observers, and pairwise similarity judgments were collected through a behavioral arrangement task. Technical validation confirms high data quality with low motion, high tSNR, and high inter-subject correlation in visual and face-processing regions. The hyperface dataset is part of a comprehensive experimental framework to investigate human face processing: all 21 participants also watched "The Grand Budapest Hotel," performed a dynamic face localizer task, and 10 participants completed an additional face identity task with personally familiar and visually familiarized faces. These datasets are publicly available and enable within-subject comparisons across paradigms. Together they provide a unique resource for characterizing human face processing under naturalistic conditions and for benchmarking computational models against human brain responses.