FoodVis: An AI-Generated Food Image Resource for Cognitive and Neuroimaging Research

FoodVis is an AI-generated food image resource developed for cognitive, behavioral, and neuroimaging research. The resource was designed to provide a broad range of food stimuli within a common image-presentation framework, reducing variability introduced by heterogeneous backgrounds, viewpoints, lighting conditions, and framing that is often present in existing food-image collections. The full dataset contains 212 food images spanning 11–898 kcal/100 g across five calorie-density tiers. All stimuli are provided as 1024 × 1024 RGB images with a uniform medium-grey background (RGB [191, 191, 191]) and are accompanied by refined binary foreground masks and alpha mattes. Calorie-density values were assigned using established food-composition databases, primarily USDA FoodData Central and CIQUAL, using the closest available match to the depicted food and preparation state. The resource includes a detailed characterization of low-level visual properties. Forty image-derived features quantify foreground geometry, luminance, colour, contrast, edge structure, texture, spatial-frequency content, and orientation structure. These measurements were used to assess visual covariance with calorie density and calorie tier. In the full set, simple foreground geometry provided little predictive information about calorie density, whereas some appearance properties retained expected relationships with food composition. For studies requiring stronger control over measured low-level visual variation, FoodVis also includes an optional visually balanced subset of 168 images. This subset was selected by directly balancing individual visual features and broader principal-component representations while preserving the overall calorie-tier composition and almost the entire calorie-density range. The balanced subset retains 99.89% of the calorie-density range of the full dataset while substantially reducing the strongest measured associations between low-level visual properties and calorie content. The release includes the stimulus images, foreground masks, alpha mattes, calorie-density and stimulus metadata, low-level visual-feature measurements, quality-control outputs, balanced-subset identifiers, analysis code, and documentation. Semantic characteristics such as processing level, taste category, preparation state, and presentation type are also documented. FoodVis is intended as a flexible stimulus resource rather than as a set of visually identical foods across nutritional conditions. Natural relationships between food appearance, identity, processing, and calorie density are retained and quantified rather than artificially removed. Further development will extend the nutritional metadata to include macronutrient information and will add behavioral validation of recognizability, name agreement, familiarity, perceived calorie content, healthiness, palatability, realism, and visible image-generation artifacts.