Yijin Wang¹ (), Chenxi Liao¹, Bei Xiao¹; ¹American University

From moment to moment, we encounter countless materials. Materials exhibit a wide range of appearances varying within and across categories. The same material can take many forms. What makes an image of material memorable? Are some materials easier to memorize than others? Memorability has been conceptualized as a stimulus-driven, stable statistical property that is consistent across observers. Understanding the visual memorability of materials is a crucial step toward revealing how the brain efficiently processes and represents visual information about them. Here, we investigate the visual memorability of materials using a traditional memorability recognition task on the STUFF dataset, which comprises 3,514 images from 200 material categories. Individual participants (N = 11) viewed a continuous stream of images and pressed a key whenever they recognized an image they had encountered; over the course of the experiment, each participant viewed the entire dataset. We computed material memorability for each image as the difference between the hit rate (proportion of participants who made a correct recognition) and the overall false alarm rate. Despite the challenging dataset, participants showed moderate accuracy (target: 56.25%; overall: 87.76%; false alarms: 8.09%; mean memorability: 48.17%) with consistent patterns—Spearman correlations between randomly split groups were high, confirming stable stimulus-driven memorability (r=0.33). Results from GLMM showed that material category strongly predicted memorability (p<0.001). Snow, tooth, and pearl were highly memorable; garnet and chlorine were least memorable. Conversely, low- to mid-level visual features extracted from pretrained neural networks (LPIPS, DreamSim) correlated only weakly with memorability. Our results suggest that memorability, through reflecting the statistical regularities of the environment, can serve as an intrinsic dimension in constructing the representation of materials. In addition to low- to mid-level visual features, an efficient store of category-driven information is fundamental for future recognition.