Shikang Peng^1,2 (), Wilma A. Bainbridge¹; ¹University of Chicago, ²University of Toronto

Visual content on social media plays a pivotal role in global entertainment and information sharing, yet some images receive significantly more likes and comments than others, raising questions about perceptual features in the images driving such engagement. While emotional and moral content has been linked to viral posts, these factors vary across populations. We propose that a post's image memorability—its ability to be remembered—might predict its viral potential. Here we collected a fully naturalistic dataset of image posts from the well-known social media platform Reddit, across three separate timepoints (N=1,247 image posts). Leveraging memorability-trained ResMem, we obtained a predicted memorability score for each image. By correlating these with the virality metrics, we found that memorable images tend to attract more comments, replicated across the three timepoints. Even after accounting for the image category of the post using ImageNet pre-trained ResNet-152, we found that ResMem’s predictions still explained unique variance in virality. A semantic analysis was then conducted on comments to investigate the mechanisms of this effect. We found that memorable images tend to elicit more neutral-affect comments, demonstrating memorability has a different pathway to virality from emotion. Further, a visual consistency measure using Google Vision AI and word embeddings showed that memorable image posts were associated with more diverse, externally-associated comments. To explore which characteristics drive memorability and virality, we analyzed image features through ResMem's convolutional layers. We found semantic distinctiveness was key, indicating a shared underlying mechanism between memorability and its viral potential. In sum, this study provides a detailed investigation surrounding the topic of 1) whether memorability can predict social media virality; 2) how and why memorability affects human online behavior leading to virality; and 3) what underlying visual features contribute to the viral potential of a post.