This work advances understandings of image-sharing behavior on Twitter, across race, gender, age, and political engagement. We infer account-level demographic measures via profile pictures of US Twitter accounts and characterize 20 types of images. Several of these types predict one's demographics using account-level logistic regression models. Around half of the learned clusters (e.g., infographics, natural scenery, sports) are predictive of the user's age, race, or gender, while several other clusters appear to be popular among politically engaged accounts (e.g., images of groups and images of single individuals, which often contain politicians). Our findings suggest it is possible to characterize certain audiences via different types of visual imagery, which has implications for information quality, online engagement, and communications.