Aside from the analysis of hate language in partisan news, relatively little is known about how media introduce bias and facilitate affective polarization. Although language and text should be secondary to images from a behavioral perspective, until recently scholarly attention was almost exclusively focused on textual analysis. In fact, when we scroll through social media newsfeeds or read a newspaper, images grab our attention first and influence our opinions well before we get to the text itself. Hence, images are meant to prime our evaluations, but do they? How does nontextual information contribute to media bias? What is the mechanism through which nontextual frames promote or mitigate affective political divide? In order to answer these questions, we select a shortlist of political events and track the manner in which ideologically diverse media reported such events on social media. We unsupervised computer vision algorithms to analyze images accompanying these news stories. After, we analyze news short descriptions to identify differences in the sentiment between liberal and conservative media outlets towards the same events. Finally, we conduct a series of online survey experiments in order to identify the exact mechanisms through which images introduce bias and increase the affective divide. Since visual information is more credible than text, we hypothesize that media outlets will attach more sentiment to images accompanying news stories than to the stories themselves. Our starting point for empirical evidence is the US, but we expect that the revealed behavioral mechanisms will be easily generalized to other political environments.