Consumption of online news generally comes with visual imagery. However, experimental evidence around news credibility perception almost universally takes images as given, leaving it unclear what it is about images that matter. This study designs a randomized survey experiment that varies the image treatment within each story while holding fixed other aspects of news articles to discover and estimate characteristics of images that can causally increase or decrease the perception of news credibility. By combining Large Vision Models and methods in NLP to discover latent treatments, I find that the general presence of an image does not uniformly change credibility perception, while images with some identified latent treatments (such as photos from press conferences, comics, or visuals of male suits) can alter credibility perception. Heterogeneous treatment effect analysis also reveals that preferences for some latent treatments are divided by gender, race, or age, suggesting paths for visual treatment targeting.