By collaborating with a number of professional photojournalists, we have obtained access to more than one million images from conflict zones – mainly Iraq. We want to use these images to examine how conflicts are depicted and narrated. From the total population of photographs in the raw shoot, we know which images where submitted to editors, and which where published. For each image we know the time and date of capture. For some images we also know the approximate location (which given the time and data stamps can be extrapolated to all images). To learn the content of the images, we manually label a subset of images. We use this subset in cohort with artificial neural networks to classify/estimate the content of all our images. Combining our image-data with the UCDP geocoded conflict event dataset, we can, e.g., asses the association between the level of violence and the content in the images. Furthermore, because we know if images were submitted to editors or not, and if they were eventually published, we can explore choices and biases throughout the publication process. For instance, we can explore whether images portraying gender roles in a particularly manner are prioritised through the publication process.