Models for depth prediction are usually trained on realistic photo/video data. However, exploring how models perform on artwork also produces some interesting results. We found that the model actually performs surprisingly well on these images. In our project, we qualitatively explore the depth interpretations of a database of art history images with an interactive visualization (depth-maps-art-and-illusions/art-history-vis), looked at how models perform on optical illusions (depth-maps-art-and-illusions/depth-visualizer), and experimented with generating adversarial examples (depth-maps-art-and-illusions/adversarial).