We propose a method for generating caustic images in real time using a deep/convolutional neural network (CNN). To do so, training images are first rendered using photon mapping, and the CNN learns the correspondences between the depth images and caustic images. After learning, the CNN generates a caustic image from a depth image within 55 milliseconds. In addition, the similarity between the generated caustic images and the ground truth shows that our method is very promising for the generation of caustic images for a number of known objects, though the method does not handle objects in which ground truth is not already known. This method can play an important role in scenes used for stage video production and interactive art in the future.