Abstract: We present an example-based image processing algorithm that uses a pair of example training images (an original and a processed version of the image), to model the processing observed. Our algorithm uses this single example pair to infer what a test image's output would be if processed in a similar manner. We create a Markov network representing the new input image and approximate the processing operation as a labeling problem, where each node requires a label corresponding to a pixel from the processed example training image. We then find the MAP labeling by using belief propagation. Our framework finds globally satisfactory labelings and can be used to generate non-photorealistic renderings, super-resolution of images, and transfer of texture from one image to another.