Markov Random Fields (MRFs) are proposed as viable stochastic models for the spatial distribution of gray levels for images of human faces. These models are trained using data bases of face and non-face images. The trained MRF models are then used for detecting human faces in test images. We investigate the performance of the face detection algorithm for two classes of MRFs given by the first- and second-order neighborhood systems. From the cross validation results and from actual detection in real images, it is shown that the second-order model makes fewer false detections. We also investigate the possibility of increasing our training data base of faces by simulating face-like images from the trained MRFs. The performance of the retrained MRFs based on added face-like images is compared to the original training data base.