<p>Single sample face and ear recognition (SSFER) is a difficult problem in the area of biometric recognition, as only a single face/ear training image is available. It becomes more challenging when the test images are captured with varying lighting, occlusions, expressions, etc. Virtual sample generation is popular for extending the training set and improving the feature extraction. This paper proposes a novel method for generating virtual samples in SSFER based on HyperBoxes to generate any number of quality virtual samples along with improving classification accuracy. We conducted extensive tests on images from five databases (ORL, YALE, and AR (illumination) face databases) and (AMI and IITD ear databases) for classification accuracy, image quality of virtual samples, and average testing time. During the evaluation, training is done with one of the variations and its virtual samples, and testing is done for the remaining variations in the database. Principal component analysis is used for feature extraction, and a K-nearest neighbour classifier is used for classifying test images. Results show that the proposed method improves the classification accuracy and virtual sample image quality when compared with several state-of-the-art methods for all the databases. The average testing time results are also found to be comparable.</p>