Facial motions produce not only facial feature points motions, but also subtle appearance changes such as wrinkles and shading changes. These subtle changes are important yet difficult issues for both analysis (tracking) and synthesis (animation). Previous approaches were mostly based on models learned from extensive training appearance examples. However, the space of all possible facial motion appearance is huge. Thus, it is not feasible to collect samples covering all possible variations due to lighting conditions, individualities, and head poses. Therefore, it is difficult to adapt such models to new conditions. In this paper, we present an adaptive technique for analyzing subtle facial appearance changes. We propose a new ratio-image based appearance feature, which is independent of a person?s face albedo. This feature is used to track face appearance variations based on exemplars. To adapt the exemplar appearance model to new people and lighting conditions, we develop an online EM-based algorithm. Experiments show that the proposed method improves classification results in a facial expression recognition task, where a variety of people and lighting conditions are involved.