Computer-generated faces, or avatars, are becoming increasingly sophisticated, but are visually unrealistic, particularly in motion, and their control remains problematic. Previous work has implemented complex three-dimensional polygonal models, often generated from laser-scans, with intricate hard-coded muscle models for actuation of speech and expression. Driving the avatar through mimicry, or performance-driven animation, involves tracking a real actor's facial movements and associating them with analogues on the model. Motion is tracked usually through markers physically attached to the actor's face or by locating natural feature boundaries. Here, complex three-dimensional models are avoided by taking an image-based approach. Novel techniques are presented for automatically creating and driving photo-realistic moveable face models, generated from example footage of a face in motion. Image changes for each frame, coupled with dense motion fields extracted using an optic flow algorithm, are analyzed to extract a set of basis actions by application of principal components analysis. These techniques yield a virtual avatar onto which the movements of an actor can automatically be projected for convincing performance-driven animation, with no need for markers.