Synthesis refers to the generation of an image of an object, usually from some pre-determined model of that object. We have adopted a trainable, example-based approach to synthesis, as described in ``Example-based Image Analysis and Synthesis'' (D. Beymer, A. Shashua, and T. Poggio). Briefly, and omitting many details, the cornerstones of the approach are as follows:

• A set of example images of the object are obtained, which capture the particular set of facial motions that need to be synthesized.
• Each of the example images are associated with a position in a high-level, multidimensional parameter space.
• Finally, a learning network learns the mapping from the parameter space to the set of example images chosen.

As a result, a synthesis module is built which is capable, for any particular set of input parameters, of synthesizing an image that corresponds to that location.

The sequences in Figures A and B were generated from the same 14-example network that could synthesize pose, eye, and mouth movements. This illustrates the versatility of our trainable, example-based synthesis approach, which can synthesize multiple pathways between the same images.

Our synthesis approach has been also demonstrated in a large set of other facial domains, such as head pose movements, and mouth expressions such as smiles, frowns, and laughs.