An important goal of character animation is to create believable, life-like movements and expressions. For film, artists spend significant amounts of effort to add sufficient complexity to a character rig to enable believable and emotionally evocative performances. However, once a complex character rig has been authored, an artist then needs to spend a significant amount of effort to animate a character and bring it to life. For video games, mesh deformations and geometry processing must be real-time, which affects the types of deformations included in a character rig for an interactive application. As a result, video game characters tend to lack some of the sophisticated deformations and motions seen in film-quality characters.

This dissertation explores applications of machine learning for improving the quality of deformations in real-time character rigs as well as applications to assist artists in producing high-quality animations. We detail a deep learning-based approach to enable complex film-quality mesh deformations to run in real-time for both a character's body and face. Our method learns mesh deformations from an existing character rig and produces an accurate approximation using significantly less computational time. In addition to mesh deformations, we present a statistical approach to synthesize novel animations from a collection of artist-created animations. Thus, single-use animations for film can be leveraged for additional applications. We also present a method for generating facial animation from a recorded performance, which provides artists with an initial animation that can be fine-tuned to meet stylistic and expressive needs.