3D Avatar Generation

3D avatar generation studies how to build digital humans that preserve a person's appearance and can be viewed or animated from new poses and viewpoints. Methods use parametric models such as FLAME and SMPL-X, neural fields and 3D Gaussian splatting, and increasingly diffusion or transformer models for feed-forward reconstruction from a single image, sparse portraits, or video. Research also targets controllable facial and body motion, speech-driven expression, and realistic clothing and hair, with evaluation on benchmarks such as NeRSemble, ZJU-MoCap, and 4D-DRESS. A central challenge is balancing identity fidelity and 3D consistency with broad generalization, realistic dynamics, and fast inference, especially when parts of the person are unseen or training data are limited.