Beyond the Scan: High-fidelity Human Digitalization at Scale

As the digital world evolves into an immersive extension of our daily lives, the digital avatar has emerged as the critical interface between human reality and digital intelligence. However, current creation methods suffer from a persistent bottleneck: the "trade-off" between the realism of expensive industrial scanning rigs and the scalability of parametric models that often lack the texture and clothing detail necessary for true visual presence. This talk presents a principled 3D avatar framework designed to bridge this gap by enabling the automated creation of realistic, animatable digital humans. We explore a technical pipeline that moves beyond traditional parametric models to capture detailed geometry and appearance from consumer-level sensors, even in complex scenarios involving multiple interacting subjects. By leveraging large-scale, unstructured 2D datasets, we encode human priors into a compressed latent space, allowing for the generation of diverse digital identities. Finally, we demonstrate how these learned generative priors resolve the inherent ambiguities of monocular reconstruction, enabling the synthesis of high-fidelity 3D avatars from a single image. This research offers a scalable roadmap for democratizing digital identity and advancing human-centric computer vision tasks at scale.