Advancing Digital Humans Toward Personalized Human-Centric AI


Loading...

Author / Producer

Date

2026

Publication Type

Doctoral Thesis

ETH Bibliography

yes

Citations

Altmetric

Data

Rights / License

Abstract

Digital human technologies have achieved impressive realism in film and interactive media. However, high-quality results still depend on specialized equipment and expert workflows, which limit their deployment outside studio and laboratory environments. In parallel, recent deep learning methods offer a path toward scalable creation and understanding of digital humans, yet, they often struggle to accommodate the biological variability of real people. This limitation is exacerbated by the scarcity of paired 3D human data, which restricts both generalization and robust adaptation. This thesis addresses this gap by proposing the paradigm of Personalized Human-Centric AI. The central hypothesis posits that explicitly integrating subject-specific biological information into learning-based frameworks can constrain the solution space, thereby mitigating ill-posed problems and reducing reliance on massive training data. To validate this hypothesis, we conduct studies across three foundational domains of digital human technology: digitalization, modeling, and perception. In the domain of digitalization, we present a pipeline that reconstructs a full-body textured mesh from a single image. By conditioning generative models and implicit neural representations on a personalized body mesh as a geometric anchor, our method completes unseen appearances and recovers high-frequency details that remain coherent with the subject’s body shape. In modeling, we introduce a representation that combines the structure of a personalized template with the expressiveness of learned implicit fields. This approach enables controlled local editing and the transfer of appearance details across subjects while maintaining 3D consistency under articulation. In perception, we propose a body fitting framework to enhance monocular 3D pose estimation. By decoupling shape calibration from pose estimation and utilizing a shape-conditioned generative prior to guide the fitting process, our method achieves more accurate poses that better respect personal anatomical constraints. Collectively, these contributions demonstrate that explicit personalization is a practical and effective mechanism for advancing high-fidelity digital humans toward human-centric AI. Looking ahead, we envision the future of this field lies in the inter-pollination of digitalization, modeling, and perception to empower physical AI agents capable of reasoning about humans within the 3D world.

Publication status

Editor

Contributors

Examiner : Marc, Pollefeys
Examiner : Siyu, Tang
Examiner : Ming-Hsuan, Yang
Examiner : Lingjie, Liu

Book title

Journal / series

Volume

Pages / Article No.

Publisher

Event

Edition / version

Methods

Software

Geographic location

Date collected

Date created

Subject

Organisational unit

03766 - Pollefeys, Marc / Pollefeys, Marc

Notes

Funding

Related publications and datasets