Advances in 3D human modeling: neural representations and generative learning
File(s)
Author(s)
Zheng, Jiali
Type
Thesis
Abstract
The creation of highly realistic and adaptable 3D digital human models has become essential to the development of virtual and augmented reality (VR/AR), gaming, digital fashion, and interactive media. While neural representations and generative learning have shown significant potential for advancing 3D human modeling, their success is still constrained by key limitations, including the scarcity of large-scale, high-fidelity 3D human datasets and the limited capacity of current models to generalize under real-world conditions marked by sparse data and visual imperfections.
This thesis addresses these critical challenges by advancing both the data and model foundations necessary for robust and scalable 3D human modeling, where the demand for photorealism, variability, and dynamic behavior is particularly acute. First, we introduce a large-scale dataset of real-world 3D garments alongside Design2Cloth, a differentiable generative model for 3D garment synthesis that enables diverse and detailed clothing generation through an intuitive 2D mask-based input mechanism. Second, we present the Imperial Light-Stage Head (ILSH) dataset, a high fidelity human head dataset captured under sparse viewpoints, designed to support research on photorealistic head synthesis and novel view generation. Lastly, we propose a robust generative and reconstruction framework for full-body clothed humans, capable of producing highly detailed and animatable avatars from a single RGB image, even under occlusions and challenging viewing conditions.
The contributions of this thesis advance state-of-the-art methods in 3D human generative modeling by providing new tools and datasets for generating and reconstructing digital humans with fine-scale detail and high realism. The presented frameworks enable practical applications in areas such as immersive content personalization, virtual try-on systems, virtual production, and telepresence, while also serving as a foundation for future research in neural 3D human modeling and reconstruction.
This thesis addresses these critical challenges by advancing both the data and model foundations necessary for robust and scalable 3D human modeling, where the demand for photorealism, variability, and dynamic behavior is particularly acute. First, we introduce a large-scale dataset of real-world 3D garments alongside Design2Cloth, a differentiable generative model for 3D garment synthesis that enables diverse and detailed clothing generation through an intuitive 2D mask-based input mechanism. Second, we present the Imperial Light-Stage Head (ILSH) dataset, a high fidelity human head dataset captured under sparse viewpoints, designed to support research on photorealistic head synthesis and novel view generation. Lastly, we propose a robust generative and reconstruction framework for full-body clothed humans, capable of producing highly detailed and animatable avatars from a single RGB image, even under occlusions and challenging viewing conditions.
The contributions of this thesis advance state-of-the-art methods in 3D human generative modeling by providing new tools and datasets for generating and reconstructing digital humans with fine-scale detail and high realism. The presented frameworks enable practical applications in areas such as immersive content personalization, virtual try-on systems, virtual production, and telepresence, while also serving as a foundation for future research in neural 3D human modeling and reconstruction.
Version
Open Access
Date Issued
2025-04-11
Date Awarded
2026-06-01
Copyright Statement
Attribution-Non Commercial-No Derivatives 4.0 International Licence (CC BY-NC-ND)
Advisor
Zafeiriou, Stefanos
Publisher Department
Department of Computing
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)