Abstract:
Modeling realistic and controllable 3D digital humans is a central focus in computer vision
and graphics, with broad applications in AR/VR, virtual try-on, entertainment, and embodied
AI. Achieving high-fidelity avatars requires accurate modeling of how people look, dress,
and move. However, this remains challenging due to high variability in clothing types, body
shapes, poses, and their interdependencies. This dissertation addresses two core aspects
of human digitization – how we dress and how we move – through novel representations,
datasets, and learning-based methods that advance the state of 3D human modeling.
Our contributions to modeling human clothing address multi-faceted challenges across
geometry representations, learning formulations, and practical applications. We explore
how clothing and body geometry are influenced by size, shape, pose, and style, proposing
specialized representations and learning frameworks—supported by high-quality datasets—to
enable more robust and generalizable models of clothed humans.
We introduce SIZER, the first model to capture clothing deformations based on clothing
size and body shape using a mesh-based representation. SizerNet models the subtle, non-
linear effects of size and shape on clothing fit and drape – factors often overlooked in prior
work. To support this, we curate a large-scale, high-quality dataset of 3D scans with diverse
subjects, clothing types, and multiple sizes per item. Rich annotations, including clothing
categories, segmentation, registrations, and accurate body shapes, enable realistic, data-driven
modeling. Additionally, we introduce ParserNet, the first learning-based method to recover
detailed clothing meshes and body shape from a single input mesh. Unlike prior methods, it
allows editing clothing or body shape without requiring segmentation or registration at test
time. Together, the dataset and models form a strong foundation for future work in clothed
human reconstruction, generation, and size-sensitive modeling.
While mesh-based approaches excel in template-specific garments, they struggle with
non-rigid deformations, topological diversity, and require costly registration and processing
pipelines. To address this, we propose Neural-GIF, an animatable model of clothed humans
learned directly from raw scans using neural implicit representations. Inspired by template-
based decomposition of motion into articulation and non-rigid deformation, we generalize
this idea to implicit shape learning. We learn a pose-dependent backward skinning field to
map input points to a canonical space, followed by a deformation field in the canonical space
to capture non-rigid, pose-dependent deformation. This enables modeling of highly complex
clothing and soft-tissue dynamics. Neural-GIF achieves fine-grained details, shape-aware
deformations, and outperforms prior methods in flexibility and accuracy. We also introduce a
4D scan dataset with loose, complex clothing to support further research in this area.
Prior work often models clothing and body as a single surface or focuses on isolated
clothing styles, failing to capture the diversity and layered nature of real-world clothing.
To address this, we propose TailorDiff, a unified, controllable, diffusion-based point cloud
viii
model that serves as an SMPL-like foundation model for clothing. By leveraging a point
cloud representation, TailorDiff is topology-agnostic, enabling unified modeling across all
categories of clothing. A hierarchical, body-aware latent space supports generation and
editing conditioned on style, body shape, and fine-grained attributes like sleeve or hem length.
To train this large-scale model, we curate a rich and diverse dataset by aggregating synthetic
clothing scans from multiple sources. TailorDiff also readily supports inverse problems, such
as fitting layered clothing to partial scans – pushing the boundaries of SMPL-based human
modeling towards more expressive, realistic representations of humans in layered clothing.
Beyond modeling how we dress, this dissertation also addresses how we move. Pose
priors are crucial for generating realistic new poses and for reconstructing accurate poses
from noisy or partial observations. Existing priors rely on Gaussian assumptions, which
bias estimates towards the mean and distort the pose manifold. We introduce PoseNDF, a
novel formulation for learning pose priors using a Neural Distance Field, without relying on
explicit density modeling or prior assumptions. PoseNDF extends implicit surface modeling
in 3D to high-dimensional pose space, learning the valid pose manifold from sets of plausible
and implausible poses via a neural network. The zero level set of the learned manifold defines
valid poses, enabling projection using the distance and its analytic gradient. Using PoseNDF
as prior yields state-of-the-art results in motion denoising, pose completion, and single-image
3-D pose estimation while increasing sample diversity.
This thesis advances the field of 3D digital humans by pushing the frontiers of human-
clothing models and human pose understanding. We provide an in-depth analysis of our ideas
and position our work in the context of related research, including approaches that emerged
concurrently. We conclude with an outline of future directions, aiming not only to address
current limitations but also to further advance the field. As the demand for immersive digital
environments continues to grow, we envision our contributions serving as a foundation for
building expressive, generalizable, and fully interactive 3D digital humans