More about HKUST
Multimodal 2D/3D/4D Avatar Generation and Editing
The Hong Kong University of Science and Technology
Department of Computer Science and Engineering
PhD Thesis Defence
Title: "Multimodal 2D/3D/4D Avatar Generation and Editing"
By
Mr. Hongyu LIU
Abstract:
The high-fidelity creation and animation of digital avatars represent a
critical frontier in modern computer vision and graphics, fueled by the
accelerating demand for immersive experiences in the Metaverse, virtual
reality (VR), and digital communication. This thesis addresses the
multifaceted challenges of avatar synthesis, specifically navigating the
complex interplay between geometric accuracy, appearance fidelity, and motion
controllability across 2D, 3D, and 4D modalities. By leveraging a paradigm
shift from traditional reconstruction pipelines to data-driven generative
approaches—specifically Generative Adversarial Networks (GANs), Diffusion
Models, and Autoregressive Models—this research proposes a series of novel
frameworks to enhance digital human representation. In the 2D domain, we
introduce CLCAE to improve GAN inversion for precise image editing and FFCLIP
for free-form, text-driven manipulation through semantic modulations. For
motion transfer, Human MotionFormer utilizes vision transformers to bridge
the gap in complex human movement synthesis. Transitioning to 3D and 4D, we
propose Follow-Your-Emoji for fine-controllable portrait animation and
HeadArtist-VL, which employs a novel Self Score Distillation (SSD) loss to
resolve the "Janus" artifact in text-to-3D head generation. Finally, we
present AvatarArtist, a pioneering open-domain 4D avatarization model that
utilizes a parameterized triplane representation and diffusion priors to
synthesize animatable avatars across diverse artistic styles. Experimental
results across standard benchmarks demonstrate that these methodologies
significantly advance the state-of-the-art in rendering quality, identity
preservation, and temporal coherence, establishing a robust foundation for
the next generation of virtual humans.
Date: Friday, 16 January 2026
Time: 2:30pm - 4:30pm
Venue: Room 5501
Lifts 25/26
Chairman: Dr. Tracy Pik Yin MOK (ISD)
Committee Members: Dr. Qifeng CHEN (Supervisor)
Dr. Xiaomin OUYANG
Dr. Dan XU
Prof. Ping TAN (ECE)
Dr. Hongsheng LI (CUHK)