In Search of Geometric Fidelity and Visual Alignment for Generating 3D Objects

PhD Qualifying Examination


Title: "In Search of Geometric Fidelity and Visual Alignment for Generating 
3D Objects"

by

Mr. Kaiyi ZHANG


Abstract:

Driven by increasing demands in virtual reality, gaming, and industrial 
design, the field of 3D object generation is advancing rapidly. However, 
achieving stable, professional- grade quality remains a significant 
challenge, primarily due to two critical issues: poor geometric fidelity and 
incoherent visual alignment. We observe that current methods often produce 
over-smoothed surfaces or structurally inconsistent shapes. Our analysis 
reveals that these issues stem from distinct fundamental causes. We attribute 
poor geometric fidelity to the limitations of mainstream latent 
representations, specifically, the trade-offs between VecSets, which discard 
high-frequency details, and Sparse Grids, which face computational hurdles. 
Furthermore, we link structural inconsistencies to the generative models’ 
insufficient semantic understanding of conditioning images, as existing 
feature extractors often fail to capture the necessary information for 
spatially coherent generation.

This survey explores strategies to address these challenges by rethinking 
data representations and integration methods. We introduce the Latent 
Flexible Grid representation as a balanced solution that robustly handles 
irregular topologies and enables local editability. Additionally, we examine 
the potential of the Sparse Grid representation for scaling up resolution. To 
fully leverage sparse grids, we propose the Watertight Geometry 
Standardization pipeline, a data-centric approach that normalizes diverse 
mesh formats into a consistent, high-quality training dataset. Finally, to 
improve visual alignment, we discuss strategies for incorporating robust 
semantic priors into the generation pipeline. We argue that leveraging 
Vision-Language Models and enforcing representation alignment with spatial 
information can effectively guide models to produce geometrically precise and 
structurally plausible results.


Date:                   Monday, 15 December 2025

Time:                   10:00am - 12:00pm

Venue:                  Room 2128A
                        Lift 19

Committee Members:      Prof. Long Quan (Supervisor)
                        Dr. Dan Xu (Chairperson)
                        Dr. Long Chen