Generative Models for Guided 3D Shape Modeling

Event details
Date | 26.08.2025 |
Hour | 14:00 › 16:00 |
Speaker | Yingxuan You |
Location | |
Category | Conferences - Seminars |
EDIC candidacy exam
Exam president: Prof. Devis Tuia
Thesis advisor: Prof. Pascal Fua
Co-examiner: Prof. Jürg Schiffmann
Abstract
With the success of generative models in the field of 2D image generation, the generation of 3D shapes has attracted increasing attention. While recent methods can produce visually compelling 3D geometry, generating shapes that also satisfy critical physical, semantic, and geometric constraints remains a major open challenge. This limitation hinders the adoption of 3D generative techniques in domains such as industrial design, aerospace, and precision engineering, where constraint satisfaction is essential. To address this issue, my research focuses on integrating guidance into the generative process. By incorporating relevant constraints directly into the shape generation pipeline, the goal is to produce 3D shapes that are not only visually realistic but also functionally valid, without the need for time-consuming post-optimization. For example, in the design of vehicles and aircraft, aerodynamic constraints will be introduced to guide the generation toward shapes with minimal drag. In garment modeling, material properties and body-cloth interactions will be considered to ensure physically accurate draping behavior. This research will involve the design of novel algorithms for constraint-aware 3D shape generation, along with extensive experiments on benchmark datasets and real-world applications to validate the effectiveness of the proposed methods.
Selected papers
Exam president: Prof. Devis Tuia
Thesis advisor: Prof. Pascal Fua
Co-examiner: Prof. Jürg Schiffmann
Abstract
With the success of generative models in the field of 2D image generation, the generation of 3D shapes has attracted increasing attention. While recent methods can produce visually compelling 3D geometry, generating shapes that also satisfy critical physical, semantic, and geometric constraints remains a major open challenge. This limitation hinders the adoption of 3D generative techniques in domains such as industrial design, aerospace, and precision engineering, where constraint satisfaction is essential. To address this issue, my research focuses on integrating guidance into the generative process. By incorporating relevant constraints directly into the shape generation pipeline, the goal is to produce 3D shapes that are not only visually realistic but also functionally valid, without the need for time-consuming post-optimization. For example, in the design of vehicles and aircraft, aerodynamic constraints will be introduced to guide the generation toward shapes with minimal drag. In garment modeling, material properties and body-cloth interactions will be considered to ensure physically accurate draping behavior. This research will involve the design of novel algorithms for constraint-aware 3D shape generation, along with extensive experiments on benchmark datasets and real-world applications to validate the effectiveness of the proposed methods.
Selected papers
- Paper1: Diffusion Models in Vision: A Survey (https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=10081412)
- Paper2: Aligning Optimization Trajectories with Diffusion Models for Constrained Design Generation (https://arxiv.org/pdf/2305.18470)
- Paper3: TripoSG: High-Fidelity 3D Shape Synthesis using Large-Scale Rectified Flow Models (https://arxiv.org/pdf/2502.06608)
Practical information
- General public
- Free
Contact
- edic@epfl.ch