Speaker: 张彪博士后

Inviter: 贾晓红研究员

Title: Generative-model-friendly representation learning for geometry data

Language: Chinese

Time & Venue: 2025.08.28 14:00-15:00 南楼N202

Abstract: Unlike image generation, where pixels provide a unified, universal representation, 3D generative models lack a single, canonical data format. This talk addresses the critical challenge of representation learning in 3D generation. We will explore how the fundamental choice of 3D representation directly influences the complexity, feasibility, and performance of generative tasks. Different representations present unique trade-offs in expressiveness, computational efficiency, and compatibility with learning algorithms. This talk will systematically examine these representation-specific challenges and their impact on model design and training difficulty. Furthermore, we will review mainstream frameworks for 3D generative modeling, analyzing how they adapt to, leverage, or circumvent the limitations imposed by various 3D representations. The discussion highlights why representation learning remains a central problem in advancing 3D content generation.