Abstract: Digital personalized avatars are essential for user identity in the metaverse, and their creation has become an exciting area of research with applications in gaming, digital twins, and more. With the rapid advancements in machine learning within both computer vision and graphics, creating personalized, game-ready avatars has become a feasible reality. In this talk, we will explore our journey in addressing this challenge by integrating and extending various research models, including stable diffusion, eg3d, and diffusion net, to build a comprehensive pipeline aimed at achieving product-level goals. We will discuss the numerous challenges encountered, such as personalization, stylization, high-quality 3D generation, geometry separation, retopology, and auto-rigging/caging, as well as potential future directions in this domain.

Bio: Haomiao Jiang is a Principal Engineer and Tech Lead Manager at Roblox, where he focuses on generative AI for avatars and content creation. He received his Ph.D. from Stanford University in 2016. Haomiao has previously worked and performed research at Google Daydream, Google Research, Meta Reality Labs, and OPPO Research US. His research interests include computational photography, machine learning in graphics, and 3D generation.