Zhengyang Geng

Abstract

Recent advances in generative modeling have shown that diffusion and flow-based approaches can generate high-quality samples, but often at the cost of many iterative steps. In this talk, I will introduce MeanFlow, a new framework for one-step generative modeling. The key idea is an identity linking instantaneous and average velocities of generative flows, enabling a residual operator whose fixed point yields the one-step sampler. I will present the theoretical foundations and practical implications of this approach, and also discuss open challenges.

Our speaker

Zhengyang Geng is a 4th-year Ph.D. student at Carnegie Mellon University, advised by Prof. J. Zico Kolter and working closely with Prof. Kaiming He. His research interest lies in bridging dynamical systems and AI, studying how dynamics can be used to construct and understand neural networks, and how learning dynamics emerge from the interaction between data, models, and environments.

To become a member of the Rough Path Interest Group, register here for free.