Event Date
Speaker: Qiang Liu, Associate Professor, Computer Science, University of Texas, Austin
Title: "Rectified flow: A straight approach to generative modeling and transport mapping"
Abstract: Rectified Flow (RF) is a simple yet general approach to generative and transfer modeling, widely applied in state-of-the-art AI tasks such as image, video, audio, and molecule generation. It provides a straightforward method for learning continuous-time transport mappings between two distributions—observed through either unpaired or paired data points—by learning neural ordinary differential equation (ODE) models that prioritizes path straightness. Straight paths are naturally preferred and allow for fast simulation with large discretization step sizes, enabling efficient one-step or few-step models. This hence yields a notion of *straight* transport, which differs from the classical notion of optimal transport. Although based solely on ODEs, RF can be extended to offer simplified perspectives on existing diffusion models.
Some related blog posts on the topic: https://www.cs.utexas.edu/~lqiang/rectflow/html/intro.html, https://rectifiedflow.github.io/
Bio: Qiang Liu is an associate professor of computer science at UT Austin. He is interested in advancing fundamental algorithms in machine learning through mathematical, statistical, physical insights.
Faculty Webpage (links to UT Austin): https://www.cs.utexas.edu/~lqiang/