Abstract
Our success with InstaFlow builds upon our previous algorithm, rectified flow. This algorithm is surprisingly simple, which addresses the problem of learning a transport mapping between two distributions using unpaired data points. Such problems include generative modeling and unsupervised domain transfer. Rectified flow is an ordinary differential equation (ODE) model that is trained to follow straight paths as much as possible, using pure supervised learning with L2 objectives. I will then introduce a special procedure, called reflow, that iteratively straightens the probability flow trajectories and simultaneously enhances coupling between noises and images.
Speaker
Xingchao Liu is a final-year PhD student in the Computer Science Department of the University of Texas at Austin. His research interests span across various domains in machine learning, with a particular focus on probabilistic inference, generative modeling and their practical applications. Recently, He is working on developing efficient diffusion models.