Improving the Training of Rectified Flows

Sangyun Lee, Zinan Lin, Giulia Fanti

Diffusion models have shown great promise for image and video generation, but sampling from state-of-the-art models requires expensive numerical integration of a generative ODE. On…