“Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow,” posted to arXiv on September 7, 2022 by Xingchao Liu, Chengyue Gong, and Qiang Liu, proposed a clean approach to fast generative modeling. Like diffusion models, it learns to transform random noise into data, but it does so by following paths that are as straight as possible.
The method trains an ordinary differential equation to transport between the noise distribution and the data distribution along the straight line connecting paired points. Curved paths force a generator to take many small integration steps to stay accurate; straight paths can be traversed accurately in very few steps, sometimes even one. The authors introduced a reflow procedure that iteratively straightens the learned paths, progressively reducing the number of steps needed to produce a high-quality sample, and showed the framework applies both to generation and to transferring between two arbitrary distributions.
Rectified flow has become a foundation for some of the most capable recent image generators; later large text-to-image systems adopted rectified-flow training to combine high quality with fast sampling. For a general reader, it represents one of the leading answers to a central engineering challenge of generative AI: how to keep the realism of step-by-step methods while making generation fast enough to be cheap and responsive.