Let us Flow Together ༄࿐࿔🚀

The Statistical Learning and AI Lab at UT Austin.

cutecat.png

Rectified flow is a unified framework for generative modeling that offers a simplified view on flow- and diffusion-based techniques. It has been applied to state-of-the-art image, audio, and video generation.

In a nutshell, rectified flow interpolates between noise and data distributions through an interpolation process. This process is then "rectified" (or "causalized") to produce a generative ODE model. The procedure can be repeated to "reflow" the system, yielding ODEs with straighter trajectories that can be discretized using fewer, or even a single, step.

This series of tutorials on rectified flow addresses topics that are often sources of confusion and highlights the connections between rectified flow and other generative modeling methods.

We provide a codebase and lecture notes with detailed theoretical derivations.

If you have questions regarding the blog posts, codebase, or notes, please feel free to reach out via this email.