lucidrains/rectified-flow-pytorch

What it solves

This project provides a PyTorch implementation of Rectified Flow, a generative modeling technique designed to make the path between noise and data more direct (straight), which simplifies the sampling process and improves efficiency.

How it works

It implements the core Rectified Flow algorithm and several subsequent research improvements. The library allows users to define a model (such as a UNet) and wrap it in a RectifiedFlow class to handle the training loss and sampling. It also supports "reflow"—a process of iteratively refining the flow to further straighten the trajectories, potentially enabling faster generation.

Who it’s for

Researchers and developers working on generative AI, specifically those interested in flow-matching and alternatives to traditional diffusion models for image synthesis.

Highlights

• Reflow Support: Includes the Reflow class to iteratively straighten the generative paths.
• Integrated Trainer: Provides a Trainer based on the accelerate library for easier model training.
• Diverse Implementations: Includes advanced variants like LSD flow, SoFlow, Split Mean Flow, Unconstrained Alignment (UA) Flow, and Laplacian multiscale flow matching.
• Ready-to-use Components: Comes with a built-in Unet and ImageDataset for quick experimentation.