Flow matching in PyTorch, from toy data to ImageNet

What it does

flow_matching is Meta’s PyTorch translation of the Flow Matching guide. It breaks the theory into practical pieces—losses, paths, schedulers, and solvers—and bundles training recipes for everything from toy 2D datasets to face-blurred ImageNet and text. Think of it as a build-your-own-generative-model kit with the math already filled in.

The interesting bit

Instead of a single reference implementation, the library treats flow matching as a configurable pipeline. You can mix and match paths and solvers, which means experimenting with Riemannian variants or discrete token flows does not require gutting an entire training script. It is explicitly designed to be read alongside the arXiv paper.

Key highlights

• Covers continuous, discrete, and Riemannian flow matching in one codebase.
• End-to-end training examples for images (CIFAR-10, face-blurred ImageNet) and text.
• Modular architecture: loss, path, scheduler, and solver are swappable components.
• No pre-trained checkpoints; every example is train-from-scratch by design.
• CC BY-NC 4.0 license, so keep it in the lab.

Caveats

• No pre-trained models are released, so reproducing larger-scale results requires significant compute.
• The CC BY-NC license prohibits commercial use unless you negotiate separately with Meta.
• Requires Python 3.9 and PyTorch 2.1 or newer.

Verdict

Ideal for researchers who want a clean, modular PyTorch baseline for flow matching experiments. Skip it if you are looking for production-ready weights or a permissive commercial license.