← AI Terminology
Flow Matching / Rectified Flow
Flow matching trains continuous normalising flows by regressing vector fields that transport noise to data; rectified flows learn straighter paths for faster sampling.
A modern generative modelling family competing with diffusion.
A modern generative modelling family competing with diffusion.
Why It Matters in AI
Diffusion dominates, but flow matching/rectified flows offer simpler training stories and efficient few-step generation. They influence image and protein generative research and some production explorations.
Key Points
| Aspect | Description |
|---|---|
| Use | Image synthesis research; science generative models |
| Idea | Learn velocity field along probability paths |
| Related | Score-based models, CNFs, diffusion |
| Sampling | ODE solvers with few function evaluations |
| Vs diffusion | Related continuous-time views; different training target |
| Rectified flow | Straighter ODE paths → fewer steps |
Simple Analogy
Learning the wind map that blows scattered dust grains into the shape of a sculpture along smooth paths — then running the wind forward quickly.
Common Usage Examples
- Rectified flow papers and open repos
- Few-step image samplers
- Compare FID vs diffusion at NFE budget
- Latent flow models
Summary
In short: Flow matching learns transport fields from noise to data — a powerful generative alternative with efficient straight-path variants.