Flow matching is a way to train a generative model that turns random noise into data by following a learned velocity field. It is closely related to diffusion but uses straighter paths, so it needs fewer steps to sample. It is used for images, audio and robot action policies. It builds on Standard Diffusion Models.

Path From Noise To Data

Take a data sample and noise , and join them with a straight line:

  • at the point is pure noise, at it is data
  • the velocity along this line is constant

Training

  • sample data , noise and a time
  • form , and train a network to predict the velocity
  • is the conditioning, for example a text prompt or a robot's observation
  • this is plain regression, so training is simple and stable

Sampling

  • start from noise
  • integrate the learned field with small steps, for example 10 Euler steps

Why Not Direct Regression

  • when there are several valid outputs (go left or right round an obstacle), MSE regression predicts their average, which may be invalid
  • flow matching learns the whole distribution, so sampling gives one valid output

Compared With Diffusion

  • diffusion adds noise along a curved schedule and learns to predict noise or the score
  • flow matching uses straight paths and predicts a velocity directly, so sampling usually needs fewer steps
  • both condition on the input in the same way, for example with cross-attention

In Robot Policies

  • the generated sample is an action chunk, a short sequence of future actions
  • an action expert conditioned on vision and language features predicts the velocity, as in SmolVLA