flow matching (FM)

professionalwidespread

A teaching method where an AI that creates images, audio, or video learns how to reach the final content from random noise via the most direct path possible.

In plain terms

Most of today's image, video, and audio-generating AI start from noise: a cat, landscape, or face must emerge from a cloud of dots resembling static on a TV screen. Flow matching teaches the model where and how fast to change the noise to create a meaningful image.

Imagine it like navigation. The previous, classic diffusion method took you to the goal via winding side streets in many small steps. Flow matching marks out a roughly straight highway between the noise and the final image, and the model learns which direction the arrow points at any point on the route.

Due to the more direct path, fewer steps are needed, meaning images can be created faster and cheaper, and training is simpler and more stable. This is why many image, video, and audio generation systems are built on this method today.

How it works

For training, pairs are formed: one side holds a random noise sample, the other a real image, audio recording, or video clip.
A path is selected between the two, in the simplest case a straight line, and at randomly chosen points along it, the direction and speed required to proceed are checked.
The neural network learns to determine this direction from any intermediate state; from millions of examples, a complete 'flow map' is assembled.
During generation, the model starts from fresh noise and follows learned directions step by step until it reaches the final content. This typically requires a few dozen steps, while specialized accelerated versions need only a few.
Your text prompt continuously influences the direction of the flow during generation, determining what appears in the image.

An everyday example

You type into an image generator, 'fox cub in an autumn forest, watercolor style,' and a few seconds later you receive the image. If the service is powered by, for example, the Stable Diffusion 3 or FLUX model, the image is generated by a model trained with the flow matching method.

Who is affected

marketerFaster and cheaper image and video generation allows you to test more variations in the same time and for the same money. → Marketing

developerIf you are training or integrating a generative model, you will likely choose flow matching-based solutions today. → Code

entrepreneur, executiveThe cost and speed of generation partly depend on this method, thus affecting how much you pay for AI-generated visual content.

everyoneA growing share of AI images, AI videos, and synthetic voices seen online originate from such models.

A common misconception

MythFlow matching is a new image generation program that is replacing diffusion models.

In realityIt is not a product but a training method under the hood. It is a close relative of diffusion models, and with certain settings, the two lead to the same mathematical result.

Limits and risks

Training still requires vast amounts of data and computing power; the method only makes it more efficient.
Fewer steps do not automatically mean better quality: with too few steps, the image may be blurred or flawed.
Generated content can be erroneous, biased, or misleading (for example, deepfakes) just like with other methods.
Copyright and data protection issues depend on the use of training data and final content; this method changes nothing about them.

Similar, but not the same

diffusion model – A diffusion model learns many small denoising steps, whereas flow matching learns the direction of travel leading directly from noise to data.
normalizing flow – The classic normalizing flow requires an invertible network with a constrained architecture, while flow matching works with a freely chosen network and much simpler training.
rectified flow – Rectified flow is a method developed independently and at the same time as flow matching, closely related to it: it specifically uses straight-line paths between noise and data, and further straightens them with a separate procedure. Today, it is mostly referred to as a variant of flow matching.

Related concepts

Broader

generative AI, generative model, machine learning

Narrower

rectified flow, conditional flow matching

On the edge

diffusion model, normalizing flow, image generation, video generation, speech synthesis

A brief history

The method was described by Yaron Lipman and co-authors in the 2022 paper 'Flow Matching for Generative Modeling'. In the same year, other research groups proposed similar ideas, including the authors of rectified flow (Xingchao Liu and colleagues). It became widely known in 2024, when both the Stable Diffusion 3 and FLUX image generation models were built on it.

Where things stand

As of late 2026, a significant share of the leading open image and video generation models is based on flow matching, and the method is also well established in speech synthesis. The research focuses mainly on one- or two-step generation, as well as on robotics and scientific applications.

Status: 2026 November

Also known as

flow matching · flow-matching · FM · flow matching · rectified flow · conditional flow matching · CFM · flow matching modell

Sources

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.