Concept
Flow Matching
A newer, simpler alternative to diffusion for generative AI - training models to move data along straight paths between noise and real samples, rather than along the curved random walks diffusion uses.
Added May 18, 2026
Diffusion models work by learning to reverse a specific stochastic process - the gradual addition of Gaussian noise over many small steps. This process is mathematically convenient but somewhat arbitrary: there is no particular reason that adding Gaussian noise is the best way to connect the data distribution to a simple noise distribution. Flow matching emerged as a more flexible and often more efficient alternative.
Flow matching trains a model to learn a vector field - a function that, at any point in space, says which direction to move to get from the noise distribution to the data distribution. Instead of learning to reverse noise addition step by step, the model learns to follow continuous "flows" that carry random noise samples to realistic data samples. The paths are simpler, often straighter, and can be traversed in fewer steps than diffusion.
The training process is elegant. For each training example, choose a random noise sample and define a straight-line path from that noise to the real data point. Train the model to predict the direction of travel (the velocity vector) at each point along this path. With enough examples, the model learns a vector field that, when followed from any noise sample, converges to the data distribution.
Why straight lines? Straight-line paths between noise and data are the shortest possible paths, which means they can be traversed in fewer steps during generation. Diffusion models typically require 20 to 1,000 sampling steps. Flow matching models can often produce high-quality samples in as few as 8 to 20 steps, with some configurations producing good results in a single step.
Meta's Voicebox speech synthesis model was one of the early high-profile uses of flow matching. Stability AI's Stable Diffusion 3 and Black Forest Labs' FLUX models use flow matching at their core, demonstrating that the approach can match or exceed diffusion quality while being more efficient. Flow matching has also shown promise in protein structure prediction, drug discovery, and other scientific generation tasks.
Analogy
Teaching someone to navigate from a randomised starting position to a known destination. Diffusion is like teaching them to reverse a random walk with many small zigzag steps - the path is convoluted and you need many corrections. Flow matching teaches them to walk in a straight line from start to destination - simpler to learn, faster to traverse.
Real-world example
FLUX.1, released by Black Forest Labs in 2024 and regarded as among the best open text-to-image models, uses flow matching. It generates higher-quality images than comparable diffusion models and does so in fewer inference steps - typically 20-28 steps compared to the 50-100 steps commonly used with older diffusion models. The quality improvement at equal step counts is one of the practical advantages of flow matching.
Why it matters
Flow matching matters because faster sampling is not just a convenience - it directly affects what AI generation is practical in real-world applications. Generation that takes 30 seconds is a thoughtful creative tool; generation that takes 2 seconds enables real-time interactive creativity. As flow matching becomes the dominant approach for generative models across images, video, audio, and scientific domains, it will enable qualitatively new applications.
In the news
Related concepts