Flow Matching

Flow matching is a simulation-free training framework for generative models. It fits a time-dependent velocity field — a neural network that says how a sample should move at each moment — so that integrating it transports samples from a simple source distribution (typically a standard Gaussian) to the data distribution along a prescribed probability path. It was introduced by Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel and Matt Le of Meta AI in "Flow Matching for Generative Modeling" (arXiv:2210.02747, October 2022), presented at ICLR 2023.^[1]

Background

Flow matching grows out of continuous normalizing flows (CNFs), introduced with Neural Ordinary Differential Equations (arXiv:1806.07366, NeurIPS 2018 best paper). A CNF defines a generator by integrating the ODE dx/dt = v(x, t): draw noise at time 0, follow the learned field, arrive at a sample at time 1. The model class is elegant — deterministic sampling, exact likelihoods — but maximum-likelihood training requires solving the ODE forward and backward on every gradient step. This simulation cost kept CNFs from scaling to large image datasets for years.^[1][2]

Key concepts

Flow matching's contribution is to replace likelihood training with vector-field regression: choose the probability path you want, then train the network to predict the target velocity along that path. The catch is that the marginal target field is intractable over a whole dataset. The paper's central trick is to express the target path as a mixture of per-sample conditional paths, for which the field is known in closed form, and to show that regressing on the conditional field yields exactly the same gradient as the intractable marginal objective — the conditional flow matching (CFM) loss. Training therefore involves no numerical ODE integration at all.^[1][3]

The choice of path is a design knob. Straight-line (optimal-transport) paths, which move each noise sample directly toward a data sample, produce straighter inference trajectories and fewer function evaluations. Training CNFs this way outperformed the standard DDPM diffusion objective on CIFAR-10 and ImageNet across FID, likelihood and sampling steps simultaneously.^[3] A close relative, rectified flow ("Flow Straight and Fast", arXiv:2209.03003, ICLR 2023 spotlight), uses the same independent-coupling target and adds a reflow step to straighten trajectories further.^[4] Viewed this way, diffusion training is a special case of the framework rather than the other way around.

Applications

By 2024 flow matching variants had become the default training objective for non-text generative models: Stable Diffusion 3 is built on a rectified-flow transformer, and open video models such as Alibaba's Wan and NVIDIA's Cosmos train with flow matching.^[3] Extensions such as Discrete Flow Matching (arXiv:2407.15595, July 2024) define probability paths over finite vocabularies, bringing the approach to text and graph generation; the authors reported 6.7% Pass@1 and 13.4% Pass@10 on HumanEval with a 1.7B-parameter non-autoregressive flow model.^[5] Practically, the framework's appeal is efficiency: straighter paths mean fewer network evaluations per generated sample, which translates directly into faster and cheaper inference.

>See Also

• Diffusion Model
• Generative AI
• Stable Diffusion
• Wan
• Seedance
• Text-to-Image Model
• Variational Autoencoder