Flow Matching is a generative modeling framework that learns continuous normalizing flows by regressing onto simple vector fields, providing a simpler and more flexible alternative to diffusion models with equivalent or better generation quality.

Architecture Overview

Flow Matching defines a time-dependent probability path p_t(x) that continuously transforms a simple prior distribution (typically Gaussian noise, p_0) into the target data distribution (p_1). This transformation is described by an ordinary differential equation (ODE): dx/dt = v_t(x), where v_t is a time-dependent velocity field.

A neural network u_θ(x, t) is trained to approximate the true velocity field v_t(x) by minimizing the flow matching objective: L = E_{t, x_t} ||u_θ(x_t, t) - v_t(x_t)||². The key insight is that while the marginal velocity field v_t is intractable, conditioning on individual data points x_1 yields simple conditional flows with known velocity fields.

The simplest conditional flow is the optimal transport path: x_t = (1-t)·x_0 + t·x_1 (linear interpolation between noise and data), giving a conditional velocity field u_t(x|x_1) = x_1 - x_0. At inference, samples are generated by integrating the learned velocity field from t=0 to t=1 using an ODE solver.

Key Innovations

  • Simulation-free training: Unlike continuous normalizing flows (CNFs), flow matching trains without expensive ODE solves during training—just simple regression on velocity vectors
  • Optimal transport paths: Straight-line interpolation paths between noise and data are shorter and simpler than diffusion paths, enabling faster sampling with fewer ODE steps
  • Unified framework: Diffusion models, stochastic interpolants, and rectified flows are all special cases of flow matching with different path choices
  • Flexibility: Any probability path connecting noise to data can be used, not just the specific forward processes required by diffusion models

Common Use Cases

Image generation, video generation, text-to-image synthesis (Stable Diffusion 3 uses flow matching), audio synthesis, molecular generation, protein structure generation, and any generative task where diffusion models were previously used.

Notable Variants & Sizes

Conditional Flow Matching (CFM), Rectified Flow (closely related formulation), Stable Diffusion 3 / SD 3.5 (flow matching with MMDiT backbone), Flux (Black Forest Labs), SiT (Scalable Interpolant Transformers), and Voicebox/Audiobox (flow matching for speech). The backbone network can be any architecture—U-Net, DiT, or custom.

Technical Details

Training: sample t ~ U(0,1), x_0 ~ N(0,I), x_1 ~ p_data, compute x_t = (1-t)x_0 + t·x_1, target velocity = x_1 - x_0. Loss: ||u_θ(x_t, t) - (x_1 - x_0)||². The neural network u_θ typically uses a DiT or U-Net architecture conditioned on timestep t (via sinusoidal embeddings + adaLN). Inference: solve dx/dt = u_θ(x,t) from t=0 to t=1 using Euler (simplest, ~20-50 steps) or adaptive ODE solvers (dopri5). Flow matching with OT paths often needs fewer steps than DDPM diffusion (10-25 vs 50+). Classifier-free guidance applies identically: u_guided = u_uncond + w·(u_cond - u_uncond). SD3 uses flow matching with a logit-normal timestep distribution that concentrates training on intermediate noise levels.