Liu, Gong, and Liu introduced Rectified Flow in Flow Straight and Fast, constructing ordinary differential equations along linear interpolations z_t = t x_1 + (1 - t) x_0 between noise and data distributions. An iterative reflow step rewires crossing paths into straighter trajectories so coarse Euler ODE solvers maintain accuracy at low step counts.
Flow Matching
Also known as: Rectified Flow · Conditional Flow Matching · CFM · MeanFlow · Self-Flow
Plain English
A generative modeling framework that learns how to smoothly transport random noise along straight paths into structured images, video, audio, or 3D shapes. Because its trajectories curve far less than classic diffusion paths, models generate sharp outputs in far fewer sampling steps.
Technical Overview
Flow matching trains a time-dependent neural velocity field v_θ(x_t, t) via simulation-free regression onto per-example conditional vector fields, avoiding the expensive ODE integration required by maximum-likelihood Continuous Normalizing Flows and replacing curved variance-preserving diffusion schedules. Using linear Optimal Transport / Rectified Flow paths (x_t = t x_1 + (1 - t) x_0), logit-normal timestep weighting, and resolution-dependent schedule shifting, DiT backbones scale cleanly across images (Stable Diffusion 3, FLUX), video (Wan), speech (F5-TTS), and 3D assets (TRELLIS) solved with Euler ODE samplers. Subsequent 2025–2026 formulations—average-velocity MeanFlow for one-step sampling without Distillation & Synthetic Data, manifold-anchored Pixel MeanFlow, and dual-timestep Self-Flow—removed external VAE and teacher-encoder bottlenecks.
Timeline (6)
2022
Paper- Paper
Lipman, Chen, Ben-Hamu, Nickel, and Le published Flow Matching for Generative Modeling, proving that Continuous Normalizing Flows can train simulation-free by regressing a neural vector field onto per-example conditional vector fields. Under Gaussian Optimal Transport conditional paths, constant-speed linear flows replace the curved noise schedules and score-matching objectives of classic Diffusion Models.
2024
PaperEsser et al. scaled rectified-flow velocity prediction to 8B-parameter MMDiT backbones for Stable Diffusion 3, replacing uniform noise-level weighting with logit-normal timestep sampling concentrated on intermediate scales and adding resolution-dependent schedule shifting. That weighting-and-shift parameterization rapidly became standard in FLUX.1 and Wan 2.1.
2025
PaperGeng et al. published Mean Flows for One-step Generative Modeling, deriving the MeanFlow Identity connecting finite-interval average velocity
u(z_t, r, t)to instantaneous flow velocity via Jacobian-vector products (JVP). Regressing directly onto average displacement lets single-step (1-NFE) generators train from scratch without a pretrained teacher or Distillation & Synthetic Data schedule.2026
PaperLu et al. introduced Pixel MeanFlow (pMF), bypassing compressed VAE tokenizers by predicting clean manifold points
xwhile optimizing an average-velocity objective parameterized in(z_t - x_pred) / tform—resolving high-dimensional velocity collapse to achieve one-step, latent-free image generation directly in pixel space.- Paper
Chefer, Esser, Rombach, and colleagues at Black Forest Labs published Self-Flow, applying Dual-Timestep Scheduling so tokens within the same sample receive heterogeneous noise levels. The resulting information asymmetry forces the generator to learn semantic visual, video, and audio representations internally without external teacher encoders, powering FLUX 3.
See also (15)
Open full graphFlow Matching
10 more connections
- WanOutgoing · EnablesIncoming · Uses
- Diffusion Transformer (DiT)Outgoing · Enables
- 3D GenerationIncoming · Uses
- Stable DiffusionOutgoing · MentionsIncoming · Uses
- Video Generation Models (Sora, Veo, Kling, Wan)Incoming · Uses
- Audio & Speech ModelsIncoming · Mentions
- Black Forest LabsOutgoing · Mentions
- Distillation & Synthetic DataOutgoing · Mentions
- Latent Space & VAEOutgoing · Mentions
- TRELLISOutgoing · Mentions