Flow Matching — Straightening the Path from Noise to Image

Diffusion arrived at "follow a velocity field from noise to data" the hard way — through Markov chains, variational bounds, and reverse-time SDEs. Flow matching asks the question that framing begs: if generation is just transporting noise to data along a path, why inherit a curved path from a noising process you never actually cared about? Why not simply choose the path — and choose it straight? The answer now trains the frontier: Stable Diffusion 3, Flux, and much of the current video generation stack are flow-matching models.

By Pratik Dhanave · November 08, 2026 · 5 min read

#Flow Matching #Diffusion Models #Rectified Flow #Generative AI #Image Generation

This is post 9 of Diffusion Models from First Principles. Post 4 ended with the probability flow ODE and a promise; post 5 hit a ~10-step floor caused by path curvature. This post cashes both in. The spirit is the one that recurs across this blog’s systems series: when a derivation gets simpler the second time a field writes it, the simpler form wins — and gets adopted.

Generation as transport

Strip away diffusion’s history and the task that remains: move samples from an easy distribution (Gaussian static) to the data distribution. Any smooth velocity field — an assignment of “which way and how fast” to every point in space at every time along the way — defines such a transport: drop a noise sample in, integrate the ODE, read out where it lands. Post 4 showed diffusion implicitly builds one such field (the probability flow ODE, with velocity derived from the learned score). Flow matching makes the field the primary object: pick the paths you want, then regress a network onto their velocities.

The construction is disarmingly direct. Pair each training image with a random noise sample. Connect them with a path — and take the simplest possible choice, the straight line: at time t, the point is just the appropriately-weighted blend of noise and image. The velocity along that path is constant: image minus noise. Training: sample an image, a noise draw, and a random time; form the blend; train the network to predict the difference vector, with mean squared error.

That is the whole recipe — compare it to post 3. No noise schedule to design (the straight-line blend is the schedule, and a particularly clean one: this is the “signal-to-noise ratio falling smoothly” framing post 2 alluded to, made literal). No variational bound; the thing regressed is the thing sampling uses. DDPM’s noise-prediction target and the flow velocity are, once again, linear relatives of each other — flow matching is not a different kind of model so much as diffusion with the cleanest possible parameterization. The frameworks converged: modern codebases implement one noising-blend + velocity-regression core and recover classic diffusion as a setting.

Straight-ish, not straight — and rectification

A subtlety keeps the theory honest. Individual training pairs are connected by straight lines, but the learned field at any point must average over every (noise, image) pair whose line passes through that point — random pairing means massive crossing, and the average of crossing straight lines is a curved marginal flow. So a flow-matching model’s actual sampling trajectories are straighter than diffusion’s, empirically, but not perfectly straight; you still integrate with a solver over multiple steps.

Rectified flow turns straightening into an explicit, iterable operation. Train a first flow; generate with it, recording which noise sample produced which image; re-pair the training data using those generated couplings — each noise point now matched to the image its own flow transported it to, eliminating the crossings — and retrain on the straightened pairs (“reflow”). Each round makes trajectories straighter; genuinely straight paths can be followed in one Euler step. This is one of the two principled routes to few-step generation (post 10’s distillation is the other), and the Stable Diffusion 3 lineage trains directly on the rectified-flow formulation, with a practical refinement: sampling intermediate times non-uniformly, concentrating training where prediction is hardest (the middle of the path), measurably improves large-model results.

Why the frontier switched

The migration (SD3, Flux, and the bulk of current video systems) was driven by compounding practical wins rather than one theoretical knockout. Simplicity: fewer coupled design choices (schedules, prediction-target variants, loss weightings collapse into the path choice), which matters more as everything else — data, architecture (post 8’s DiT), parallelism — gets harder. Trainability at scale: the well-conditioned constant-velocity target trains stably in exactly the large-batch, high-resolution regimes frontier systems occupy. A cleaner route to speed: straighter paths ease the solver’s job today and define the one-step limit via reflow. The same interfaces: CFG (post 7) applies verbatim — run the velocity network conditioned and unconditioned, extrapolate; DiT backbones (post 8) slot in unchanged; post 6’s latent spaces host it all. Everything the series built survives; only the path got straighter.

The takeaway: flow matching reframes generation as transport along a chosen velocity field — pair noise with data, connect with straight lines, regress the network onto the constant velocity (image minus noise) at random blend times. Crossing paths keep the learned marginal flow merely straight-ish, so rectified flow re-pairs data with the model’s own couplings and retrains, straightening toward one-step generation. With simpler training, scale-friendly conditioning of the objective, and full compatibility with CFG, DiT, and latents, it became the frontier’s default formulation — diffusion, rewritten the simple way.

Key takeaways

Generation = transport: a velocity field carrying Gaussian noise to data; diffusion’s probability flow ODE (post 4) was one such field, derived the hard way. Flow matching chooses the field directly.
The recipe: blend noise and image at a random time along a straight line; regress the network onto the constant velocity image − noise with MSE. Schedule, prediction target, and loss weighting collapse into one path choice — and classic diffusion reappears as a special setting.
Random pairing makes lines cross, so the learned marginal flow is curved anyway — straighter than diffusion’s, still multi-step.
Rectified flow / reflow: re-pair noise→image using the model’s own generations and retrain; each round straightens trajectories, approaching one-step sampling — the SD3-lineage training formulation (with hardest-in-the-middle time sampling as a scale refinement).
Adoption logic: fewer knobs, stable large-scale training, a principled speed path — while CFG, DiT backbones, and latent spaces carry over unchanged.