The paper in four sentences
Picture generation as moving a cloud of random points (noise) until it takes the shape of the data, like wind blowing sand into the shape of dunes.
A neural network learns the "wind": at every place and moment, which direction and how fast a point should move. Generating means releasing a noise point and letting it drift.
The trick that makes training easy: pair a random noise point with a random data point, draw a path between them, and teach the network the speed along that path. Averaged over many pairs, it learns the right wind for the whole cloud.
Choosing straight-line paths, rather than the curved paths diffusion implies, gives a simpler wind that needs fewer steps to follow. Recent image models and the robot model π0 use this.
Idea · §1–2
Generating as moving a cloud of points
Every generative model has to turn something simple and random into something that looks like real data. Diffusion models do it by learning to remove noise step by step. There is another way to see the same job: as transport. Start with a cloud of random points, then move every point smoothly over a fixed stretch of time until the cloud has the shape of the data.
A rule that says, for every location and every moment, which way and how fast to move is called a velocity field (or vector field). If a neural network can represent that field, generating a sample is just following it: start at a random point and take small steps in the direction the network says. Models built this way are called continuous normalising flows. They were known before this paper, but they were slow and awkward to train, because training required simulating the whole journey.
TakeawayGeneration can be seen as moving noise into data along a learned velocity field; the challenge is training that field efficiently.
Idea · §4.1 · Figs. 2–4
Straight paths beat curved ones
Many different velocity fields can carry the noise cloud to the data. The paper's flexible framework lets you choose the paths. One choice reproduces the curved paths that diffusion models follow. Another, which the paper calls the optimal transport path, moves each point in a straight line at constant speed.
Straight paths matter because generating means following the field in a limited number of steps. On a curved path, each straight step overshoots the bend; on a straight path, even one big step lands close to where it should. Below, both kinds of flow carry the same noise to the same three data clusters, using the exact velocity field for each. Lower the number of steps and see which holds up.
TakeawayStraight-line paths make the velocity field simple, so it can be followed accurately in few steps.
Check yourself
- Why does a curved route need more steps to follow?
- Each step moves in a straight line in the current direction. On a curve the direction keeps changing, so big steps cut corners and go off course; on a straight route the direction doesn't change.
Idea · §3 · Theorems 1–2
The training trick: learn from single pairs
Here is the problem. To train the network you would want to show it the correct overall velocity at each place and time, but nobody knows that field; it is exactly what we're trying to find.
The paper's insight: pick one noise point and one data point, join them with a simple path, and use the velocity along that path as the training target. For a straight path the target is simply "data point minus noise point". Each individual target is "wrong", since a point in the middle of the cloud could be heading to many places. But the network, trained with a squared-error loss over millions of random pairs, learns the average velocity at each place and time. The paper proves that this average is exactly the field that moves the whole cloud correctly.
In the figure, time runs left to right. Grey lines are individual training pairs: they cross each other freely. The amber curves are what the learned, averaged field actually does: they never cross, and each noise point is routed smoothly to one of the two data clusters.
TakeawayTrain on simple single-pair targets; the squared-error loss averages them into the correct velocity field for the whole distribution.
Check yourself
- Why can the learned paths not cross, even though the training pairs do?
- The learned field gives one velocity per place and time. Two paths meeting at the same place and time would have to continue identically, so they cannot cross.
- What is the training target for a straight path from noise point A to data point B?
- The constant velocity B − A (with a tiny correction in the paper for a minimum noise level).
Idea · §6 · Tables 1–2
How well it worked
The authors trained one U-Net architecture with different objectives on CIFAR-10 and downsampled ImageNet (32, 64 and 128 pixels). Flow matching with straight paths gave the best likelihood and image-quality scores among their trained baselines, and generated images with fewer network evaluations: on ImageNet 64 × 64, an FID of 14.45 using about 138 evaluations, against 17.36 and about 264 for the DDPM baseline. It also trained faster and handled 4× super-resolution well.
Even training the network with diffusion-shaped paths, but via flow matching, worked better than the usual score-matching loss, showing the training method itself is robust.
TakeawayFlow matching with straight paths matched or beat diffusion training in quality while sampling with fewer steps.
Idea · §5, §7 and legacy
A simpler alternative to diffusion
Flow matching does not throw diffusion away; it contains it. Diffusion paths are one choice inside the framework, and the training loss is a cousin of diffusion's noise-prediction loss. What it adds is freedom to choose the path, and a clean, simulation-free way to train any of them.
The paper was one of several in 2022 that reached closely related ideas at the same time, notably "rectified flow" (Liu and colleagues) and "stochastic interpolants" (Albergo and Vanden-Eijnden).
TakeawayFlow matching reframed diffusion as learning velocities along chosen paths, and straight paths became the default for many state-of-the-art generators.
Abstract, restated
A new paradigm for generative modelling built on continuous normalising flows (CNFs), training CNFs at unprecedented scale. Flow Matching (FM) is a simulation-free approach that regresses vector fields of fixed conditional probability paths. It is compatible with a general family of Gaussian probability paths that includes diffusion paths as a special case; using FM with diffusion paths gives a more robust and stable alternative for training diffusion models. FM also opens the door to non-diffusion paths, such as optimal transport (OT) displacement interpolation, which are more efficient than diffusion paths, train and sample faster, and generalise better. On ImageNet, FM-trained CNFs beat diffusion-based alternatives in likelihood and sample quality, and allow fast, reliable sampling with off-the-shelf ODE solvers. ICLR 2023; arXiv:2210.02747.
Section · §2 · Eqs. 1–4
Continuous normalising flows
A time-dependent vector field vt : [0, 1] × ℝd → ℝd defines a flow φt by the ODE
and pushes a simple prior p0 forward to a probability path pt = [φt]*p0 via the change-of-variables formula (Eqs. 3–4). A vector field ut generates pt if its flow pushes p0 to pt; equivalently, the continuity equation ∂tpt + div(ptut) = 0 holds (App. B, Eq. 26). Convention: t = 0 is noise (p0 = N(0, I)) and t = 1 is data, the reverse of the DDPM time direction. Earlier CNF training by maximum likelihood needed ODE simulation through the network during training, which limited scale.
Section · §3 · Eqs. 5–9 · Theorems 1–2
Flow Matching and Conditional Flow Matching
Given a target path pt with p1 ≈ q and a generating field ut, Flow Matching regresses
which is intractable because neither pt nor ut is known. Build them from per-sample conditional paths pt(x | x1) with p0(x | x1) = p(x) and p1(x | x1) concentrated at x1:
Theorem 1: if each ut(· | x1) generates pt(· | x1), the marginal field (8) generates the marginal path (6). So ut(x) = E[ut(x | x1) | xt = x], a posterior average. Then define
Theorem 2: LFM and LCFM differ by a constant independent of θ, so ∇θLFM = ∇θLCFM. The proof is the usual fact that the minimiser of a squared loss is the conditional expectation, here of ut(x | x1) given x. Training needs only samples of x1 and of the conditional path; no ODE is simulated.
Check yourself
- Why is LCFM not zero at its optimum?
- The conditional targets for a given (x, t) vary with the unknown x1; the network can only output their mean, so the residual variance remains. That variance is the constant separating LCFM from LFM.
Section · §4 · Eqs. 10–23 · Theorem 3 · Figs. 2–3, 7
Gaussian conditional paths: diffusion vs optimal transport
Take pt(x | x1) = N(x | μt(x1), σt(x1)2I) (Eq. 10) with μ0 = 0, σ0 = 1, μ1 = x1, σ1 = σmin. The simplest flow generating it is ψt(x) = σt(x1)x + μt(x1) (Eq. 11), and Theorem 3 gives its field:
Reparameterising x = ψt(x0) with x0 ~ N(0, I) gives LCFM = E‖vt(ψt(x0)) − (d/dt)ψt(x0)‖2 (Eq. 14).
Example I, diffusion paths. Variance-exploding: pt(x | x1) = N(x1, σ1−t2I) (Eq. 16). Variance-preserving: N(α1−tx1, (1 − α1−t2)I) with αs = e−½T(s), T(s) = ∫0sβ(r)dr (Eq. 18). Their fields (Eqs. 17, 19) coincide with the probability-flow ODE of Song et al. (2021); these paths only approach noise asymptotically, not at finite time.
Example II, optimal-transport paths. μt(x) = tx1, σt(x) = 1 − (1 − σmin)t (Eq. 20), so
The conditional flow is the OT displacement map between the two Gaussians p0(x | x1) and p1(x | x1) (McCann's interpolation). Its conditional field has constant direction in time (Fig. 2) and its conditional trajectories are straight, whereas diffusion conditional trajectories can overshoot and backtrack (Fig. 3). The authors caution that this "by no means" implies the marginal field is an OT solution.
The figure measures the practical consequence on the 2-D toy above: Euler sampling with N function evaluations against a 2,000-step reference, for the exact marginal fields of both paths (VP with β from 0.1 to 20, as in Song et al.; both integrated to t = 0.999).
Section · §6 · Tables 1–2 · Figs. 5–7
Experiments
The authors trained FM-OT, FM with diffusion paths and score matching with diffusion paths using an identical U-Net (from Dhariwal and Nichol, with minimal changes); likelihood via the instantaneous change of variables and sampling with the adaptive dopri5 solver at tolerance 10−5; NFE is the number of function evaluations it used.
| Table 1 | CIFAR-10 NLL / FID / NFE | ImageNet 32 NLL / FID / NFE | ImageNet 64 NLL / FID / NFE |
|---|---|---|---|
| DDPM | 3.12 / 7.48 / 274 | 3.54 / 6.99 / 262 | 3.32 / 17.36 / 264 |
| Score matching | 3.16 / 19.94 / 242 | 3.56 / 5.68 / 178 | 3.40 / 19.74 / 441 |
| ScoreFlow | 3.09 / 20.78 / 428 | 3.55 / 14.14 / 195 | 3.36 / 24.95 / 601 |
| FM w/ diffusion paths | 3.10 / 8.06 / 183 | 3.54 / 6.37 / 193 | 3.33 / 16.88 / 187 |
| FM w/ OT paths | 2.99 / 6.35 / 142 | 3.53 / 5.02 / 122 | 3.31 / 14.45 / 138 |
NLL in bits per dimension. On ImageNet 128 × 128, FM-OT reaches NLL 2.90 and FID 20.9, which the authors call state of the art apart from IC-GAN. They note the CIFAR-10 FIDs are higher than usual, perhaps because the architecture was not tuned for CIFAR-10. FM-OT also lowers FID faster during training (Fig. 5), and its sampling cost stays roughly constant over training, unlike score matching (Fig. 10). Super-resolution from 64 to 256 pixels (Table 2): FM-OT FID 3.4, IS 200.8, against SR3's 5.2 and 180.1, with slightly lower PSNR (24.7 vs 26.4) and SSIM (0.747 vs 0.762).
Check yourself
- "FM w/ diffusion" and DDPM use the same probability path. What, then, does their comparison isolate?
- The training objective and sampler: regressing the path's vector field (and sampling with an ODE solver) versus DDPM's noise-prediction loss and ancestral sampling. FM with diffusion paths is comparable in quality and uses fewer NFEs.
Section · §5, §7 and lineage
Limitations and lineage
- Marginal paths still curve. Only conditional paths are straight. Where conditional paths from different x1 cross, the averaged field bends, so few-step sampling remains approximate.
- Gaussian, isotropic paths only. The paper restricts itself to isotropic Gaussian conditional paths and lists richer paths as future work.
- Evaluation. Experiments are at modest resolutions with a fixed architecture; there is no dedicated limitations section.
Reference
Glossary
- CNF
- Continuous normalising flow: a generative model defined by integrating a learned vector field from a simple prior.
- Probability path pt
- A time-indexed family of densities from noise (t = 0) to data (t = 1).
- FM / CFM
- Flow Matching (regress the marginal field, intractable) / Conditional Flow Matching (regress per-sample fields, same gradients).
- OT path
- Conditional path with linear mean and standard deviation in t; straight conditional trajectories.
- σmin
- Small standard deviation of the conditional path at t = 1.
- NFE
- Number of function (network) evaluations used by the ODE solver per sample.
- dopri5
- Dormand–Prince adaptive Runge–Kutta ODE solver.
- Rectified flow
- Concurrent formulation with the same straight interpolation (σmin = 0), plus an optional "reflow" step.