Flow matching: why straight interpolation does not imply one-step generation

Technical monograph · Continuous generative models · 21 September 2026

Abstract

Flow matching can construct training examples by interpolating along straight lines between noise and data. This simplicity does not imply that the generator follows straight trajectories or that one numerical step suffices to obtain the desired distribution. We examine the transition from conditional velocity to its conditional mean and develop a fully solvable Gaussian case. The exact field transforms N(0,1) into N(3,4), yet one Euler step collapses every sample to the point 3. We compare Euler and Heun under a declared field-evaluation budget and separate learning error, probability paths, and integration. The result is a verifiable example of how a generator can have the correct field while still producing the wrong distribution.

Prerequisites and contribution. Derivatives, conditional probability, and basic Gaussian distributions are required. This is a pedagogical derivation with executed calculations, not a new generative method or a reproduction of an image benchmark. We train no network: we know the exact field and study only transport and numerical sampling.

1. What is learned, and what is integrated?

A continuous generative model starts with a simple variable X₀, such as Gaussian noise, and transforms it into a variable whose distribution resembles the data. Time t belongs to [0,1]; it is not the physical time of the depicted scene. A field v(t,x) assigns a velocity to every position x. The generated sample is the endpoint of a trajectory satisfying dX/dt = v(t,X).

This description already contains two different objects: the sample distribution at each time and the path followed by each sample. A distribution can be transported by different fields. Specifying where probability mass must go does not uniquely determine every trajectory. In multiple dimensions, some motions can change trajectories without changing a particular density.

The foundational flow-matching paper formulates learning as velocity-field regression and provides tractable conditional objectives. We focus on the distinction between conditional and marginal fields made explicit in that method. Our analysis develops a scalar case chosen to make every step calculable. [1, Sections 3–4]

A terminological shortcut should be avoided: “simulation-free training” means constructing a training example does not require integrating the model’s ODE. It does not mean generation requires no integration. Producing a new sample still requires solving a dynamical problem, either exactly or approximately.

2. From sample pairs to velocity regression

Choose X₀ from the initial distribution and X₁ from the target distribution. For now, sample them independently. Define an interpolation and its derivative:

Time is sampled uniformly. Target U is available because both endpoints of the pair are known. The network, however, receives only t and X_t, not the hidden endpoints. Different pairs can yield the same intermediate position with different velocities. Reconstructing every individual velocity exactly can therefore be impossible even with unlimited capacity.

The minimizer of squared loss is the conditional mean. At a fixed time and position, write U = E[U|X_t] + R, where residual R has zero conditional mean. Expanding the square makes the cross term vanish. We obtain:

The second term is independent of network parameters. A positive training loss can therefore coexist with a perfect marginal field. Comparing absolute loss values across different paths while ignoring this irreducible component can lead to incorrect conclusions about model quality.

Why does the conditional mean transport the right distribution? Take a smooth test function f. The chain rule gives dE[f(X_t)]/dt = E[∇f(X_t)·U]. Conditioning on X_t allows U to be replaced with v*. This is the weak form of the continuity equation associated with the marginal field. The argument requires sufficient regularity and integrability to exchange differentiation and expectation; the regression formula alone does not justify ignoring these assumptions.

3. A Gaussian case we can solve exactly

Consider independent X₀ ∼ N(0,1) and X₁ ∼ N(μ,σ²). The interpolation is also Gaussian. Its mean is m(t) = tμ and its variance is D(t) = (1−t)² + t²σ². For velocity U, E[U] = μ and Cov(U,X_t) = tσ²−(1−t).

The conditional mean of jointly Gaussian variables is affine. Substituting the quantities just calculated yields:

For σ > 0, the denominator remains positive throughout the interval. The marginal field in this example has no endpoint singularity. Dependence on x is linear, but its coefficient changes with time. “Linear in the state” and “a straight trajectory through time” are different properties.

The ODE solution starting from value z is explicit:

To verify it, note that D′(t)=2c(t). Differentiating the solution gives μ + c(t)z/√D(t). Evaluating (3) at x=φ(t,z) produces the same expression. If z is standard normal, the solution has exactly mean tμ and variance D(t). We have checked both dynamics and distribution without relying on how a point cloud looks.

Our calculation uses μ=3 and σ=2. Variance initially decreases: D(t)=1−2t+5t² reaches its minimum of 0.8 at t=0.2, then rises to 4. The flow first contracts and then expands. For nonzero z, the square root of D(t) makes the trajectory non-affine in time, despite the straight conditional interpolations used to construct it.

4. Pairwise straight lines are not the generator’s trajectories

The line (1−t)x₀+tx₁ preserves the identity of an initial pair. The marginal field instead aggregates velocities compatible with the current position. Its trajectories generally do not preserve the original pairing of x₀ and x₁. In our example the ODE always ends at μ+σx₀, whereas the training pair used an x₁ independent of x₀.

There is no contradiction: both constructions have the same intermediate marginal distributions but different dependencies between endpoints. A generator must produce the final distribution; it need not reproduce the random coupling used to build its targets. Confusing a distribution with a coupling makes a straightforward conditioning operation seem mysterious.

We can also change the coupling. Choose X₁=μ+σX₀ rather than sampling it independently. The interpolation becomes X_t=tμ+[1+t(σ−1)]X₀. Every trajectory is now straight, intermediate variance is [1+t(σ−1)]², and conditional velocity is unambiguous given X_t. Endpoint distributions are unchanged; the route between them differs.

This comparison clarifies the role of pair selection. In practice, we generally lack the exact transport map between noise and complex data. The ease of the Gaussian case does not automatically carry over to images. It is nevertheless enough to show that the target distribution alone does not determine the difficulty of regression or sampling.

5. Positive loss with a perfect field

Under independent coupling, the conditional variance of velocity is constant in position but changes with time. The Gaussian formula gives:

For σ=2 this is 4 initially, 5 at the minimum of D, and 1 at the end. These quantities are not network errors: they describe target ambiguity created by independent coupling. The final identity follows by expanding (1+σ²)D−c², which reduces to σ². Comparison with deterministic coupling, where this variance is zero, makes clear why loss is not a universal metric that can be compared without context.

A network can still learn a useful mean from noisy targets. Equation (2) explains why: minimizing conditional loss reduces deviation from the marginal field even when an irreducible component remains. This is a statistical regression phenomenon; it does not require every training example to be exactly reconstructible.

6. Executed experiment: one step can destroy a distribution

We integrate the exact field (3) using explicit Euler and Heun with uniform steps. Euler uses one field evaluation per step; Heun uses two. With h=1/n, their respective updates are:

For n=1, Euler evaluates v(0,z)=3−z. Its output is therefore z+(3−z)=3 for every z. Noise disappears and the result is a point mass, although the target is N(3,4). The mean is correct, but variance is zero. A mean-only check would declare a completely incorrect sampler successful.

Because both the field and integrators are affine in the state, numerical output remains an affine transformation of z. Its mean and standard deviation can be determined exactly by integrating z=0 and z=1; Monte Carlo sampling is unnecessary. We measure the order-2 Wasserstein distance between univariate Gaussians: W₂²=(m−3)²+(s−2)², where m and s are the numerical mean and standard deviation. This includes the degenerate case s=0.

The calculation was executed with Python 3.14.0 using only the standard library. No seed is needed because operations are deterministic. A finite-difference check of the exact solution’s derivative returned a maximum discrepancy of approximately 6.07×10⁻¹⁰. Complete results include intermediate step counts and accompany the code.

At 16 evaluations, Heun reduces error to approximately 0.01915 compared with Euler’s 0.17727. This is a result for this smooth problem, not a guarantee that Heun is always more efficient in neural generators. Here field evaluation is trivial; a real model introduces batch size, precision, architecture, memory, and per-evaluation cost. NFE is a useful structural measure, not a stopwatch.

7. Separate three errors before choosing a generator

The first error is statistical: finite data and model capacity can prevent learning v*. The second concerns the objective and path: the final distribution defined by the construction may be a regularized version of the data, or coupling may make the field difficult. The third is numerical: even a perfect field produces incorrect samples if the solver is too coarse. Our experiment eliminates the first error and isolates the third.

Lower loss does not by itself identify which error has decreased. More steps cannot systematically fix an incorrect field; they integrate the wrong dynamics more accurately. Conversely, training a better network may produce invisible benefits when the sampler dominates final error. Model, path, and integration must therefore be varied separately.

An informative experiment first fixes the field and studies numerical convergence, then fixes a sufficiently accurate solver and compares learned fields. For images, distributions and quality do not reduce to mean and variance; our analytical W₂ does not replace domain-appropriate metrics or assessment. The pedagogical advantage is precisely that exact ground truth is available to check every step.

The 2025 MIT notes develop the relationship between probability paths, marginal fields, and learning, providing a reference for extending this analysis. We do not compare recent generative architectures or declare a state-of-the-art winner: we study a mechanism needed to interpret them. [2, Sections 2–4]

8. Reproducibility materials and conclusion

The attached program implements D(t), the field, the exact solution, and both integrators. Run python experiment.py --out results.json. Plotting requires Matplotlib; instructions and versions are in the archive. No model weights, datasets, or external services are required.

def variance(t):
    return (1-t)**2 + 4*t*t

def velocity(t, x):
    return 3 + (-1+5*t)/variance(t) * (x-3*t)

def exact(t, z):
    return 3*t + math.sqrt(variance(t))*z

Download code, results, and instructions · Numerical results in JSON.

The conclusion is precise: making conditional examples straight does not suffice to make the marginal flow straight. The model learns a conditional mean, and the generator integrates that field. Intermediate distributions, pair coupling, regression ambiguity, and numerical accuracy are distinct levels. In our case the field is exact and the final mean is correct, yet one step destroys all desired variability. Checking these assumptions, beyond the apparent simplicity of interpolation, is what makes rigorous reasoning about generative models possible.

References

1. Lipman, Y., Chen, R. T. Q., Ben-Hamu, H., Nickel, M., Le, M. Flow Matching for Generative Modeling. arXiv:2210.02747v2, 2023. Methods: Sections 3–4; experimental protocol: Section 6.
2. Holderrieth, P., Erives, E. An Introduction to Flow Matching and Diffusion Models. MIT 6.S184 notes, arXiv:2506.02070v1, 2025. Paths, fields, and regression: Sections 2–4.

Analysis prepared with AI assistance. Synthetic educational experiment, not EL-AI corporate research. Illustrative AI-generated cover; the scientific figure comes from the attached calculations.