Source
Pooladian, Ben-Hamu, Domingo-Enrich, Amos, Lipman and Chen (2023), LIT-789, §3–4, Lemma 3.2, Eqs. 18–19 and Thm. 4.2. Tong, Fatras, Malkin, Huguet, Zhang, Rector-Brooks, Wolf and Bengio (2023), LIT-798, §3.2 and the proof of Prop. 3.4. Liu, Gong and Liu (2022), LIT-636, Thms. 3.3–3.7. Lee, Lin and Fanti (2024), LIT-791, §2–3. Liu, Zhang, Ma, Peng and Liu (2023), LIT-796, §2.2 and Fig. 6.
The account
The learned velocity is an average. Flow matching draws a pair (x₀, x₁), places a point on the straight line between them, and regresses the network onto that line’s velocity x₁ − x₀. The minimizer at (x, t) is the conditional expectation of x₁ − x₀ over all pairs whose line passes through x at time t. Lee et al. (LIT-791, Eq. 1) write the same field through the posterior mean of the data endpoint given x_t.
Crossings are what make it curve. If two lines with different directions pass through the same (x, t), the field there is neither, and a particle integrating it turns. Multisample Flow Matching (LIT-789) makes this exact in two places. The joint objective’s optimal value is the spread of the conditional velocities about their mean, and it bounds the gradient variance at fixed (x, t) (Lemma 3.2). Under independent pairing it “generally cannot achieve value zero … since there are an infinite number of pairs (x0, x1) whose conditional path crosses any particular x at a time t”. Its straightness measure S (Eq. 18) can be rewritten as the variance of the velocity along a trajectory (Eq. 19). It is zero exactly when every trajectory is a straight line. Lee et al.’s Eq. 4 is the same fact read as a loss: the training error has a floor, the posterior variance of the endpoint given x_t, which no network can reduce.
A coupling without crossings gives a straight flow. Two constructions remove the crossings while keeping the endpoint distributions.
- Optimal transport. OT-CFM (LIT-798) proves that with the exact OT plan and vanishing path width the learned field solves the dynamic OT problem (Prop. 3.4). Its proof turns on the fact that paths from an OT map, the gradient of a convex function, cannot cross. So the field at each point is that point’s own line, T(x) − x. Multisample Flow Matching proves the minibatch version in the limit: as the batch size k grows, the objective’s optimum and S go to zero and the transport cost goes to W₂² (Thm. 4.2).
- Reflow. Rectified Flow (LIT-636) pairs each noise with the point the current flow carries it to and retrains. Marginals are preserved (Thm. 3.3), no convex transport cost increases (Thm. 3.5), and straightness improves at rate O(1/K) over K rounds (Thm. 3.7). All three are about exact minimizers. InstaFlow (LIT-796, §2.2) restates the three properties for its text-conditioned reflow.
Where this sits against THEORY-106. That account says a straight conditional path does not make the marginal ODE straight, and that “straight against curved path” comparisons change a weighting, a network output and a sampling schedule. This one says what does decide the marginal’s shape when the conditional path is held straight: the coupling. OT-CFM states the premise in §3.2.1: the conditional path is OT, “the marginal path pt(x) is not in general an OT path”.
What was measured
- In 2-D the effect is an order of magnitude. OT-CFM’s normalized path energy is 0.018–0.087 with minibatch OT against 0.222–2.738 with independent pairs, over five seeds. A 2-rectified flow, one reflow on the same tasks, reaches 0.069–0.149 (LIT-798, Table 2).
- On ImageNet at 32 and 64 px it saves Euler steps. In the controlled comparison of Multisample Flow Matching (LIT-789, Table 1, read off Fig. 3), FID 20 on ImageNet-64 takes 12 Euler steps with batch OT against 29 with independent pairing. Likelihood is unchanged. Samples from the same noise move less as the step count drops (Table 3).
- On a diffusion teacher, reflow measurably straightens. InstaFlow (LIT-796, Fig. 6) estimates S on Stable Diffusion’s probability-flow ODE and on the reflowed model, and S falls.
What it explains
SOTA-444’s mechanism. That practice pairs noise and data by exact minibatch OT before the ordinary regression. This account is why that can help and why it cannot hurt the target. The marginals are untouched, and fewer crossings mean a straighter field and a smaller error per coarse step. It also accounts for the practice’s limit. The OT-CFM proof is about the exact plan, and the minibatch plan’s crossings vanish only as k grows (Thm. 4.2). Nothing bounds them at image dimension and realistic batch size, which is the practice’s own Condition.
Why reflow helps a one-step student, and when it should matter more. SOTA-392 says to distil a rectified flow and treat reflow as optional, and InstaFlow contests it from a diffusion teacher. On this account a one-step distiller must reproduce in one step a map the teacher reaches along a curved path, and reflow replaces that map with one whose straight-line pairs barely cross. InstaFlow puts it as a suggestion: “the new coupling might be easier for the student network to learn” (§2.2). Its matched comparison fits that reading: from Stable Diffusion, reflow first gives one-step FID-5k 31.0 against 40.9. So does Rectified Flow’s own table from a straight-path teacher: 4.85 against 6.18. The account predicts a larger increment from a more curved teacher. That is the direction of the two results, but they differ in teacher, scale, data and distiller, so it is a prediction and not a finding.
Where it stops
- It does not say reflow is free at many steps. Rectified Flow’s own 2- and 3-rectified flows lose full-simulation quality, 2.58 to 3.36 and 3.96 FID. The theorems are about exact minimizers, and each round trains on generated pairs with its own error. SOTA-392’s last clause, keep the pre-reflow model for many steps, rests on that measurement, not on this account.
- “One reflow is enough” is an argument, not a result. Lee et al. (LIT-791, §3) argue that for two 1-rectified-flow pairs to cross, one noise must equal another plus a scaled data difference. Such a noise sits off the Gaussian annulus and is autocorrelated, so the optimal 2-rectified flow is nearly straight. It is the 2-rectified flow, after one reflow, that they call nearly straight, not the base model. The evidence is qualitative (Fig. 2), and the checklist says the claim “is intuitive, and we do not prove it”. Their follow-on choice of a U-shaped timestep density reads the training loss as the reducible error, which is sound only if the floor above is near zero.
- A straighter marginal is not what path comparisons measure. On CIFAR-10, under one recipe, changing the coupling moves 100-step Euler FID from 4.461 to 4.443, and changing the conditional path from VP to straight moves it from 7.772 to 4.640 (LIT-798, Table 5). So the gain SOTA-266 records for the straight path is not, on this evidence, a gain from a straighter marginal ODE. That agrees with THEORY-106.
- It does not make a few-step sampler. Straightening lowers the step count of an ODE solve. Four-step Euler FID with batch OT is still 38.86 on ImageNet-32 (LIT-789, Tables 7–8).
- Unconditional only. No paper here pairs under a class or text condition. Pairing across a batch makes a noise depend on the other images, and whether the conditional marginals survive is not addressed.
- Straight is necessary for one-step accuracy, not sufficient for quality. Rectified Flow’s reflowed model at one undistilled step scores 12.21 FID, against 4.85 once distilled. Distillation does work that straightening does not.