Beckmann Transport Models: From Autonomous Flows to One-Step Maps

Eric Vanden-Eijnden; Florentin Coeurdoux; Lee Cheuk-Kit; Michael Samuel Albergo; Peter Potaptchik; Yilun Du

REVIEW 4 major objections 6 minor 13 references

Beckmann Transport Models: From Autonomous Flows to One-Step Maps

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A time-independent drift transports a base distribution onto a lower-dimensional target manifold, and the one-step map is the unique solution of a conservation equation.

desk verdict A genuinely interesting framework with a real proof gap: the central transport theorem rests on an unproved assertion that divergence-free perturbations of the drift don't change the basins of attraction. read the letter →

arxiv 2608.01692 v2 pith:PIFJOQGN submitted 2026-08-03 cs.LG

Lee Cheuk-Kit , Florentin Coeurdoux , Peter Potaptchik , Yilun Du , Michael Samuel Albergo , Eric Vanden-Eijnden This is my paper

classification cs.LG MSC 49Q2235Q4968T07

keywords autonomousflowmatchingdivergenceequationBeckmanntransportproblemone-stepgenerativemapconservationequilibriumPoisson

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper proves that a velocity field learned by flow matching remains a valid transport even when it is frozen in time, provided the target distribution lives on a lower-dimensional manifold. The time-independent drift obeys a divergence equation that encodes the base distribution as a source and the target manifold as a sink, so every trajectory of the autonomous flow converges to the manifold and the endpoint map pushes the base to the target. The same equation is the flux constraint of Beckmann's transportation problem, which gives a dynamic interpretation that was previously missing. As a corollary, the one-step transport map is the unique solution of a conservation equation, which can be learned directly from samples and used for single-pass generation. The construction recovers Poisson-flow generative models and corrects a mass-allocation bias in the original equilibrium-matching loss.

What carries the argument

The machinery is the pair formed by the divergence equation $\nabla \cdot (\nu b) = \mu_0 - \mu_1$ and the conservation equation $b \cdot \nabla T = 0$ with boundary $T = \mathrm{id}$ on $M_1$. The divergence equation makes the positive weight $\nu$ a design parameter: $\nu$ equal to the time-averaged interpolant law gives the flow-matching drift, while $\nu \equiv 1$ recovers the Coulomb field of Poisson-flow models. The conservation equation states that the endpoint of an autonomous trajectory is constant along that trajectory, which uniquely characterizes $T$ and, after discretizing the Eulerian flow equation, supplies the direct map-learning objective with stop-gradient updates.

What would settle it

Simulate a two-atom target in the plane, add a divergence-free swirl to a gradient drift that satisfies the divergence equation, and count which atom each initial condition reaches; if any positive-measure set of trajectories switches basins, the pushforward masses will not equal the target atom weights.

Watch

Extended reading notes

Core claim

The central claim is that the minimizer $b$ of the time-independent flow-matching regression loss satisfies $\nabla \cdot (\nu b) = \mu_0 - \mu_1$ with $\nu$ the time-averaged interpolant density, and this stationary current has $\mu_0$ as source and the singular target support $M_1$ as sink. Consequently the autonomous ODE $\dot{X}_t = b(X_t)$ carries almost every initial point into $M_1$, and the first-hitting limit $T(x_0) = \lim_{t \uparrow \tau} X_t(x_0)$ defines a map with $T_\sharp \mu_0 = \mu_1$; $T$ is the unique continuous solution, up to $\mu_0$-null sets, of $b \cdot \nabla T = 0$ with $T = \mathrm{id}$ on $M_1$. The map can be learned through a residual objective derived from the Eulerian form of the flow, giving one-step generation that improves when composed iteratively.

Load-bearing premise

The argument assumes that the swirling, divergence-free part of the learned drift cannot move any trajectory across a basin boundary, so the endpoint map is fully determined by the gradient part; if that fails, the autonomous flow may not push the base distribution to the target.

Editorial extensions

If this is right

The autonomous flow from the time-independent flow-matching drift reaches the singular target, so generation can use a fully frozen drift and trajectory-dependent hitting times instead of a time-dependent field.
The endpoint map $T$ can be trained directly from flow-matching samples with one-step inference, and iterating a partially trained $T$ sharpens the result because each application advances the trajectory along the autonomous flow.
Equilibrium matching's original loss is biased whenever its schedule is not identically one; using an interpolant-consistent target removes that bias at no extra training cost.
Poisson-flow generative models arise as the $\nu \equiv 1$ instance of the same divergence-equation construction, unifying them with equilibrium matching.
The autonomous action $\int |b|^2 \nu$ lies between the squared Wasserstein cost and the time-dependent flow-matching action, so freezing the drift cannot increase the transport cost.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The framework suggests testing whether optimizing over the pair $(\nu, b)$ subject to the divergence equation can tighten the upper bound on the squared Wasserstein cost beyond the flow-matching-derived pair.
Because atomic targets are the $k=0$ instance of the singular-support assumption, the conservation equation applies directly to discrete and token-based generation once tokens are embedded as vertices of a simplex.
A testable consequence is that the corrected loss should improve mass allocation most on classes with non-uniform frequencies, since the ImageNet gain is small where classes are near-uniform.
The iteration count at inference is a dial on effective flow time, so a single learned map can trade a few extra forward passes for the training burden of full convergence.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The paper has a good idea and a real gap. The identification of the time-independent flow-matching minimizer with the divergence equation ∇·(νb)=µ0−µ1 is new and clean, and the recovery of PFGM as the ν=1 case plus the correction to Equilibrium Matching's loss are genuine contributions. The basin-weight experiments make the EM bias concrete, and the direct map-learning objective is worth taking seriously.

The problem is Proposition 1, the load-bearing result. The proof in Appendix B, Step 3 writes the FM minimizer b=j/ν in a ν-weighted Helmholtz decomposition, νb=ν∇φ+q⊥ with ∇·q⊥=0, and then asserts that the basin structure of the flow ˙X=b(X) coincides with that of ˙X=∇φ(X) because q⊥/ν "generates rotations without changing the mass transport." That assertion is not proven and is not generally true. A divergence-free perturbation of a gradient field changes the non-conservative part of the ODE and can push orbits across separatrices, altering which atom of a singular target a trajectory lands on. The divergence condition (9) controls only ∇·(νb), not q⊥. So the basin mass-accounting in Theorem 1, which integrates over basins defined by the ∇φ flow, cannot be transferred to b as written. This is the bridge between the actual FM minimizer and the advertised transport guarantee, and it is missing.

The second issue is smaller but real. Theorem 2 states that the transport map T is the unique continuous solution of b·∇T=0 with T=id on M1. For atomic targets, T is piecewise constant on basins and jumps across basin boundaries; those boundaries may be µ0-null, but the jump is a genuine discontinuity. Redefining T on a null set does not make it continuous in the usual sense, so the statement is false as written. The proof only establishes uniqueness up to null sets, which is a weaker and more honest claim.

What the paper does well: the divergence-equation framework is a useful unification, the EM loss correction is convincingly demonstrated on the 2D atomic example, and the map-learning objective (Theorem 3) is a sensible residual scheme. The ImageNet gains are modest (1.87 vs 1.90 FID) and no code is released, but those are secondary.

Who this is for: people working on flow matching, equilibrium matching, and one-step generative models. It deserves a serious referee because the framework is interesting and the proof gap is potentially repairable — either by proving that the specific FM minimizer has the needed gradient structure, or by adding a condition that rules out the problematic divergence-free part. I'd recommend engaging with it, but the authors need to fix the central proof or weaken the claim before the transport theorem can be trusted.

Referee Report

4 major / 6 minor

Summary. The paper introduces Beckmann Transport Models (BTM), a generative modeling framework based on time-independent drift fields b and positive weights ν satisfying the divergence equation ∇·(νb)=µ0−µ1. The central theoretical claims are: (i) Proposition 1, that the minimizer of the time-independent flow-matching regression loss transports an absolutely continuous base µ0 to a singular target µ1 through the autonomous ODE Ẋ=b(X); (ii) Theorem 1, extending this to any gradient drift satisfying the divergence equation; (iii) Theorem 2, a conservation equation b·∇T=0 with boundary condition T=id on M1 that uniquely characterizes the one-step transport map; and (iv) Theorem 3, an objective for learning T directly from FM samples. The paper also positions BTM as a dynamical realization of Beckmann's optimal transport problem, claims to recover PFGM and a corrected Equilibrium Matching loss, and reports experiments on atomic 2D targets, a spiral dataset, and ImageNet 256×256.

Significance. If the theoretical claims were established, the paper would offer a valuable unifying perspective connecting flow matching, Poisson-flow generative models, and Beckmann transport, and the direct map-learning objective would be practically appealing for one-step generation. The paper also identifies a genuine bias in the original Equilibrium Matching loss and reports a modest but consistent FID improvement on ImageNet, with reproducible experimental details. However, the central proof of Proposition 1 contains an unproved and generally false equivalence, so the main theoretical contribution is not currently supported. The experiments, while suggestive, do not compensate for the missing foundation of the paper's core theorem.

major comments (4)

[Appendix B, Step 3 (Eq. (23))] The reduction of the FM drift b=j/ν to the gradient field ∇ϕ via the ν-weighted Helmholtz decomposition is asserted, not proved: the claim that the divergence-free remainder q⊥/ν 'generates rotations without changing the mass transport' is generally false. A divergence-free perturbation of a gradient field can change the separatrix structure and the ω-limit sets of the autonomous ODE (for example, a rotational component can create limit cycles or shift basin boundaries), so the basins BA defined by the ∇ϕ-flow in the proof of Theorem 1 cannot be used to account for the mass transported by the b-flow. Since the divergence condition (9) only controls ∇·(νb), it imposes no constraint on q⊥, and the conclusion T♯µ0=µ1 for the actual FM minimizer b is unsupported. This is the load-bearing bridge of Proposition 1.
[Appendix B, proof of Theorem 1(i)] The claim that 'any limit point would be a local maximum of ϕ along the trajectory, hence must lie on M1' is not justified. Along a gradient flow, ϕ is nondecreasing, so a limit point is a critical point, but it need not be a local maximum; a trajectory may converge to a saddle (or to another non-maximum critical point) in a set that is not obviously µ0-null. The strong maximum principle only rules out interior local maxima, not convergence to saddles. The proof therefore requires an additional argument showing that the union of such exceptional stable manifolds has µ0-measure zero, which is absent.
[Theorem 2 (Section 2.4)] Theorem 2 asserts the uniqueness of a continuous solution to (11) up to a µ0-null set. The proof assumes that every trajectory of Ẋ=b(X) converges to M1 and that the boundary condition T=id on M1 determines the value along the trajectory. For a general b satisfying only (9), convergence to M1 is not guaranteed (the Lyapunov argument of Theorem 1 applies only to gradient b), and if b is not Lipschitz at M1 the characteristic flow may be non-unique. Thus the uniqueness statement is not established outside the gradient setting, and even there it inherits the unresolved convergence issues of Theorem 1.
[Theorem 3 (Section 2.6, Eq. (15))] The stationary-point analysis of Theorem 3 is incorrect. The loss in (15) is E|T(It)−sg(T(It)+˙It·∇T(It))|² = E|˙It·∇T(It)|² (plus the boundary term). The gradient condition is E[(˙It·∇T(It)) δT(It)] = 0 for all variations δT, which is the weak form of ˙It·∇T=0 almost surely, not the conditional expectation condition E[˙It·∇T(It)|It=x]=b(x)·∇T(x)=0 stated in the proof. The claimed uniqueness of the stationary point therefore does not follow from the given argument, and the objective may have a different set of stationary points.

minor comments (6)

[Appendices F and G] Sections F and G of the appendix are nearly identical; one of them should be removed or clearly differentiated to avoid duplication.
[Appendix E] Appendix E contains unresolved cross-references to 'Section ??' in two places.
[Appendix C] Appendix C contains a typo: 'basus' should be 'basis'.
[Figure 5 caption and Section 2.6] The Figure 5 caption contains a stray 'left::', and the main text after 'Figure 2' is incomplete: 'in which we learn the autonomous map for' is not finished.
[Algorithm 2] Algorithm 2 uses the symbol bTk without definition; consider renaming it to avoid confusion with the drift b.
[Appendix B, explicit kernels] The text refers to 'equation (11) of the original note' in the explicit-kernel paragraph; this reference is not defined in the manuscript.

Circularity Check

1 steps flagged · score 6.0 of 10

Proposition 1's bridge from the FM drift to gradient transport is asserted, not derived: the divergence-free remainder q⊥ is declared mass-transport-inert, which is exactly the conclusion needed.

other [Appendix B, Proof of Proposition 1, Step 3 (Eq. (23))]
"The basin structure of the autonomous flow ˙Xt =b(X t) coincides with that of the gradient flow ˙Xt =∇ϕ(X t) on µ0-a.e. initial conditions: the divergence-free part q⊥/ν generates rotations within changing the mass transport."

Theorem 1 applies only to gradient drifts, but the actual FM minimizer b=j/ν is admitted not to be a gradient. The only bridge from b to Theorem 1 is the Helmholtz decomposition: the gradient part ∇φ satisfies the divergence condition, and q⊥ is declared to 'generate rotations without changing the mass transport.' That declaration is precisely the missing content: it says b and ∇φ have the same pushforward, i.e. that Proposition 1's conclusion T♯μ0=μ1 already holds for b. The decomposition and the divergence condition impose only ∇·q⊥=0, which does not imply dynamical inertness; a divergence-free perturbation can move trajectories across basin boundaries. The proof therefore reduces the central claim to an assertion equivalent to the claim, rather than deriving it.

full rationale

The rest of the derivation is largely self-contained and not circular. The identities ∇·(νb)=μ0−μ1 and b=j/ν are obtained by direct calculation from the time-independent FM loss; Theorem 1 is a genuine gradient-flow proof; Theorem 2's conservation equation is proven by differentiating along trajectories; Theorem 3's objective is constructed so its stationary condition is exactly that conservation equation. Recoveries of PFGM and the EM correction are instances of the divergence equation, not renamed fits. Prior work by the present authors is cited for background and comparison, not as the authority for Proposition 1. The one serious circular step is Appendix B Step 3: the paper asserts that the divergence-free remainder q⊥/ν does not change the basin structure or mass transport of the flow, and this assertion is exactly what is needed to transfer Theorem 1 from ∇φ to the actual non-gradient FM minimizer b. No equation proves it, and the asserted invariance is equivalent to the conclusion of Proposition 1. Because that step is the load-bearing bridge for the paper's central claim, the paper is partially circular despite the independent content of the gradient-case theorem and the map characterization.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The theoretical construction introduces no free parameters in the mathematical claims; the free parameters listed are experimental hyperparameters. The axioms include the unproven regularity assumptions needed for the analytic arguments, and the two ad hoc assumptions that are load-bearing for the central proof. No invented physical entities are introduced.

free parameters (4)

boundary weight λ = not specified (Algorithm 2)
Regularization weight in map objective (15), chosen by hand; affects learned map quality.
breakpoint t_c = not specified in text, used in piecewise interpolant (57)
Tunable scalar controlling bulk/tail trade-off in image experiments.
adaptive weighting p, c = p=1, c=0.01
Chosen hyperparameters in (59) for stabilization of direct-map training.
boundary loss threshold τ = τ=0.8
Applied boundary loss for times beyond 0.8 in image experiments.

assumptions (4)

domain assumption ν and φ satisfy enough regularity for the strong maximum principle and divergence theorem in the weighted elliptic operator ∇·(ν∇φ)
Used in proof of Theorem 1 (Appendix B) to conclude trajectories converge to M1.
ad hoc to paper The divergence-free part q⊥ in the ν-weighted Helmholtz decomposition of the FM minimizer does not change the basins of attraction of the autonomous flow
Assumed in Proposition 1 proof Step 3 (Appendix B) without proof; required to reduce b to a gradient flow.
ad hoc to paper The constructed endpoint map T admits a continuous representative
Theorem 2 asserts uniqueness among continuous solutions; but for atomic targets T is piecewise constant and discontinuous on basin boundaries.
ad hoc to paper Solution to ∇·(bν̄)=µ0−µ1 with vanishing at infinity is unique
Used in Appendix D to identify the autonomous flow's occupation measure with ν.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Beckmann Transport Models: From Autonomous Flows to One-Step Maps." pith.science (2026). https://pith.science/paper/PIFJOQGN

@misc{pith2026260801692,
  author       = {Pith},
  title        = {Pith review of: Beckmann Transport Models: From Autonomous Flows to One-Step Maps},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PIFJOQGN}},
  note         = {Machine review of arXiv:2608.01692}
}

read the original abstract

We propose an instantiation of flow matching that relies on a time-independent velocity field (an \emph{autonomous flow}) to exactly map between two distributions, so long as the target is singular, i.e.\ supported on a lower-dimensional data manifold. We also show that the one-step generative map associated with this flow is the unique solution of a simple conservation equation, which can be used to learn the map directly from samples. These autonomous flows and maps give a dynamical meaning to the flux constraint of Beckmann's transportation problem. Their construction provides a unifying framework that recovers, for instance, the closed-form Poisson-flow generative model and equilibrium matching with a quadratic flow-matching regression loss. We illustrate how this theory corrects inconsistencies in existing methods and demonstrate the effectiveness of the autonomous flow and the one-step map on ImageNet 256x256.

Figures

Figures reproduced from arXiv: 2608.01692 by the authors.

**Figure 2.** Autonomous transport map learning on 5-mode atomic distribution learned with [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

**Figure 3.** Basins of attraction for BTM (left) and EqM (right), schedule ct = (1 − t) 0.8 , J = 5 atoms. Each color identifies the set of initial conditions that converge to the corresponding atom; white lines are the separatrices. EqM introduces a bias because of an inconsistency in its loss that BTM corrects at no cost. Flow vs. map generation ( [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Figures from the paper (7 more)

**Figure 4.** Figure 4: Flow vs. map generation. Spiral target embedded in d = 6 with µ0 = N (0, I6). From left: target µ1; autonomous flow X˙ t = b(Xt); one-step map T 1 = T (1-NFE); three-fold iterate T 3 (3-NFE). Iteration sharpens a partially trained map; a fully trained Tθ matches the fl…

**Figure 5.** Figure 5: Left: Curated class-conditional samples from EqM-XL/2 architecture on ImageNet 256 × 256. Each image corresponds to a distinct ImageNet class. Right: : FID comparison against recent baselines, All numbers use the same evaluation protocol BTM correction improves EqM ima…

**Figure 6.** Figure 6: Comparison between the proposed Beckmann transport models (BTMs) and Drifting Deng et al. [2026]. Depending on the choice of kernel the training of drifting model can be unstable, whereas the proposed Beckmann transport model trains stably across common choices of inte…

**Figure 7.** Figure 7: Left: mass-allocation MAE vs. schedule exponent [PITH_FULL_IMAGE:figures/full_fig_p023_7.png]

**Figure 8.** Figure 8: Left: fraction of frozen particles vs. ODE time for [PITH_FULL_IMAGE:figures/full_fig_p023_8.png]

**Figure 9.** Figure 9: KDE of reference µ1 (left) and Coulomb-transported particles (right). µ0 = N (0, I5), N = 5000 particles, M = 1000 mini-batch, T = 30. No training, no score function. K.1 MNIST digit generation We consider a class-conditional MNIST digit generation task, where the tran…

**Figure 10.** Figure 10: Generated MNIST digits. From left, one-step and two-step generated results from the learnt transport map T. K.2 Latent ImageNet 256 × 256 generation We now detail the experimental setup and ablation results for our image experiments. We utilize the SiT architecture Ma…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

13 extracted references · 5 canonical work pages

[1]

The condition’s role in higher codim is to promote oblique streamlines to normal — a stronger but inessential form of regularity

In codim ≥2 it is not required: T∈C 1 for any f. The condition’s role in higher codim is to promote oblique streamlines to normal — a stronger but inessential form of regularity. G.10 Atomic case (k=d) Ty(x)M1 ={0},y(x) ∥ ≡0, hencej ∥ ≡0trivially. The codim-dformula (43) forj ⊥ applies. 21 H Self-stopping interpolant for BTM image experiments Motivated by...

work page 2025
[5]

Jonathan Ho, Ajay Jain, and Pieter Abbeel

Oral; arXiv:2505.13447. Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. InAdvances in Neural Information Processing Systems, volume 33, pages 6840–6851,

arXiv
[9]

Yilun Xu, Ziming Liu, Yonglong Tian, Shangyuan Tong, Max Tegmark, and Tommi Jaakkola

arXiv:2209.11178. Yilun Xu, Ziming Liu, Yonglong Tian, Shangyuan Tong, Max Tegmark, and Tommi Jaakkola. PFGM++: Unlocking the potential of physics-inspired generative models. InProceedings of the 40th International Conference on Machine Learning, ICML,

arXiv
[10]

URL https://arxiv.org/ abs/2302.04265. A Related work Flow matching, diffusion, and equilibrium matching.Flow matching [Lipman et al., 2023, Albergo and Vanden-Eijnden, 2023, Albergo et al., 2023, Liu et al., 2023] and diffusion models [Ho et al., 2020, Song et al., 2021] construct differential equations with time-dependent drifts that transport between b...

arXiv 2023
[11]

exists and lies in M1, so the conclusion T♯µ0 =µ 1 from Step 3 holds. Explicit kernels for discreteµ 1 For µ0 =N(0, Id), µ1 = P j pjδxj , and a general interpolant It =α tx0 +β tx1 (with αt +β t = 1 assumed for brevity), the contribution of atomjtoνandjis, withϵ=x−x j ̸= 0: ν(j)(x) = pj (2π)d/2 Z 1 0 α−d t exp − |ϵ+α txj|2 2α2 t dt,(24) j(j)(x) = pj (2π)d...

work page 2022
[2000]

Boffi, Michael S

Nicholas M. Boffi, Michael S. Albergo, and Eric Vanden-Eijnden. Flow map matching with stochastic interpolants: A mathematical framework for consistency models.Transactions on Machine Learning Research, 2025a. arXiv:2406.07507. Nicholas M. Boffi, Michael S. Albergo, and Eric Vanden-Eijnden. How to build a consistency model: Learning flow maps via self-dis...

arXiv
[2009]

Equilibrium matching: Generative modeling with implicit energy-based models.arXiv preprint arXiv:2510.02300,

Runqian Wang and Yilun Du. Equilibrium matching: Generative modeling with implicit energy-based models.arXiv preprint arXiv:2510.02300,

arXiv
[2019]

Dynamical computation of the density of states and Bayes factors using nonequilibrium importance sampling

arXiv:1809.11132. 10 Filippo Santambrogio.Optimal Transport for Applied Mathematicians, volume 87 ofProgress in Nonlinear Differential Equations and Their Applications. Birkhäuser,

work page Pith review arXiv

Show all 13 references

[2022]

Mingyang Deng, He Li, Tianhong Li, Yilun Du, and Kaiming He

arXiv:2206.09908. Mingyang Deng, He Li, Tianhong Li, Yilun Du, and Kaiming He. Generative modeling via drifting. arXiv preprint arXiv:2602.04770,

arXiv
[2023]

Albergo, Nicholas M

Michael S. Albergo, Nicholas M. Boffi, and Eric Vanden-Eijnden. Stochastic interpolants: A unifying framework for flows and diffusions.arXiv preprint arXiv:2303.08797,

arXiv
[2024]

URLhttps://arxiv.org/abs/2401.08740. Grant M. Rotskoff and Eric Vanden-Eijnden. Dynamical computation of the density of states and Bayes factors using nonequilibrium importance sampling.Physical Review Letters, 122(15): 150602,

arXiv
[2025]

Zhengyang Geng, Mingyang Deng, Xingjian Bai, J

Oral; arXiv:2410.12557. Zhengyang Geng, Mingyang Deng, Xingjian Bai, J. Zico Kolter, and Kaiming He. Mean flows for one-step generative modeling. InAdvances in Neural Information Processing Systems (NeurIPS),

arXiv
[2026]

Drifting relies on learning the fixed-point of a McKean-Vlasov dynamics, driving samples from measure µ0 to a target probability measureµ ∗ ast→ ∞

both approximate the fixed points of transport maps, they differ fundamentally in implementation. Drifting relies on learning the fixed-point of a McKean-Vlasov dynamics, driving samples from measure µ0 to a target probability measureµ ∗ ast→ ∞. In equation, this is written as...

2026

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

The condition’s role in higher codim is to promote oblique streamlines to normal — a stronger but inessential form of regularity

In codim ≥2 it is not required: T∈C 1 for any f. The condition’s role in higher codim is to promote oblique streamlines to normal — a stronger but inessential form of regularity. G.10 Atomic case (k=d) Ty(x)M1 ={0},y(x) ∥ ≡0, hencej ∥ ≡0trivially. The codim-dformula (43) forj ⊥ applies. 21 H Self-stopping interpolant for BTM image experiments Motivated by...

work page 2025

[2] [5]

Jonathan Ho, Ajay Jain, and Pieter Abbeel

Oral; arXiv:2505.13447. Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. InAdvances in Neural Information Processing Systems, volume 33, pages 6840–6851,

arXiv

[3] [9]

Yilun Xu, Ziming Liu, Yonglong Tian, Shangyuan Tong, Max Tegmark, and Tommi Jaakkola

arXiv:2209.11178. Yilun Xu, Ziming Liu, Yonglong Tian, Shangyuan Tong, Max Tegmark, and Tommi Jaakkola. PFGM++: Unlocking the potential of physics-inspired generative models. InProceedings of the 40th International Conference on Machine Learning, ICML,

arXiv

[4] [10]

URL https://arxiv.org/ abs/2302.04265. A Related work Flow matching, diffusion, and equilibrium matching.Flow matching [Lipman et al., 2023, Albergo and Vanden-Eijnden, 2023, Albergo et al., 2023, Liu et al., 2023] and diffusion models [Ho et al., 2020, Song et al., 2021] construct differential equations with time-dependent drifts that transport between b...

arXiv 2023

[5] [11]

exists and lies in M1, so the conclusion T♯µ0 =µ 1 from Step 3 holds. Explicit kernels for discreteµ 1 For µ0 =N(0, Id), µ1 = P j pjδxj , and a general interpolant It =α tx0 +β tx1 (with αt +β t = 1 assumed for brevity), the contribution of atomjtoνandjis, withϵ=x−x j ̸= 0: ν(j)(x) = pj (2π)d/2 Z 1 0 α−d t exp − |ϵ+α txj|2 2α2 t dt,(24) j(j)(x) = pj (2π)d...

work page 2022

[6] [2000]

Boffi, Michael S

Nicholas M. Boffi, Michael S. Albergo, and Eric Vanden-Eijnden. Flow map matching with stochastic interpolants: A mathematical framework for consistency models.Transactions on Machine Learning Research, 2025a. arXiv:2406.07507. Nicholas M. Boffi, Michael S. Albergo, and Eric Vanden-Eijnden. How to build a consistency model: Learning flow maps via self-dis...

arXiv

[7] [2009]

Equilibrium matching: Generative modeling with implicit energy-based models.arXiv preprint arXiv:2510.02300,

Runqian Wang and Yilun Du. Equilibrium matching: Generative modeling with implicit energy-based models.arXiv preprint arXiv:2510.02300,

arXiv

[8] [2019]

Dynamical computation of the density of states and Bayes factors using nonequilibrium importance sampling

arXiv:1809.11132. 10 Filippo Santambrogio.Optimal Transport for Applied Mathematicians, volume 87 ofProgress in Nonlinear Differential Equations and Their Applications. Birkhäuser,

work page Pith review arXiv