pith. sign in

arxiv: 2606.05327 · v1 · pith:7R2RBR4Znew · submitted 2026-06-03 · 💻 cs.LG · q-bio.QM· stat.ML

Multimarginal flow matching with optimal transport potentials

Raghav Kansal , David Crair , Nghia Nguyen , Scott Pope , Bradley Parry This is my paper

Pith reviewed 2026-06-28 07:26 UTC · model grok-4.3

classification 💻 cs.LG q-bio.QMstat.ML
keywords flow matchingoptimal transportmultimarginal transportgenerative modelingsingle-cell RNA sequencingdynamical systems
0
0 comments X

The pith

Optimal transport potentials extend flow matching to multiple observed marginals while keeping training simulation-free.

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

The paper shows that dynamic optimal transport potentials can be added to the conditional flow matching objective to steer generated paths toward any number of intermediate observed distributions. This produces a training loss that stays closed-form and simulation-free yet enforces soft consistency with the extra marginals at their prescribed times. A sympathetic reader cares because sequential snapshot data appear routinely in single-cell biology, ocean currents, and weather, yet prior flow matching handled only the two endpoints. The resulting OTP-FM algorithm is reported to achieve state-of-the-art accuracy and speed on those three application domains.

Core claim

By extending the conditional flow matching loss with potential terms drawn from the dynamic optimal transport action, the authors obtain a simulation-free objective whose minimizers are flows that match both the endpoint distributions and any supplied intermediate marginals.

What carries the argument

Optimal transport potentials, which are added to the dynamic OT action to softly penalize deviation from intermediate marginals and are then folded directly into the conditional flow matching training target.

If this is right

  • Flows can be trained to respect any number of observed time-point distributions without extra simulation cost.
  • The learned vector fields remain flexible in their spatiotemporal evolution between the fixed points.
  • Training scales to the same regime as ordinary conditional flow matching because the extra loss terms are evaluated from samples.
  • The same construction applies to any conditional flow matching variant that already admits a simulation-free objective.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The potential construction could be reused inside other transport-based generative models that currently handle only two marginals.
  • One could test whether the same potentials improve performance when the intermediate marginals are noisy or partially observed.
  • The method opens a route to continuous-time interpolation tasks where the data supply more than start and end snapshots.

Load-bearing premise

The potentials can be inserted into the flow matching loss without destroying its closed-form, simulation-free character or its ability to match the endpoints exactly.

What would settle it

A controlled synthetic experiment in which OTP-FM is trained on three known marginals and then checked to see whether the generated paths actually pass near the middle marginal at the prescribed time; failure would falsify the steering claim.

Figures

Figures reproduced from arXiv: 2606.05327 by Bradley Parry, David Crair, Nghia Nguyen, Raghav Kansal, Scott Pope.

Figure 1
Figure 1. Figure 1: (Left) Comparing standard CFM — straight-line trajectories ignoring intermediate marginals; multimarginal CFM — stitching CFM trajectories piecewise between consecutive marginals; prescriptive approaches such as MMFM and 3MSBM — using fixed interpolation strategies to smooth kinks; and OTP-FM, whose soft potential-driven dynamics with tunable strength w, temporal width τ , and λ shape yields smooth and fle… view at source ↗
Figure 2
Figure 2. Figure 2: Top: Exact solutions to the marginal dynamic OTP problem for 1D Gaussian marginals for varying potentials, strengths, and λk(t). Bottom: OTP-FM solutions for the same marginals and potentials, except the rightmost plot, which demonstrates D = W∞2 . kernels λk : [0, 1] → R + centered around tk with charac￾teristic width τ ∈ R +. The choice of statistical distance D remains free; in fact, in the soft setting… view at source ↗
Figure 3
Figure 3. Figure 3: Training time vs. performance of different methods for the CITE 5D L1O (top) and EB 100D L2O (bottom) experiments. Red stars denote OTP-FM. Top right is better [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: A characteristic example of the exact dynamic OT solution (under the Gaussian ansatz) with MMD and a polynomial kernel failing to enforce the intermediate marginal at high strength. Hence, combining all terms with E[(x ⊤y + c) 2 ] = E[(x ⊤y) 2 ] + 2cE[x ⊤y] + c 2 : DMMD2 [ρt, µtk ] = ∥mρt ∥ 4 + dσ4 ρt + 2σ 2 ρt ∥mρt ∥ 2 + 2c∥mρt ∥ 2 + c 2 − 2 h (m⊤ ρtmµtk ) 2 + dσ2 ρt σ 2 µtk + σ 2 ρt ∥mµtk ∥ 2 + σ 2 µtk ∥… view at source ↗
Figure 5
Figure 5. Figure 5: Left: Example loss curves and α(i) schedule. Center: Progression of target conditional trajectories during training for the W2 2 distance based on example samples from four 1D marginals. Right: Example time sampling distribution pt1,t2 . 26 [PITH_FULL_IMAGE:figures/full_fig_p026_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Example OTP-FM trajectories inferred from 1, 2, 5, 10, and 50 steps per marginal (left to right), demonstrating the effectiveness of the MeanFlow consistency model. F. Experimental details and studies on synthetic data F.1. Architecture We use a three-layer MLP for the MeanFlow velocity model with 256 nodes per hidden layer and the SiLU activation function after each hidden layer. As stated in Sec. 4.3, we… view at source ↗
Figure 7
Figure 7. Figure 7: Example OTP-FM trajectories trained with the LSD consistency objective and inferred from 1, 2, 5, 10, and 50 steps per marginal (left to right), demonstrating comparable performance to the MeanFlow model. On the other hand, the W2 2 , W∞ 2 , and KLD with a Gaussian score estimator are able to learn this. This suggests that kernel-based estimates of the MMD and KLD ∇g provide too noisy and unstable a traini… view at source ↗
Figure 8
Figure 8. Figure 8: The exact dynamic OT solution with MMD and the RBF kernel (top left) demonstrates that the intermediate marginal is effectively enforced by the potential. However, kernel-based estimates of the MMD (top center) and KLD (top right) ∇g do not allow OTP-FM to learn the intermediate marginal well (even for very high strength potentials). Finally, the W2 2 (bottom left), W∞2 (bottom center), and KLD with a Gaus… view at source ↗
Figure 9
Figure 9. Figure 9: Top: Varying the half-width τ of the λk(t)s, with Gaussian shape, from 0.01, 0.05, 0.1, 0.2, to 0.4 (left to right). Bottom: Varying the shape of the λk(t)s, with τ = 0.1, between Gaussian, triangular, and rectangular (left to right). 31 [PITH_FULL_IMAGE:figures/full_fig_p031_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Top: Example loss curves for the three different loss functions: squared L2 (top left, log scale), adaptive weights with p = 1 (top center), and learnt log-variance (top right). Bottom: The corresponding OTP-FM trajectories for each loss function (for the same marginals and potential strengths). 32 [PITH_FULL_IMAGE:figures/full_fig_p032_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Ground truth and simulated trajectories for the EB 100D dataset in the leave-two-out setting (t1 and t3 marginals held out) for different OTP-FM potential parameters. The “baseline” model uses D = W2 2 with a Gaussian λ(t), with half-width τ = 0.33 and strength w = 1000. All other models use the same parameters except for the difference specified. and CITE 50D runs because of the prohibitively slow traini… view at source ↗
Figure 12
Figure 12. Figure 12: Training time vs. performance of different methods for the EB 5D L1O (left), EB 100D L1O (center), and CITE 50D L1O (right) experiments. Top right is better. DMSB DMSB (Chen et al., 2023) is a neural SDE method that lifts multi-marginal Schrodinger bridges to phase space to ¨ smooth the dynamics. We use the provided code14 with the authors’ EB 100D and CITE 50D configurations; the EB 5D L1O value is repor… view at source ↗
Figure 13
Figure 13. Figure 13: Ground truth and simulated trajectories for the EB 100D dataset in the leave-two-out setting (t1 and t3 marginals held out) in all four PCs across all five time intervals, comparing OTP-FM with different potentials with baseline models. 38 [PITH_FULL_IMAGE:figures/full_fig_p038_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Ground truth and simulated trajectories for the GoM dataset, comparing MMFM and OTP-FM with different potentials. H. Experimental details on the Gulf of Mexico dataset H.1. Dataset This dataset comprises high-resolution bathymetric measurements of the Gulf of Mexico (GoM) in the region between 98◦E and 77◦E in longitude and from 18◦N to 32◦N in latitude. The original dataset was released to the public by … view at source ↗
Figure 15
Figure 15. Figure 15: Visualizing the Beijing PM2.5 air quality dataset. (a) Ground truth marginal distributions as violin plots showing the density at each timepoint, the median, and the interquartile range. The marginals are heavily concentrated at low PM2.5 but with long tails. (b) Cubic spline interpolation through OT-coupled ground-truth samples, excluding held-out marginals; this represents MMFM’s training targets. (c) T… view at source ↗
read the original abstract

Flow matching (FM) has emerged as a powerful framework for learning dynamic transport maps between two empirical distributions. However, less explored is the setting with intermediate observed marginals that can help constrain the flows between the endpoints. This "multimarginal" regime is central to modeling temporal evolution in dynamical systems in many scientific domains that can sample sequential distributions. We tackle this problem with a novel approach that leverages the connection between FM and dynamic optimal transport (OT), softly steering the flow towards the intermediate marginals through potential terms in the dynamic OT action. By extending the conditional FM learning target to incorporate these potentials, we derive an efficient, simulation-free algorithm for multimarginal FM that offers considerable flexibility in the spatiotemporal dynamics of the learned flows. We demonstrate state-of-the-art performance and training efficiency of OT-potential FM (OTP-FM) on diverse single-cell RNA sequencing, oceanographic, and meteorological datasets. Our code is available at https://github.com/Bexorg-Inc/OTP-FM.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 0 minor

Summary. The paper proposes OT-potential Flow Matching (OTP-FM), which extends conditional flow matching by incorporating potential terms derived from the dynamic optimal transport action. This modification is claimed to softly steer learned flows toward observed intermediate marginals while preserving endpoint matching, yielding an efficient simulation-free training objective for multimarginal problems. The method is demonstrated on single-cell RNA sequencing, oceanographic, and meteorological datasets with reported state-of-the-art performance and training efficiency.

Significance. If the central derivation holds and the simulation-free property is preserved, the approach would supply a flexible, computationally attractive framework for learning flows constrained by multiple observed marginals, addressing a practically relevant gap in scientific applications involving sequential distributions.

major comments (1)
  1. [Abstract] Abstract: the claim that extending the conditional FM target with OT potentials yields a simulation-free algorithm is load-bearing for the central contribution, yet the abstract supplies no indication that the potentials are restricted to a form (e.g., quadratic or separable) that restores closed-form conditional path velocities; without such restriction the target velocity generally requires solving the continuity equation or per-sample optimization of the augmented action.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the detailed reading and the focus on the simulation-free claim. The comment correctly identifies that the abstract does not explicitly flag the restrictions on the OT potentials needed to retain closed-form conditional velocities. We address this below and will revise the abstract accordingly.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim that extending the conditional FM target with OT potentials yields a simulation-free algorithm is load-bearing for the central contribution, yet the abstract supplies no indication that the potentials are restricted to a form (e.g., quadratic or separable) that restores closed-form conditional path velocities; without such restriction the target velocity generally requires solving the continuity equation or per-sample optimization of the augmented action.

    Authors: We agree that the abstract should make the restriction on the potentials explicit, as this is necessary to preserve the simulation-free property. In the full derivation (Section 3), the dynamic OT potentials are restricted to quadratic forms in the position variable (or separable in time and space) so that the augmented conditional vector field admits an analytic expression; the resulting training objective therefore remains a simple regression against the modified target velocity without requiring numerical integration or per-sample optimization. We will revise the abstract to state: “By extending the conditional FM learning target to incorporate quadratic OT potentials, we derive an efficient, simulation-free algorithm...” This change clarifies the load-bearing assumption without altering the technical contribution. revision: yes

Circularity Check

0 steps flagged

No circularity; derivation extends FM target independently

full rationale

The abstract and description present an extension of conditional flow matching by adding OT potential terms to the learning target, yielding a claimed simulation-free multimarginal objective. No equations, self-citations, or fitted quantities are shown that reduce the central result to a redefinition of the input data or a parameter fit by construction. The connection to dynamic OT is invoked as an external link, and the simulation-free property is asserted via the extension rather than tautologically assumed. This matches the default case of a self-contained methodological proposal with no load-bearing circular steps.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract only; no explicit free parameters, axioms, or invented entities are stated.

pith-pipeline@v0.9.1-grok · 5709 in / 990 out tokens · 26757 ms · 2026-06-28T07:26:40.070424+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

80 extracted references · 12 canonical work pages

  1. [1]

    The Eleventh International Conference on Learning Representations , year=

    Flow Matching for Generative Modeling , author=. The Eleventh International Conference on Learning Representations , year=

  2. [2]

    2025 , eprint=

    Stochastic Interpolants: A Unifying Framework for Flows and Diffusions , author=. 2025 , eprint=

    2025

  3. [3]

    Transactions on Machine Learning Research , issn=

    Improving and generalizing flow-based generative models with minibatch optimal transport , author=. Transactions on Machine Learning Research , issn=. 2024 , url=

    2024

  4. [4]

    The Eleventh International Conference on Learning Representations , year=

    Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow , author=. The Eleventh International Conference on Learning Representations , year=

  5. [5]

    Proceedings of the 40th International Conference on Machine Learning (ICML) , pages =

    Multisample Flow Matching: Straightening Flows with Minibatch Couplings , author =. Proceedings of the 40th International Conference on Machine Learning (ICML) , pages =. 2023 , volume =

    2023

  6. [6]

    Numerische Mathematik , year =

    Jean-David Benamou and Yann Brenier , title =. Numerische Mathematik , year =. doi:10.1007/s002110050002 , url =

    work page doi:10.1007/s002110050002

  7. [7]

    Springer Science & Business Media, 2008

    Cédric Villani , title =. 2009 , series =. doi:10.1007/978-3-540-71050-9 , url =

    work page doi:10.1007/978-3-540-71050-9 2009

  8. [8]

    Integral Probability Metrics and Their Generating Classes of Functions , journal =

    M. Integral Probability Metrics and Their Generating Classes of Functions , journal =. 1997 , publisher =. doi:10.2307/1428011 , url =

    work page doi:10.2307/1428011 1997

  9. [9]

    Advances in Neural Information Processing Systems , year =

    Neural Ordinary Differential Equations , author =. Advances in Neural Information Processing Systems , year =

  10. [10]

    Journal of Machine Learning Research , volume =

    A Kernel Two-Sample Test , author =. Journal of Machine Learning Research , volume =. 2012 , url =

    2012

  11. [11]

    International Conference on Learning Representations (ICLR) , year =

    Progressive Distillation for Fast Sampling of Diffusion Models , author =. International Conference on Learning Representations (ICLR) , year =. 2202.00512 , archivePrefix=

    Pith/arXiv arXiv

  12. [12]

    Advances in Neural Information Processing Systems (NeurIPS) , year =

    Elucidating the Design Space of Diffusion-Based Generative Models , author =. Advances in Neural Information Processing Systems (NeurIPS) , year =. 2206.00364 , archivePrefix=

    Pith/arXiv arXiv

  13. [13]

    2023 , eprint =

    Latent Consistency Models: Synthesizing High-Resolution Images with Few-Step Inference , author =. 2023 , eprint =

    2023

  14. [14]

    Proceedings of the 40th International Conference on Machine Learning (ICML) , year =

    Consistency Models , author =. Proceedings of the 40th International Conference on Machine Learning (ICML) , year =. 2303.01469 , archivePrefix=

    Pith/arXiv arXiv

  15. [15]

    arXiv preprint arXiv:2310.14189 , year =

    Improved Techniques for Training Consistency Models , author =. arXiv preprint arXiv:2310.14189 , year =. 2310.14189 , archivePrefix=

    Pith/arXiv arXiv

  16. [16]

    Transactions on Machine Learning Research , year =

    Flow map matching with stochastic interpolants: A mathematical framework for consistency models , author =. Transactions on Machine Learning Research , year =

  17. [17]

    The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

    How to build a consistency model: Learning flow maps via self-distillation , author=. The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

  18. [18]

    The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

    Mean Flows for One-step Generative Modeling , author=. The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

  19. [19]

    International Conference on Learning Representations , year =

    Consistency Trajectory Models: Learning Probability Flow ODE Trajectory of Diffusion , author =. International Conference on Learning Representations , year =. 2310.02279 , url =

    arXiv

  20. [20]

    arXiv preprint arXiv:2402.19159 , year =

    Trajectory Consistency Distillation: Improved Latent Consistency Distillation by Semi-Linear Consistency Function with Trajectory Mapping , author =. arXiv preprint arXiv:2402.19159 , year =. 2402.19159 , archivePrefix=

    arXiv

  21. [21]

    International Conference on Learning Representations , year =

    One Step Diffusion via Shortcut Models , author =. International Conference on Learning Representations , year =. 2410.12557 , url =

    Pith/arXiv arXiv

  22. [22]

    The Thirty-ninth Annual Conference on Neural Information Processing Systems , year =

    Align Your Flow: Scaling Continuous-Time Flow Map Distillation , author =. The Thirty-ninth Annual Conference on Neural Information Processing Systems , year =

  23. [23]

    Advances in Neural Information Processing Systems , volume=

    Attention is All You Need , author=. Advances in Neural Information Processing Systems , volume=. 2017 , eprint=

    2017

  24. [24]

    and Georg, Kurt , title =

    Allgower, Eugene L. and Georg, Kurt , title =. 2003 , doi =

    2003

  25. [25]

    Journal of the

    Iterative Procedures for Nonlinear Integral Equations , author =. Journal of the. 1965 , doi =

    1965

  26. [26]

    Soviet Physics Uspekhi , volume =

    The Vibrational Properties of an Electron Gas , author =. Soviet Physics Uspekhi , volume =. 1968 , doi =

    1968

  27. [27]

    1983 , isbn =

    Introduction to Plasma Theory , author =. 1983 , isbn =

    1983

  28. [28]

    Proceedings of the National Academy of Sciences , volume =

    A Class of Markov Processes Associated with Nonlinear Parabolic Equations , author =. Proceedings of the National Academy of Sciences , volume =. 1966 , doi =

    1966

  29. [29]

    The Annals of Applied Probability , volume =

    Propagation of Chaos in Neural Fields , author =. The Annals of Applied Probability , volume =. 2014 , doi =

    2014

  30. [30]

    Advances in Neural Information Processing Systems , editor =

    On the Properties of Kullback-Leibler Divergence Between Multivariate Gaussian Distributions , author =. Advances in Neural Information Processing Systems , editor =

  31. [31]

    2006 , isbn =

    Statistical Inference Based on Divergence Measures , author =. 2006 , isbn =. doi:10.1201/9781420034813 , url =

    work page doi:10.1201/9781420034813 2006

  32. [32]

    1980 , series =

    Introduction to Numerical Analysis , author =. 1980 , series =. doi:10.1007/978-1-4757-5592-3 , url =

    work page doi:10.1007/978-1-4757-5592-3 1980

  33. [33]

    , school =

    van de Rotten, Bastiaan A. , school =. A Limited Memory. 2003 , address =

    2003

  34. [34]

    Solving Ordinary Differential Equations

    Hairer, Ernst and Wanner, Gerhard , publisher =. Solving Ordinary Differential Equations. 1996 , edition =. doi:10.1007/978-3-642-05221-7 , url =

    work page doi:10.1007/978-3-642-05221-7 1996

  35. [35]

    and others , journal =

    Virtanen, Pauli and Gommers, Ralf and Oliphant, Travis E. and others , journal =. 2020 , doi =

    2020

  36. [36]

    Advances in Neural Information Processing Systems , volume =

    Sinkhorn Distances: Lightspeed Computation of Optimal Transport , author =. Advances in Neural Information Processing Systems , volume =. 2013 , url =

    2013

  37. [37]

    Proceedings of the Twenty Third International Conference on Artificial Intelligence and Statistics , pages =

    Learning with minibatch Wasserstein: asymptotic and gradient properties , author =. Proceedings of the Twenty Third International Conference on Artificial Intelligence and Statistics , pages =. 2020 , volume =

    2020

  38. [38]

    2019 , doi =

    Computational Optimal Transport , author =. 2019 , doi =

    2019

  39. [39]

    Advances in Neural Information Processing Systems , year =

    Near-linear time approximation algorithms for optimal transport via Sinkhorn iteration , author =. Advances in Neural Information Processing Systems , year =. 1705.09634 , archivePrefix=

    Pith/arXiv arXiv

  40. [40]

    2024 , url =

    Karras, Tero and Aittala, Miika and Lehtinen, Jaakko and Hellsten, Janne and Aila, Timo and Laine, Samuli , booktitle =. 2024 , volume =. doi:10.1109/CVPR52733.2024.02282 , url =

    work page doi:10.1109/cvpr52733.2024.02282 2024

  41. [41]

    Nature , year=

    Mapping cells through time and space with moscot , author=. Nature , year=. doi:10.1038/s41586-024-08453-2 , url=

    work page doi:10.1038/s41586-024-08453-2

  42. [42]

    Nature Methods , volume=

    Learning single-cell perturbation responses using neural optimal transport , author=. Nature Methods , volume=. 2023 , doi=

    2023

  43. [43]

    Nature Reviews Methods Primers , volume=

    Optimal transport for single-cell and spatial omics , author=. Nature Reviews Methods Primers , volume=. 2024 , doi=

    2024

  44. [44]

    Nature Biotechnology , volume=

    Visualizing structure and transitions in high-dimensional biological data , author=. Nature Biotechnology , volume=. 2019 , doi=

    2019

  45. [45]

    Advances in Neural Information Processing Systems , year=

    Manifold Interpolating Optimal-Transport Flows for Trajectory Inference , author=. Advances in Neural Information Processing Systems , year=

  46. [46]

    Proceedings of the 37th International Conference on Machine Learning , pages=

    TrajectoryNet: A Dynamic Optimal Transport Network for Modeling Cellular Dynamics , author=. Proceedings of the 37th International Conference on Machine Learning , pages=. 2020 , volume=

    2020

  47. [47]

    , booktitle=

    Chen, Tianrong and Liu, Guan-Horng and Theodorou, Evangelos A. , booktitle=. Likelihood Training of Schr. 2022 , url=

    2022

  48. [48]

    2023 , url=

    Tianrong Chen and Guan-Horng Liu and Molei Tao and Evangelos Theodorou , booktitle=. 2023 , url=

    2023

  49. [49]

    2023 , url=

    Takeshi Koshizuka and Issei Sato , booktitle=. 2023 , url=

    2023

  50. [50]

    Diffusion Schr

    Shi, Yuyang and De Bortoli, Valentin and Campbell, Andrew and Doucet, Arnaud , booktitle=. Diffusion Schr. 2023 , url=

    2023

  51. [51]

    Trajectory Inference with Smooth Schr

    Hong, Wanli and Shi, Yuliang and Niles-Weed, Jonathan , journal=. Trajectory Inference with Smooth Schr. 2025 , url=

    2025

  52. [52]

    International Conference on Learning Representations , year=

    Multi-Modal and Multi-Attribute Generation of Single Cells with CFGen , author=. International Conference on Learning Representations , year=. 2407.11734 , url=

    arXiv

  53. [53]

    bioRxiv , year=

    CellFlow enables generative single-cell phenotype modeling with flow matching , author=. bioRxiv , year=. doi:10.1101/2025.04.11.648220 , url=

    work page doi:10.1101/2025.04.11.648220 2025

  54. [54]

    International Conference on Learning Representations , year=

    Modeling Complex System Dynamics with Flow Matching Across Time and Conditions , author=. International Conference on Learning Representations , year=

  55. [55]

    The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

    Momentum Multi-Marginal Schr\"odinger Bridge Matching , author=. The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

  56. [56]

    Proceedings of The 24th International Conference on Artificial Intelligence and Statistics , pages=

    Fast and Smooth Interpolation on Wasserstein Space , author=. Proceedings of The 24th International Conference on Artificial Intelligence and Statistics , pages=. 2021 , volume=

    2021

  57. [57]

    Physical Review D , volume=

    Evaluating generative models in high energy physics , author=. Physical Review D , volume=. 2023 , doi=. 2211.10295 , url=

    arXiv 2023

  58. [58]

    Advances in Neural Information Processing Systems , volume=

    GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium , author=. Advances in Neural Information Processing Systems , volume=. 2017 , eprint=

    2017

  59. [59]

    Multi-marginal Schr

    Shen, Yunyi and Berlinghieri, Renato and Broderick, Tamara , booktitle=. Multi-marginal Schr. 2025 , eprint=

    2025

  60. [60]

    2024 , howpublished=

    HYCOM-TSIS 1/100 ^ Gulf of Mexico Reanalysis (GOMb0.01) , author=. 2024 , howpublished=

    2024

  61. [61]

    2017 , howpublished =

    Chen, Song , title =. 2017 , howpublished =. doi:10.24432/C5RK5G , url=

    work page doi:10.24432/c5rk5g 2017

  62. [62]

    Preuer, Kristina and Renz, Philipp and Unterthiner, Thomas and Hochreiter, Sepp and Klambauer, G. Fr. Journal of Chemical Information and Modeling , year =. doi:10.1021/acs.jcim.8b00234 , url =

    work page doi:10.1021/acs.jcim.8b00234

  63. [63]

    Particle Cloud Generation with Message Passing Generative Adversarial Networks , url =

    Kansal, Raghav and Duarte, Javier and Su, Hao and Orzari, Breno and Tomei, Thiago and Pierini, Maurizio and Touranakou, Mary and vlimant, jean-roch and Gunopulos, Dimitrios , booktitle =. Particle Cloud Generation with Message Passing Generative Adversarial Networks , url =. 2021 , eprint =

    2021

  64. [64]

    and Jonsson, Bror and Kulk, Gemma and Forget, Gael and Racault, Marie-Fanny and Jackson, Thomas and Dutkiewicz, Stephanie and Müller, Christian L

    Hyun, Sangwon and Mishra, Aditya and Follett, Christopher L. and Jonsson, Bror and Kulk, Gemma and Forget, Gael and Racault, Marie-Fanny and Jackson, Thomas and Dutkiewicz, Stephanie and Müller, Christian L. and Bien, Jacob , title =. Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences , volume =. 2022 , month =. doi:10.109...

    work page doi:10.1098/rspa.2021.0875 2022

  65. [65]

    The Thirty-eighth Annual Conference on Neural Information Processing Systems , year=

    Metric Flow Matching for Smooth Interpolations on the Data Manifold , author=. The Thirty-eighth Annual Conference on Neural Information Processing Systems , year=

  66. [66]

    A Computational Framework for Solving

    Neklyudov, Kirill and Brekelmans, Rob and Tong, Alexander and Atanackovic, Lazar and Liu, Qiang and Makhzani, Alireza , booktitle =. A Computational Framework for Solving. 2024 , editor =

    2024

  67. [67]

    Learning of Population Dynamics: Inverse Optimization Meets

    Mikhail Persiianov and Jiawei Chen and Petr Mokrov and Alexander Tyurin and Evgeny Burnaev and Alexander Korotin , booktitle=. Learning of Population Dynamics: Inverse Optimization Meets. 2026 , url=

    2026

  68. [68]

    The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

    Joint Velocity-Growth Flow Matching for Single-Cell Dynamics Modeling , author=. The Thirty-ninth Annual Conference on Neural Information Processing Systems , year=

  69. [69]

    Simulation-Free Schr\ "odinger Bridges via Score and Flow Matching , urldate =

    Tong, Alexander and Malkin, Nikolay and Fatras, Kilian and Atanackovic, Lazar and Zhang, Yanlei and Huguet, Guillaume and Wolf, Guy and Bengio, Yoshua , eprint =. Simulation-Free Schr\ "odinger Bridges via Score and Flow Matching , urldate =. 2023 , booktitle =

    2023

  70. [70]

    The Thirteenth International Conference on Learning Representations , year=

    Learning stochastic dynamics from snapshots through regularized unbalanced optimal transport , author=. The Thirteenth International Conference on Learning Representations , year=

  71. [71]

    Gradient Flows in Metric Spaces and in the Space of Probability Measures , publisher =

    Luigi Ambrosio and Nicola Gigli and Giuseppe Savar. Gradient Flows in Metric Spaces and in the Space of Probability Measures , publisher =. 2008 , doi =

    2008

  72. [72]

    An Introduction to -Convergence , publisher =

    Gianni. An Introduction to -Convergence , publisher =. 1993 , doi =

    1993

  73. [73]

    Proceedings of The 25th International Conference on Artificial Intelligence and Statistics , year=

    Proximal Optimal Transport Modeling of Population Dynamics , author=. Proceedings of The 25th International Conference on Artificial Intelligence and Statistics , year=. 2106.06345 , archivePrefix=

    arXiv

  74. [74]

    2006 , URL =

    Braides, Andrea , title =. 2006 , URL =

    2006

  75. [75]

    2015 , doi =

    Filippo Santambrogio , title =. 2015 , doi =

    2015

  76. [76]

    2026 , eprint=

    Improved Mean Flows: On the Challenges of Fastforward Generative Models , author=. 2026 , eprint=

    2026

  77. [77]

    The Thirty-eighth Annual Conference on Neural Information Processing Systems , year=

    Learning diffusion at lightspeed , author=. The Thirty-eighth Annual Conference on Neural Information Processing Systems , year=

  78. [78]

    2022 , howpublished =

    Daniel Burkhardt and Malte Luecken and Andrew Benz and Peter Holderrieth and Jonathan Bloom and Christopher Lance and Ashley Chow and Ryan Holbrook , title =. 2022 , howpublished =

    2022

  79. [79]

    2026 , url =

    Peter Holderrieth and Ezra Erives , title =. 2026 , url =. 2506.02070 , archivePrefix =

    arXiv 2026

  80. [80]

    2026 , eprint=

    One-step Latent-free Image Generation with Pixel Mean Flows , author=. 2026 , eprint=

    2026