pith. machine review for the scientific record. sign in

arxiv: 2605.07319 · v1 · submitted 2026-05-08 · 💻 cs.LG · cs.AI

Recognition: no theorem link

Generative Modeling with Flux Matching

Peter Pao-Huang , Xiaojie Qiu , Stefano Ermon
Authors on Pith no claims yet

Pith reviewed 2026-05-11 01:25 UTC · model grok-4.3

classification 💻 cs.LG cs.AI
keywords generative modelingflux matchingvector fieldsstationary distributionscore-based modelsinductive biasesdynamics design
0
0 comments X

The pith

Flux Matching relaxes score matching to admit any vector field whose stationary distribution is the data.

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

The paper introduces Flux Matching to train generative models using a weaker condition than matching the data score. Instead of forcing the vector field to equal the gradient of the log-density, the objective only requires that the data distribution remains unchanged under the dynamics in the long run. This change opens up infinitely many possible vector fields that can still generate samples from the target distribution. The added freedom lets designers directly specify or optimize properties of the dynamics such as sampling speed, interpretability, or variable dependencies that score matching cannot accommodate.

Core claim

Flux Matching trains a neural network to produce a vector field satisfying a divergence condition that makes the data distribution stationary, without requiring the field to be conservative. This generalization admits a much larger family of dynamics than score-based methods and turns the vector field itself into a tunable design choice rather than a fixed target to match.

What carries the argument

The Flux Matching objective, which enforces that the probability flux out of any region equals the change in probability mass inside it so the data distribution stays stationary.

If this is right

  • Sampling can be made faster by choosing vector fields with favorable flow properties.
  • Inductive biases and structural priors can be imposed directly on the dynamics.
  • Directed dependencies between variables can be encoded in the learned vector field.
  • Models become possible that are mechanistic and interpretable by construction.

Where Pith is reading between the lines

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

  • This approach could let generative models borrow vector fields from physics simulators that already satisfy conservation laws or other constraints.
  • One could optimize the vector field explicitly for minimal integration time or for staying on a learned manifold.

Load-bearing premise

Optimizing the weaker objective will produce vector fields that actually keep the data as their long-run stationary distribution without introducing instabilities or needing extra constraints.

What would settle it

Train a Flux Matching model on a low-dimensional mixture of Gaussians, then integrate the learned dynamics for many steps and check whether the generated points converge to the training distribution regardless of starting point.

Figures

Figures reproduced from arXiv: 2605.07319 by Peter Pao-Huang, Stefano Ermon, Xiaojie Qiu.

Figure 1
Figure 1. Figure 1: Ω is the space of vector fields ∈ L 2 (pdata). Score matching learns ∇ log pdata, a single point in this space. In contrast, Flux Matching can learn any vector field inside the rectangle. The narrow focus on the score overlooks a large space of alternative vector fields whose diffusion processes share the same target distribution. We refer to these vector fields as generative vector fields, or generative f… view at source ↗
Figure 2
Figure 2. Figure 2: From Section 4.1 where we maximize different vector field attributes that generate the same [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Geometric interpretation of Flux Matching. Colors are detailed in the accompanying Algorithm 1. Intuitively, rθ (aka the Langevin-Stein operator) is invari￾ant to any pdata-preserving dynamics, but applies the local differential operator ∇·, which sits one derivative beyond the Fisher divergence geometry. Step 1 closes this gap by using diffusion simulations dxt = ∇ log pdata(xt) dt+ √ 2 dWt to propagate t… view at source ↗
Figure 4
Figure 4. Figure 4: Normalized score matching and Flux Matching losses as we vary properties of the vector [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: (Left) Learned RNA velocity using Flux Matching; blue arrows indicate ground-truth biological progression between cell-types. (Right) CBC and consistency means across datasets. leave the distribution unchanged. Flux Matching instead treats the entire distribution preserving family as equivalent. Setup. On a 2D three-component Gaussian mixture, we construct three one-parameter families of distribution prese… view at source ↗
Figure 6
Figure 6. Figure 6: FID (calculated with 1K generated samples) as a function of the number of sampling steps. fitting procedure rather than to added expressivity. We foresee that Flux Matching can be applied to other newly developed RNA velocity models with more sophisticated biological parameterizations. 4.3 Unrestricted Generative Fields Flux Matching’s main value is in imposing structure on the learned vector field, but th… view at source ↗
Figure 7
Figure 7. Figure 7: (Left) Causal attention mask used for trajectory generation. Rows index the output at each trajectory time, and columns index the input states at each trajectory time. The upper-triangular mask allows f σ θ,n to depend only on states xm with m ≤ n, enforcing autoregressive structure while still evaluating all outputs in parallel. (Right) Relative change in Wasserstein distance from adding a causal attentio… view at source ↗
Figure 12
Figure 12. Figure 12: 22 [PITH_FULL_IMAGE:figures/full_fig_p022_12.png] view at source ↗
Figure 8
Figure 8. Figure 8: Bone Marrow Dataset. (Left half) inferred RNA velocity (Right half) ground truth biological progression given by arrows [PITH_FULL_IMAGE:figures/full_fig_p023_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Dentate Gyrus Dataset. (Left half) inferred RNA velocity (Right half) ground truth biological progression given by arrows 23 [PITH_FULL_IMAGE:figures/full_fig_p023_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Gastrulation Dataset. (Left half) inferred RNA velocity (Right half) ground truth biological progression given by arrows [PITH_FULL_IMAGE:figures/full_fig_p024_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Hindbrain Dataset. (Left half) inferred RNA velocity (Right half) ground truth biological progression given by arrows 24 [PITH_FULL_IMAGE:figures/full_fig_p024_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Pancreas Dataset. (Left half) inferred RNA velocity (Right half) ground truth biological progression given by arrows [PITH_FULL_IMAGE:figures/full_fig_p025_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Random CIFAR-10 samples from Flux Matching (left) and DSM (right). [PITH_FULL_IMAGE:figures/full_fig_p027_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Random CelebA samples from Flux Matching (left) and DSM (right). [PITH_FULL_IMAGE:figures/full_fig_p028_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: A single sample from the spring simulation. The system contains two masses, shown [PITH_FULL_IMAGE:figures/full_fig_p030_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Mode crossing and loss variance in a three-component Gaussian mixture. The component [PITH_FULL_IMAGE:figures/full_fig_p033_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Distribution flows can have the same marginal evolution but different particle trajectories. [PITH_FULL_IMAGE:figures/full_fig_p034_17.png] view at source ↗
read the original abstract

We introduce Flux Matching, a new paradigm for generative modeling that generalizes existing score-based models to a broader family of vector fields that need not be conservative. Rather than requiring the model to equal the data score, the Flux Matching objective imposes a weaker condition that admits infinitely many vector fields whose stationary distribution is the data. This flexibility enables a class of generative models that cannot be learned under score matching, in which inductive biases, structural priors, and properties of the dynamics can be directly imposed or optimized. We show that Flux Matching performs strongly on high-dimensional image datasets and, more importantly, that our added freedom unlocks a range of applications including faster sampling, interpretable and mechanistic models, and dynamics that encode directed dependencies between variables. More broadly, Flux Matching opens a new dimension in generative modeling by turning the vector field itself into a design choice rather than a fixed target. Code is available at https://github.com/peterpaohuang/flux_matching.

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

3 major / 2 minor

Summary. The paper introduces Flux Matching as a generalization of score-based generative models. Rather than matching the data score, it proposes a weaker objective on the vector field that still admits the data distribution as a stationary measure, thereby allowing non-conservative dynamics. The authors claim this flexibility enables new model classes with inductive biases, faster sampling, interpretable dynamics, and directed dependencies, and they report strong empirical performance on high-dimensional image datasets.

Significance. If the central theoretical claim holds—that the population Flux Matching objective guarantees the data as an invariant measure for a broader family of vector fields than score matching—and if the finite-sample neural implementation reliably realizes this property, the work would meaningfully expand the design space of generative models by treating the vector field as a tunable object rather than a fixed target. The public code release supports reproducibility and further exploration of the claimed applications.

major comments (3)
  1. [Abstract, §3] Abstract and §3 (theoretical development): the manuscript asserts that the Flux Matching objective imposes a weaker condition than score matching while still ensuring the data distribution is stationary for non-conservative vector fields, yet provides no derivation of the stationary continuity equation or proof that solutions to the population objective satisfy the required divergence condition. Without this, it is unclear whether the claimed family of vector fields is non-empty or whether the objective is sufficient.
  2. [§4] §4 (empirical validation) and experimental results: the reported image-generation metrics evaluate short-horizon sampling quality but do not include direct checks (long-rollout histograms, empirical divergence from the target measure, or probability-current verification) that the learned non-conservative fields actually possess the data distribution as their invariant measure. This leaves the central novelty unverified beyond what score-matching baselines already achieve.
  3. [§4.2] §4.2 (applications): the claims of faster sampling, interpretable mechanistic models, and dynamics encoding directed dependencies rest on the assumption that optimizing the weaker objective reliably produces stable long-term behavior; no ablation or stability analysis is shown for the non-conservative cases that constitute the claimed advantage over score matching.
minor comments (2)
  1. [Abstract, §2] Notation for the vector field and the Flux Matching loss should be introduced with explicit definitions before use in the abstract and early sections to improve readability.
  2. [§4] The GitHub link is provided but the manuscript does not specify which exact experimental configurations (hyperparameters, architectures, non-conservative parameterizations) correspond to the reported numbers.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed report. We address each major comment below with clarifications and proposed revisions to improve the rigor and completeness of the manuscript.

read point-by-point responses

  1. Referee: [Abstract, §3] Abstract and §3 (theoretical development): the manuscript asserts that the Flux Matching objective imposes a weaker condition than score matching while still ensuring the data distribution is stationary for non-conservative vector fields, yet provides no derivation of the stationary continuity equation or proof that solutions to the population objective satisfy the required divergence condition. Without this, it is unclear whether the claimed family of vector fields is non-empty or whether the objective is sufficient.

    Authors: We agree that an explicit derivation would strengthen the theoretical section. Although the stationarity condition follows from the continuity equation and the objective being constructed to enforce a divergence-free probability current at the data distribution, the exposition in §3 is too concise. In the revised manuscript we will expand §3 with a complete derivation: starting from the Fokker-Planck/continuity equation, showing that the population Flux Matching loss is equivalent to requiring the probability current to vanish at the data measure, and explicitly constructing a family of non-conservative vector fields (including a simple 2-D example) that satisfy the condition while differing from the score. revision: yes

  2. Referee: [§4] §4 (empirical validation) and experimental results: the reported image-generation metrics evaluate short-horizon sampling quality but do not include direct checks (long-rollout histograms, empirical divergence from the target measure, or probability-current verification) that the learned non-conservative fields actually possess the data distribution as their invariant measure. This leaves the central novelty unverified beyond what score-matching baselines already achieve.

    Authors: We acknowledge that direct verification of the invariant-measure property is important for substantiating the central claim. For high-dimensional image data, exhaustive long-rollout histograms and full probability-current computations are computationally prohibitive. In the revision we will add, as supplementary material, explicit verification experiments on lower-dimensional problems (Gaussian mixtures and MNIST) that include long-horizon sampling trajectories, empirical estimation of the divergence of the learned vector field, and checks that the data distribution remains stationary. For the main high-dimensional results we will include qualitative long-rollout visualizations demonstrating stability. revision: partial

  3. Referee: [§4.2] §4.2 (applications): the claims of faster sampling, interpretable mechanistic models, and dynamics encoding directed dependencies rest on the assumption that optimizing the weaker objective reliably produces stable long-term behavior; no ablation or stability analysis is shown for the non-conservative cases that constitute the claimed advantage over score matching.

    Authors: The referee is correct that explicit stability analysis for the non-conservative regime is missing. We will add to the revised §4.2 an ablation comparing conservative (score-equivalent) and non-conservative Flux Matching models on long-horizon sampling stability, including quantitative metrics (e.g., KL divergence or MMD after extended rollouts) and qualitative demonstrations of directed-dependency encoding. These results will be used to support the claims of faster sampling and mechanistic interpretability. revision: yes

Circularity Check

0 steps flagged

No significant circularity in the Flux Matching derivation

full rationale

The derivation starts from the stationary-distribution property of a vector field and relaxes the score-matching condition to a weaker flux-matching objective that by construction admits non-conservative fields. This relaxation is presented as an explicit mathematical choice rather than a fit or self-definition; the objective is not obtained by renaming or by fitting a parameter to the target data and then relabeling the fit as a prediction. No load-bearing self-citation, uniqueness theorem imported from the same authors, or ansatz smuggled via prior work is required for the central claim. The paper remains self-contained: the population objective is derived from the continuity equation, the finite-sample loss is a direct Monte-Carlo estimate of that objective, and downstream empirical results on image datasets constitute independent validation rather than tautological confirmation.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the mathematical fact that non-conservative vector fields can have a prescribed stationary distribution, which is a standard property from dynamical systems but is treated as the enabling assumption here.

axioms (1)
  • domain assumption There exist infinitely many vector fields (not necessarily conservative) whose stationary distribution equals the data distribution.
    This is the explicit weaker condition stated in the abstract that replaces the score-matching requirement.

pith-pipeline@v0.9.0 · 5458 in / 1245 out tokens · 36798 ms · 2026-05-11T01:25:48.992601+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

68 extracted references · 68 canonical work pages · 7 internal anchors

  1. [1]

    Abramson, J

    J. Abramson, J. Adler, J. Dunger, R. Evans, T. Green, A. Pritzel, O. Ronneberger, L. Willmore, A. J. Ballard, J. Bambrick, et al. Accurate structure prediction of biomolecular interactions with alphafold 3.Nature, 630(8016):493–500, 2024

    work page 2024

  2. [2]

    M. S. Albergo and E. Vanden-Eijnden. Building normalizing flows with stochastic interpolants. arXiv preprint arXiv:2209.15571, 2022

    work page internal anchor Pith review arXiv 2022

  3. [3]

    F. Bao, S. Nie, K. Xue, Y . Cao, C. Li, H. Su, and J. Zhu. All are worth words: A vit backbone for diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22669–22679, 2023

    work page 2023

  4. [4]

    J. Baxter. A model of inductive bias learning.Journal of artificial intelligence research, 12:149–198, 2000

    work page 2000

  5. [5]

    Bergen, M

    V . Bergen, M. Lange, S. Peidli, F. A. Wolf, and F. J. Theis. Generalizing rna velocity to transient cell states through dynamical modeling.Nature biotechnology, 38(12):1408–1414, 2020

    work page 2020

  6. [6]

    Bleile, S

    F. Bleile, S. Lumpp, and M. Drton. Efficient learning of stationary diffusions with stein-type discrepancies.arXiv preprint arXiv:2601.16597, 2026

    work page arXiv 2026

  7. [7]

    Chartrand

    R. Chartrand. Numerical differentiation of noisy, nonsmooth data.International Scholarly Research Notices, 2011(1):164564, 2011

    work page 2011

  8. [8]

    R. T. Chen, Y . Rubanova, J. Bettencourt, and D. K. Duvenaud. Neural ordinary differential equations.Advances in neural information processing systems, 31, 2018

    work page 2018

  9. [9]

    C. Chi, Z. Xu, S. Feng, E. Cousineau, Y . Du, B. Burchfiel, R. Tedrake, and S. Song. Diffusion policy: Visuomotor policy learning via action diffusion.The International Journal of Robotics Research, 44(10-11):1684–1704, 2025

    work page 2025

  10. [10]

    K. Choi, C. Meng, Y . Song, and S. Ermon. Density ratio estimation via infinitesimal classifica- tion. InInternational Conference on Artificial Intelligence and Statistics, pages 2552–2573. PMLR, 2022

    work page 2022

  11. [11]

    Diffdock: Diffusion steps, twists, and turns for molecular docking.arXiv preprint arXiv:2210.01776, 2022

    G. Corso, H. Stärk, B. Jing, R. Barzilay, and T. Jaakkola. Diffdock: Diffusion steps, twists, and turns for molecular docking.arXiv preprint arXiv:2210.01776, 2022

    work page arXiv 2022

  12. [12]

    Dhariwal and A

    P. Dhariwal and A. Nichol. Diffusion models beat gans on image synthesis.Advances in neural information processing systems, 34:8780–8794, 2021

    work page 2021

  13. [13]

    L. Dinh, J. Sohl-Dickstein, and S. Bengio. Density estimation using real nvp.arXiv preprint arXiv:1605.08803, 2016

    work page internal anchor Pith review arXiv 2016

  14. [14]

    Dixit, O

    A. Dixit, O. Parnas, B. Li, J. Chen, C. P. Fulco, L. Jerby-Arnon, N. D. Marjanovic, D. Dionne, T. Burks, R. Raychowdhury, et al. Perturb-seq: dissecting molecular circuits with scalable single-cell rna profiling of pooled genetic screens.cell, 167(7):1853–1866, 2016

    work page 2016

  15. [15]

    Du and I

    Y . Du and I. Mordatch. Implicit generation and modeling with energy based models.Advances in neural information processing systems, 32, 2019

    work page 2019

  16. [16]

    A. B. Duncan, T. Lelievre, and G. A. Pavliotis. Variance reduction using nonreversible langevin samplers.Journal of statistical physics, 163(3):457–491, 2016. 10

    work page 2016

  17. [17]

    A. B. Duncan, G. A. Pavliotis, and K. Zygalakis. Nonreversible langevin samplers: Splitting schemes, analysis and implementation.arXiv preprint arXiv:1701.04247, 2017

    work page arXiv 2017

  18. [18]

    Fishman, L

    N. Fishman, L. Klarner, E. Mathieu, M. Hutchinson, and V . De Bortoli. Metropolis sampling for constrained diffusion models.Advances in Neural Information Processing Systems, 36:62296– 62331, 2023

    work page 2023

  19. [19]

    Diffusion: Minimal multi-gpu implementation of diffusion models with classifier-free guidance (cfg)

    FutureXiang. Diffusion: Minimal multi-gpu implementation of diffusion models with classifier-free guidance (cfg). https://github.com/FutureXiang/Diffusion/tree/ master, 2023

    work page 2023

  20. [20]

    Z. Geng, M. Deng, X. Bai, J. Z. Kolter, and K. He. Mean flows for one-step generative modeling. arXiv preprint arXiv:2505.13447, 2025

    work page internal anchor Pith review arXiv 2025

  21. [21]

    S. Gong, M. Li, J. Feng, Z. Wu, and L. Kong. Diffuseq: Sequence to sequence text generation with diffusion models.arXiv preprint arXiv:2210.08933, 2022

    work page arXiv 2022

  22. [22]

    Gutmann and A

    M. Gutmann and A. Hyvärinen. Noise-contrastive estimation: A new estimation principle for unnormalized statistical models. InProceedings of the thirteenth international conference on artificial intelligence and statistics, pages 297–304. JMLR Workshop and Conference Proceedings, 2010

    work page 2010

  23. [23]

    Hansen and A

    N. Hansen and A. Sokol. Causal interpretation of stochastic differential equations. 2014

    work page 2014

  24. [24]

    G. E. Hinton. Training products of experts by minimizing contrastive divergence.Neural computation, 14(8):1771–1800, 2002

    work page 2002

  25. [25]

    J. Ho, A. Jain, and P. Abbeel. Denoising diffusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    work page 2020

  26. [26]

    J. Ho, T. Salimans, A. Gritsenko, W. Chan, M. Norouzi, and D. J. Fleet. Video diffusion models. Advances in neural information processing systems, 35:8633–8646, 2022

    work page 2022

  27. [27]

    Horvat and J.-P

    C. Horvat and J.-P. Pfister. On gauge freedom, conservativity and intrinsic dimensionality estimation in diffusion models.arXiv preprint arXiv:2402.03845, 2024

    work page arXiv 2024

  28. [28]

    Huang, T

    Y . Huang, T. Transue, S.-H. Wang, W. Feldman, H. Zhang, and B. Wang. Improving flow matching by aligning flow divergence.arXiv preprint arXiv:2602.00869, 2026

    work page arXiv 2026

  29. [29]

    M. F. Hutchinson. A stochastic estimator of the trace of the influence matrix for laplacian smoothing splines.Communications in Statistics-Simulation and Computation, 18(3):1059– 1076, 1989

    work page 1989

  30. [30]

    Hwang, S.-Y

    C.-R. Hwang, S.-Y . Hwang-Ma, and S.-J. Sheu. Accelerating diffusions. 2005

    work page 2005

  31. [31]

    Hyvärinen and P

    A. Hyvärinen and P. Dayan. Estimation of non-normalized statistical models by score matching. Journal of Machine Learning Research, 6(4), 2005

    work page 2005

  32. [32]

    B. Jing, G. Corso, J. Chang, R. Barzilay, and T. Jaakkola. Torsional diffusion for molecular conformer generation.Advances in neural information processing systems, 35:24240–24253, 2022

    work page 2022

  33. [33]

    Karras, M

    T. Karras, M. Aittala, T. Aila, and S. Laine. Elucidating the design space of diffusion-based generative models.Advances in neural information processing systems, 35:26565–26577, 2022

    work page 2022

  34. [34]

    Karras, M

    T. Karras, M. Aittala, J. Lehtinen, J. Hellsten, T. Aila, and S. Laine. Analyzing and improving the training dynamics of diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 24174–24184, 2024

    work page 2024

  35. [35]

    Z. Kong, W. Ping, J. Huang, K. Zhao, and B. Catanzaro. Diffwave: A versatile diffusion model for audio synthesis.arXiv preprint arXiv:2009.09761, 2020

    work page arXiv 2009

  36. [36]

    La Manno, R

    G. La Manno, R. Soldatov, A. Zeisel, E. Braun, H. Hochgerner, V . Petukhov, K. Lidschreiber, M. E. Kastriti, P. Lönnerberg, A. Furlan, et al. Rna velocity of single cells.Nature, 560(7719):494–498, 2018. 11

    work page 2018

  37. [37]

    C.-H. Lai, B. Nguyen, N. Murata, Y . Takida, T. Uesaka, Y . Mitsufuji, S. Ermon, and M. Tao. A unified view of drifting and score-based models.arXiv preprint arXiv:2603.07514, 2026

    work page arXiv 2026

  38. [38]

    C.-H. Lai, Y . Takida, N. Murata, T. Uesaka, Y . Mitsufuji, and S. Ermon. Fp-diffusion: Improving score-based diffusion models by enforcing the underlying score fokker-planck equation. In International Conference on Machine Learning, pages 18365–18398. PMLR, 2023

    work page 2023

  39. [39]

    LeCun, S

    Y . LeCun, S. Chopra, R. Hadsell, M. Ranzato, F. Huang, et al. A tutorial on energy-based learning.Predicting structured data, 1(0), 2006

    work page 2006

  40. [40]

    Flow Matching for Generative Modeling

    Y . Lipman, R. T. Chen, H. Ben-Hamu, M. Nickel, and M. Le. Flow matching for generative modeling.arXiv preprint arXiv:2210.02747, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  41. [41]

    X. Liu, C. Gong, and Q. Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow.arXiv preprint arXiv:2209.03003, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  42. [42]

    Lorch, A

    L. Lorch, A. Krause, and B. Schölkopf. Causal modeling with stationary diffusions. In International Conference on Artificial Intelligence and Statistics, pages 1927–1935. PMLR, 2024

    work page 1927

  43. [43]

    Lou and S

    A. Lou and S. Ermon. Reflected diffusion models. InInternational Conference on Machine Learning, pages 22675–22701. PMLR, 2023

    work page 2023

  44. [44]

    Y .-A. Ma, T. Chen, and E. Fox. A complete recipe for stochastic gradient mcmc.Advances in neural information processing systems, 28, 2015

    work page 2015

  45. [45]

    Neklyudov, R

    K. Neklyudov, R. Brekelmans, D. Severo, and A. Makhzani. Action matching: Learning stochastic dynamics from samples. InInternational conference on machine learning, pages 25858–25889. PMLR, 2023

    work page 2023

  46. [46]

    arXiv preprint arXiv:2310.10649 , year=

    K. Neklyudov, R. Brekelmans, A. Tong, L. Atanackovic, Q. Liu, and A. Makhzani. A computa- tional framework for solving wasserstein lagrangian flows.arXiv preprint arXiv:2310.10649, 2023

    work page arXiv 2023

  47. [47]

    Papamakarios, E

    G. Papamakarios, E. Nalisnick, D. J. Rezende, S. Mohamed, and B. Lakshminarayanan. Nor- malizing flows for probabilistic modeling and inference.Journal of Machine Learning Research, 22(57):1–64, 2021

    work page 2021

  48. [48]

    G. A. Pavliotis. Stochastic processes and applications.Texts in applied mathematics, 60:41–43, 2014

    work page 2014

  49. [49]

    G. A. Pavliotis and A. M. Stuart. Multiscale methods, volume 53 of texts in applied mathematics, 2008

    work page 2008

  50. [50]

    Imitating human behaviour with diffusion models

    T. Pearce, T. Rashid, A. Kanervisto, D. Bignell, M. Sun, R. Georgescu, S. V . Macua, S. Z. Tan, I. Momennejad, K. Hofmann, et al. Imitating human behaviour with diffusion models.arXiv preprint arXiv:2301.10677, 2023

    work page arXiv 2023

  51. [51]

    Peebles and S

    W. Peebles and S. Xie. Scalable diffusion models with transformers. InProceedings of the IEEE/CVF international conference on computer vision, pages 4195–4205, 2023

    work page 2023

  52. [52]

    arXiv preprint arXiv:2510.26645 , year=

    K. Petrovi´c, L. Atanackovic, V . Moro, K. Kapu´sniak, I. I. Ceylan, M. Bronstein, A. J. Bose, and A. Tong. Curly flow matching for learning non-gradient field dynamics.arXiv preprint arXiv:2510.26645, 2025

    work page arXiv 2025

  53. [53]

    Rey-Bellet and K

    L. Rey-Bellet and K. Spiliopoulos. Irreversible langevin samplers and variance reduction: a large deviations approach.Nonlinearity, 28(7):2081–2103, 2015

    work page 2081

  54. [54]

    Rombach, A

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022

    work page 2022

  55. [55]

    P. K. Rubenstein, S. Bongers, B. Schölkopf, and J. M. Mooij. From deterministic odes to dynamic structural causal models.arXiv preprint arXiv:1608.08028, 2016. 12

    work page arXiv 2016

  56. [56]

    Sohl-Dickstein, E

    J. Sohl-Dickstein, E. Weiss, N. Maheswaranathan, and S. Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. InInternational conference on machine learning, pages 2256–2265. pmlr, 2015

    work page 2015

  57. [57]

    Y . Song, P. Dhariwal, M. Chen, and I. Sutskever. Consistency models. 2023

    work page 2023

  58. [58]

    Song and S

    Y . Song and S. Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019

    work page 2019

  59. [59]

    Y . Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole. Score-based generative modeling through stochastic differential equations.arXiv preprint arXiv:2011.13456, 2020

    work page internal anchor Pith review Pith/arXiv arXiv 2011

  60. [60]

    A. Tong, K. Fatras, N. Malkin, G. Huguet, Y . Zhang, J. Rector-Brooks, G. Wolf, and Y . Bengio. Improving and generalizing flow-based generative models with minibatch optimal transport. arXiv preprint arXiv:2302.00482, 2023

    work page internal anchor Pith review arXiv 2023

  61. [61]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin. Attention is all you need.Advances in neural information processing systems, 30, 2017

    work page 2017

  62. [62]

    P. Vincent. A connection between score matching and denoising autoencoders.Neural compu- tation, 23(7):1661–1674, 2011

    work page 2011

  63. [63]

    J. L. Watson, D. Juergens, N. R. Bennett, B. L. Trippe, J. Yim, H. E. Eisenach, W. Ahern, A. J. Borst, R. J. Ragotte, L. F. Milles, et al. De novo design of protein structure and function with rfdiffusion.Nature, 620(7976):1089–1100, 2023

    work page 2023

  64. [64]

    M. Xu, L. Yu, Y . Song, C. Shi, S. Ermon, and J. Tang. Geodiff: A geometric diffusion model for molecular conformation generation.arXiv preprint arXiv:2203.02923, 2022

    work page arXiv 2022

  65. [65]

    Y . Xu, S. Tong, and T. Jaakkola. Stable target field for reduced variance score estimation in diffusion models.arXiv preprint arXiv:2302.00670, 2023

    work page arXiv 2023

  66. [66]

    bone_marrow

    Y . Zhang and M. Levin. Equilibrium flow: from snapshots to dynamics.arXiv preprint arXiv:2509.17990, 2025. 13 Appendix Overview A Proofs 15 B Experiment Details 18 B.1 Application 1:ControllableGenerative Fields . . . . . . . . . . . . . . . . . . . . 18 B.2 Application 2:InterpretableGenerative Fields . . . . . . . . . . . . . . . . . . . . 21 B.3 Appli...

    work page arXiv 2025

  67. [67]

    Crucially, their method cannot learn non-gradient fields given a single distribution

    frame their method as learning non-gradient field dynamics, yet it fundamentally matches a prior to a terminal distribution via a learned (rather than predefined) interpolation, a special case of [46]. Crucially, their method cannot learn non-gradient fields given a single distribution. The two approaches are in fact complementary, with [52] producing a b...

    work page

  68. [68]

    In contrast, Flux Matching is a standalone generative objective: [38, 28] add a regularizer to a generative loss, whereas Flux Matchingisthe generative loss

    and [28] enforce the Fokker–Planck equation (respectively, the continuity equation) as a reg- ularizer on top of a primary score matching or flow matching objective, providing tighter control over the induced PDE. In contrast, Flux Matching is a standalone generative objective: [38, 28] add a regularizer to a generative loss, whereas Flux Matchingisthe ge...

    work page