Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

This paper proves that rectified flow—a method that learns a deterministic ODE from noise to data—can be lifted to infinite-dimensional Hilbert spaces, and that the induced flow preserves the marginal distribution of the original stochastic

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Functional rectified flow is defined and proved to preserve marginals in separable Hilbert spaces, with functional flow matching and probability-flow ODEs as special cases.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection Solid conditional theory for rectified flow in Hilbert space; the main theorem is genuine, but Assumption 2 is never checked in experiments and the empirical gains are thinner than claimed. the 3 major comments →

arxiv 2509.10384 v1 pith:MZ3MWVEG submitted 2025-09-12 cs.LG stat.ML

Flow Straight and Fast in Hilbert Space: Functional Rectified Flow

classification cs.LG stat.ML
keywords rectified flowfunctional generative modelinfinite-dimensional Hilbert spacemarginal preservationsuperposition principleexpected velocityflow matchingstraightening effect
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

Rectified flow is a generative technique that learns a deterministic ordinary differential equation (ODE) to turn noise into data. This paper extends it to infinite-dimensional Hilbert spaces, where data points are functions like images, time series, or PDE solutions. The core result is a theorem stating that if the original stochastic process is 'rectifiable'—roughly, its expected velocity is integrable and defines a well-posed ODE—then the learned flow reproduces the process's marginal distributions at every time t. That gives a principled way to sample functional data with a single ODE solve, avoiding stochastic reverse diffusion. The paper also shows that existing functional flow matching and probability-flow-ODE methods are special cases, and reports strong sample quality on image and fluid-dynamics benchmarks.

Core claim

The central claim is Theorem 5 (restated as Theorem 20 in the appendix): for any H-valued stochastic process X that is rectifiable, the rectified flow Z defined by Z_t = Z_0 + ∫₀ᵗ v_X(s, Z_s) ds, with v_X(t,x) = E[Ẋ_t | X_t = x] and Z_0 ~ X_0, has the same marginal distribution as X_t for every t ∈ [0,1]. The proof builds a probability measure over paths from the continuity equation via the superposition principle in infinite-dimensional Hilbert space, then shows this measure must equal the law of the ODE solution. A key step is showing the expected velocity field v_X solves the continuity equation for the marginal measures; the superposition principle then yields a path measure with those m

What carries the argument

The engine is the expected velocity field v_X(t,x) = E[Ẋ_t | X_t = x], the conditional expectation of the process's time derivative given its current value. The paper proves that the pair (v_X, µ_t), where µ_t is the marginal law of X_t, satisfies the continuity equation; then an infinite-dimensional superposition principle guarantees there is a distribution over trajectories solving the ODE ẋ = v_X(t,x) that has µ_t as its time-marginals. Uniqueness of the initial-value problem then identifies that trajectory distribution with the rectified flow. This bypasses the absolute-continuity condition needed in prior functional flow-matching theory.

Load-bearing premise

The whole marginal-preservation guarantee rests on the assumption that the ordinary differential equation driven by the expected velocity has a unique solution and a continuous solution map; the paper itself notes that ordinary neural-network velocity models are not guaranteed to satisfy this.

What would settle it

For a trained functional rectified-flow model, estimate the Lipschitz constant of the learned velocity field on the empirical support of the data (for instance, by measuring Jacobian spectral norms at many random test functions). If the velocity is not Lipschitz—or the continuity of the solution map fails—then the theorem's Assumption 2 is violated, meaning the deployed sampler is not covered by the paper's marginal-preservation guarantee.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Deterministic sampling from functional generative models becomes theoretically grounded: one numerical ODE integration with a random initial condition draws from the target functional distribution, avoiding reverse-SDE simulation.
  • Functional flow matching and functional probability-flow ODE are unified as nonlinear rectified flows, so algorithms and intuitions transfer across these methods.
  • Transport-cost reduction and the straightening effect hold in Hilbert space, meaning repeated rectification can make functional trajectories more nearly straight and enable few-step or one-step sampling.
  • The framework works with point-evaluation-based architectures—implicit neural representations, transformers, and neural operators—so it applies to images, time series, and PDE fields without specialized training loops.
  • The restrictive absolute-continuity assumption of prior functional flow matching is removed, allowing a larger class of couplings and interpolation paths.

Where Pith is reading between the lines

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

  • The same rectification recipe could be applied to stochastic processes arising in neural SDEs or Bayesian inverse problems, giving deterministic samplers with preserved finite-dimensional distributions; the paper does not test this.
  • Because the theorem does not require Gaussian noise or independence between X0 and X1, the method may extend naturally to conditional generation (e.g., PDE surrogate models) by using paired function samples as endpoints; this is left implicit.
  • The straightening bound suggests a practical stopping rule: measure the variance term V and stop rectifying when it is near zero, which the paper does not implement.
  • A useful diagnostic, not discussed in the paper, is to check the Lipschitz constant of the trained velocity; if it is not finite, practitioners can add spectral normalization or Jacobian regularization to bring the deployed model back inside Assumption 2.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper develops a functional generalization of rectified flow for separable Hilbert spaces. It defines the expected velocity of a pathwise differentiable H-valued process, introduces a rectifiability assumption (Assumptions 2 and 3), and proves that the induced deterministic rectified flow preserves the marginal distributions of the original process (Theorem 5 / Theorem 20). The proof proceeds through a continuity-equation/superposition-principle argument in infinite dimensions. The paper further shows that functional flow matching (Kerrigan et al.) and the functional probability-flow ODE (Na et al.) can be viewed as nonlinear rectified flows, and reports experiments on MNIST, CelebA, and Navier-Stokes data using INR, transformer, and neural-operator backbones.

Significance. If the theorem is read as a conditional statement, this is a solid and useful contribution: it provides a rigorous infinite-dimensional analogue of rectified flow, with a detailed appendix proof that is internally consistent and avoids the absolute-continuity condition used by Kerrigan et al. The unification of functional flow matching and functional probability-flow ODE is conceptually valuable. However, the central marginal-preservation guarantee is conditional on Assumption 2, which is stated for the exact expected velocity v_X and is not verified for any trained network used in the experiments. The empirical section also lacks error bars for the two main image benchmarks, and the reported MNIST gap is very small. These gaps sit at the boundary between the theory and the paper's stronger claims, but they are local rather than fatal: the core derivation is sound, and the issues can be addressed by enforcing/checking the assumption and by more careful empirical reporting.

major comments (3)
  1. [Section 3.1, Assumption 2; Proposition 18; Section 4.1] Theorem 20 is explicitly conditional on Assumption 2, which requires the IVP driven by the exact expected velocity v_X to admit a unique solution with a continuous solution map. The only sufficient condition supplied is Proposition 18, requiring v_X to be globally Lipschitz with an integrable Lipschitz constant. The paper itself notes that generic neural networks need not satisfy this condition, and the experimental appendix (Section C) does not mention spectral normalization, Jacobian regularization, or any other Lipschitz enforcement. Moreover, for the linear interpolation X_t = tX_1+(1-t)X_0, v_X can be discontinuous where interpolation segments collide, so Assumption 2 may fail in the tested settings. Thus the claim in Section 4.1 that the framework 'removes restrictive assumptions' and replaces them with 'more plausible and readily satisfied' conditions is overstated: the absolute-c
  2. [Section 6, Tables 2 and 3] The superiority claims rest on single runs without error bars. On MNIST, FRF achieves FID 0.41 versus FDP 0.43; this difference is well within typical run-to-run variability and cannot support the claim of improved sample quality. The CelebA differences are larger, but without multiple seeds and confidence intervals they remain point estimates. I recommend reporting mean +/- std over at least three seeds, and for completeness reporting all metrics for all baselines in each table (several entries are currently missing, e.g., Inf-Diff FID and FD2F FID-CLIP).
  3. [Section 4.2, Proposition 7] The proposition states that Equation (6) 'is the rectified flow induced by the process Y'_t', but as written Equation (6) is solved backward in time with terminal condition Y_1 ~ N(0,Q), whereas Definition 4 evolves forward from an initial condition Z_0 ~ X_0. The following sentence then correctly describes Equation (6) as the time-reversal of the nonlinear rectified flow. This is an inconsistency in the main statement. Please restate Proposition 7 in terms of a time-reversed rectified flow, or introduce a forward-time convention that makes the direction of integration unambiguous.
minor comments (4)
  1. [Section 5] The statement that elements of L2(M) are characterized by pointwise evaluations relies on 'Theorem 2 in Franzese et al.', but this theorem is not stated. Please include the precise conditions, since not every L2 function is determined by pointwise values.
  2. [Table 1] The VP path is listed with alpha_t 'arbitrary in [0,1]'. If alpha_t is truly arbitrary, it may not be continuously differentiable; please specify the regularity required for the entries of the table.
  3. [Section 3.1] The extension v_X(t,x)=0 outside supp(X_t) can create discontinuities at the boundary of the support. This is not itself an error if Assumption 2 is checked, but it is worth a remark because it interacts with the well-posedness assumption.
  4. [Abstract and Section 1] The abstract says the framework 'removes restrictive measure-theoretic assumptions'; given Assumption 2, the abstract should be phrased as replacing those assumptions with a well-posedness condition, not removing them entirely.

Circularity Check

0 steps flagged

No significant circularity: the main theorem is a conditional mathematical result derived from an external superposition principle and explicit well-posedness assumptions, not a restatement of its inputs.

full rationale

The central claim (Theorem 5/20) is that the rectified flow Z, defined as the unique solution of z(t)=u+∫_0^t v_X(s,z(s))ds with u∼X_0, has the same marginals as X. This is not assumed: it is proved by (i) observing that the marginal curve μ_t solves the continuity equation with v_X (an algebraic consequence of the definition of conditional expectation, Eqs. (22)-(23) in Appendix A.4), (ii) invoking the Stepanov–Trevisan superposition principle to produce a path measure with marginals μ_t, and (iii) using Assumption 2 (unique IVP solution and continuous solution map) in Theorem 19 to identify that path measure with the law of the rectified flow. The well-posedness Assumption 2 is an explicit hypothesis, not a conclusion smuggled in; the paper even gives Proposition 18 as a sufficient Lipschitz condition and notes generic neural networks need not satisfy it. That caveat makes the experimental validation of the formal guarantee weaker, but it is a correctness/robustness limitation, not circularity. The connections to functional flow matching and probability-flow ODEs are algebraic identifications (Table 1, Proposition 7), and no fitted parameter is renamed as a prediction. No load-bearing self-citation appears; cited external results (Stepanov–Trevisan, Da Prato–Zabczyk, etc.) are standard and independent.

Axiom & Free-Parameter Ledger

0 free parameters · 5 axioms · 0 invented entities

The central theoretical claims introduce no fitted parameters. They rest on standard results (superposition principle, Doob-Dynkin lemma, Gaussian measure theory) and on the paper's explicit regularity assumptions (Assumptions 2 and 3), which are not verified for the trained networks.

axioms (5)
  • standard math Superposition principle in Hilbert space (Theorem 3.4 of Stepanov and Trevisan [75])
    Invoked in Appendix A.3 to lift continuity equations to ODE path measures; the paper verifies its hypotheses.
  • standard math Existence of conditional expectation in separable Banach spaces and the Doob-Dynkin lemma
    Used in Appendix A.1 to define the expected velocity field v_X.
  • ad hoc to paper Assumption 2: IVP (1) with v_X admits a unique solution with continuous solution map Φ
    Introduced specifically for this work; not guaranteed for arbitrary distributions, and not verified for the trained networks in experiments.
  • domain assumption Assumption 3: integrability of v_X and E sup_t ||X_dot_t|| < infinity
    Mild regularity condition needed to justify interchange of derivative and expectation; holds for C^1 interpolation paths with finite first moments.
  • domain assumption Pathwise continuous differentiability of the process X
    Needed to define X_dot_t and apply the chain rule; holds automatically for the linear interpolation paths (3) and (5) with C^1 coefficients.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Flow Straight and Fast in Hilbert Space: Functional Rectified Flow." pith.science (2026). https://pith.science/paper/MZ3MWVEG

@misc{pith2026250910384,
  author       = {Pith},
  title        = {Pith review of: Flow Straight and Fast in Hilbert Space: Functional Rectified Flow},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MZ3MWVEG}},
  note         = {Machine review of arXiv:2509.10384}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Many generative models originally developed in finite-dimensional Euclidean space have functional generalizations in infinite-dimensional settings. However, the extension of rectified flow to infinite-dimensional spaces remains unexplored. In this work, we establish a rigorous functional formulation of rectified flow in an infinite-dimensional Hilbert space. Our approach builds upon the superposition principle for continuity equations in an infinite-dimensional space. We further show that this framework extends naturally to functional flow matching and functional probability flow ODEs, interpreting them as nonlinear generalizations of rectified flow. Notably, our extension to functional flow matching removes the restrictive measure-theoretic assumptions in the existing theory of \citet{kerrigan2024functional}. Furthermore, we demonstrate experimentally that our method achieves superior performance compared to existing functional generative models.

Figures

Figures reproduced from arXiv: 2509.10384 by Clayton Scott, Jianxin Zhang.

Figure 1
Figure 1. Figure 1: Qualitative results on MNIST: (a) samples generated at the original [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Qualitative results of functional rectified flow with vision transformer. [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Additional CelebA samples generated by FRF. [PITH_FULL_IMAGE:figures/full_fig_p032_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Additional CelebA samples generated by FRF. [PITH_FULL_IMAGE:figures/full_fig_p033_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Additional CelebA samples generated by FRF. [PITH_FULL_IMAGE:figures/full_fig_p033_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Additional CelebA samples generated by FRF. [PITH_FULL_IMAGE:figures/full_fig_p034_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Additional CelebA samples generated by FRF. [PITH_FULL_IMAGE:figures/full_fig_p034_7.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Optimal-Transport-Guided Functional Flow Matching for Turbulent Field Generation in Hilbert Space

    cs.LG 2026-04 unverdicted novelty 6.0

    FOT-CFM generates turbulent fields in function space with superior high-order statistics and energy spectra on Navier-Stokes, Kolmogorov flow, and Hasegawa-Wakatani equations compared to baselines.

Reference graph

Works this paper leans on

89 extracted references · 15 linked inside Pith · cited by 1 Pith paper

  1. [1]

    (2024).Lectures on Optimal Transport

    Ambrosio, L., Brué, E., and Semola, D. (2024).Lectures on Optimal Transport. UNITEXT. Springer Cham, 2 edition. eBook published: 28 December 2024; Softcover published: 29 December 2024

  2. [2]

    (2008).Gradient Flows: In Metric Spaces and in the Space of Probability Measures

    Ambrosio, L., Gigli, N., and Savaré, G. (2008).Gradient Flows: In Metric Spaces and in the Space of Probability Measures. Lectures in Mathematics. ETH Zürich. Birkhäuser Basel, 2 edition. Published: 13 March 2008 (softcover), 29 October 2008 (eBook)

  3. [3]

    Bao, F., Nie, S., Xue, K., Cao, Y ., Li, C., Su, H., and Zhu, J. (2023). All are worth words: A vit backbone for diffusion models. InCVPR

  4. [4]

    Bogachev, V . I. (1998).Gaussian Measures, volume 62 ofMathematical Surveys and Mono- graphs. American Mathematical Society, Providence, RI. Primary 60B11; Secondary 28C20, 46G12, 60G15, 60H07

  5. [5]

    and Willcocks, C

    Bond-Taylor, S. and Willcocks, C. G. (2024). $\infty$-diff: Infinite resolution diffusion with subsampled mollified states. InThe Twelfth International Conference on Learning Representations

  6. [6]

    J., Leary, C., Maclaurin, D., Necula, G., Paszke, A., VanderPlas, J., Wanderman-Milne, S., et al

    Bradbury, J., Frostig, R., Hawkins, P., Johnson, M. J., Leary, C., Maclaurin, D., Necula, G., Paszke, A., VanderPlas, J., Wanderman-Milne, S., et al. (2018). Jax: composable transformations of python+ numpy programs

  7. [7]

    Cao, S. (2021). Choose a transformer: Fourier or galerkin. In Beygelzimer, A., Dauphin, Y ., Liang, P., and Vaughan, J. W., editors,Advances in Neural Information Processing Systems

  8. [8]

    Chen, R. T. Q., Rubanova, Y ., Bettencourt, J., and Duvenaud, D. (2018). Neural ordinary differential equations. InAdvances in Neural Information Processing Systems, volume 31

  9. [9]

    Cohn, D. L. (2013).Measure Theory, volume XXI, 457 ofBirkhäuser Advanced Texts Basler Lehrbücher. Birkhäuser, New York, NY , 2 edition. Springer Science+Business Media New York 2013

  10. [10]

    (2012).Calculus on Normed Vector Spaces

    Coleman, R. (2012).Calculus on Normed Vector Spaces. Universitext. Springer, New York, NY , 1 edition. Published: 24 July 2012 (print), 25 July 2012 (online)

  11. [11]

    Conway, J. B. (1990).A Course in Functional Analysis, volume 96 ofGraduate Texts in Mathematics. Springer New York, NY , 2 edition. eBook ISBN: 978-1-4757-4383-8; Softcover ISBN: 978-1-4419-3092-7 (2010); Copyright © Springer-Verlag New York 2007

  12. [12]

    and Zabczyk, J

    Da Prato, G. and Zabczyk, J. (1992).Probability measures, page 30–69. Encyclopedia of Mathematics and its Applications. Cambridge University Press

  13. [13]

    and Zabczyk, J

    Da Prato, G. and Zabczyk, J. (2014).Stochastic Equations in Infinite Dimensions. Encyclopedia of Mathematics and its Applications. Cambridge University Press, 2 edition

  14. [14]

    Dalva, Y ., Venkatesh, K., and Yanardag, P. (2024). Fluxspace: Disentangled semantic editing in rectified flow transformers.arXiv preprint arXiv:2412.09611

  15. [15]

    and Nichol, A

    Dhariwal, P. and Nichol, A. Q. (2021). Diffusion models beat GANs on image synthesis. In Beygelzimer, A., Dauphin, Y ., Liang, P., and Vaughan, J. W., editors,Advances in Neural Information Processing Systems

  16. [16]

    Dupont, E., Kim, H., Eslami, S. M. A., Rezende, D. J., and Rosenbaum, D. (2022). From data to functa: Your data point is a function and you can treat it like one. In39th International Conference on Machine Learning (ICML)

  17. [17]

    Dutordoir, V ., Saul, A., Ghahramani, Z., and Simpson, F. (2023). Neural diffusion processes. In International Conference on Machine Learning, pages 8990–9012. PMLR

  18. [18]

    Evans, L. C. (2010).Partial Differential Equations, volume 19 ofGraduate Studies in Mathe- matics. American Mathematical Society, Providence, RI, 2nd edition. 10

  19. [19]

    Finn, C., Abbeel, P., and Levine, S. (2017). Model-agnostic meta-learning for fast adaptation of deep networks. InInternational conference on machine learning, pages 1126–1135. PMLR

  20. [20]

    Franzese, G., Corallo, G., Rossi, S., Heinonen, M., Filippone, M., and Michiardi, P. (2023). Continuous-time functional diffusion processes. InThirty-seventh Conference on Neural Informa- tion Processing Systems

  21. [21]

    Go, H., Park, B., Jang, J., Kim, J.-Y ., Kwon, S., and Kim, C. (2024). Splatflow: Multi-view rectified flow model for 3d gaussian splatting synthesis.arXiv preprint arXiv:2411.16443

  22. [22]

    Goel, K., Gu, A., Donahue, C., and Ré, C. (2022). It’s raw! audio generation with state-space models.International Conference on Machine Learning (ICML)

  23. [23]

    J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., and Bengio, Y

    Goodfellow, I. J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., and Bengio, Y . (2014). Generative adversarial networks

  24. [24]

    Gouk, H., Frank, E., Pfahringer, B., and Cree, M. J. (2021). Regularisation of neural networks by enforcing lipschitz continuity.Machine Learning, 110(2):393–416

  25. [25]

    Guan, W., Su, Q., Zhou, H., Miao, S., Xie, X., Li, L., and Hong, Q. (2024). Reflow-tts: A rectified flow model for high-fidelity text-to-speech. InICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 10501–10505. IEEE

  26. [26]

    Guo, Y ., Du, C., Ma, Z., Chen, X., and Yu, K. (2024). V oiceflow: Efficient text-to-speech with rectified flow matching. InICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 11121–11125. IEEE

  27. [27]

    Hagemann, P., Mildenberger, S., Ruthotto, L., Steidl, G., and Yang, N. T. (2023). Multilevel diffusion: Infinite dimensional score-based diffusion models for image generation.arXiv preprint arXiv:2303.04772

  28. [28]

    and Burger, M

    Helin, T. and Burger, M. (2015). Maximum a posteriori probability estimates in infinite- dimensional bayesian inverse problems.Inverse Problems, 31(8):085009

  29. [29]

    Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., and Hochreiter, S. (2017). Gans trained by a two time-scale update rule converge to a local nash equilibrium.Advances in neural information processing systems, 30

  30. [30]

    P., Poole, B., Norouzi, M., Fleet, D

    Ho, J., Chan, W., Saharia, C., Whang, J., Gao, R., Gritsenko, A., Kingma, D. P., Poole, B., Norouzi, M., Fleet, D. J., and Salimans, T. (2022). Imagen video: High definition video generation with diffusion models

  31. [31]

    Ho, J., Jain, A., and Abbeel, P. (2020). Denoising diffusion probabilistic models.arXiv preprint arxiv:2006.11239

  32. [32]

    and Nguyen, D

    Holt, W. and Nguyen, D. (2023). Essential aspects of bayesian data imputation. SSRN. Available at SSRN: https://ssrn.com/abstract=4494314 or http://dx.doi.org/10.2139/ssrn. 4494314

  33. [33]

    Issa, Z., Horvath, B., Lemercier, M., and Salvi, C. (2023). Non-adversarial training of neural SDEs with signature kernel scores. InThirty-seventh Conference on Neural Information Processing Systems

  34. [34]

    (2021).Foundations of Modern Probability

    Kallenberg, O. (2021).Foundations of Modern Probability. Number 75 in Probability Theory and Stochastic Modelling. Springer Cham, 3 edition

  35. [35]

    Kang, M., Zhu, J.-Y ., Zhang, R., Park, J., Shechtman, E., Paris, S., and Park, T. (2023). Scaling up gans for text-to-image synthesis. In2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10124–10134

  36. [36]

    Kerrigan, G., Ley, J., and Smyth, P. (2023). Diffusion generative models in infinite dimensions. In Ruiz, F., Dy, J., and van de Meent, J.-W., editors,Proceedings of The 26th International Conference on Artificial Intelligence and Statistics, volume 206 ofProceedings of Machine Learning Research, pages 9538–9563. PMLR. 11

  37. [37]

    Kerrigan, G., Migliorini, G., and Smyth, P. (2024). Functional flow matching. InThe 27th International Conference on AI and Statistics (AISTATS)

  38. [38]

    Kidger, P., Foster, J., Li, X., Oberhauser, H., and Lyons, T. (2021). Neural sdes as infinite- dimensional gans

  39. [39]

    Kong, Z., Ping, W., Huang, J., Zhao, K., and Catanzaro, B. (2021). Diffwave: A versatile diffusion model for audio synthesis. InInternational Conference on Learning Representations

  40. [40]

    J., Bonev, B., Aziz- zadenesheli, K., Berner, J., and Anandkumar, A

    Kossaifi, J., Kovachki, N., Li, Z., Pitt, D., Liu-Schiaffini, M., George, R. J., Bonev, B., Aziz- zadenesheli, K., Berner, J., and Anandkumar, A. (2024). A library for learning neural operators

  41. [41]

    B., Li, Z., Liu, B., Azizzadenesheli, K., Bhattacharya, K., Stuart, A

    Kovachki, N. B., Li, Z., Liu, B., Azizzadenesheli, K., Bhattacharya, K., Stuart, A. M., and Anandkumar, A. (2021). Neural operator: Learning maps between function spaces.CoRR, abs/2108.08481

  42. [42]

    Kynkäänniemi, T., Karras, T., Aittala, M., Aila, T., and Lehtinen, J. (2023). The role of imagenet classes in fréchet inception distance. InProc. ICLR

  43. [43]

    Lasanen, S. (2012). Non-gaussian statistical inverse problems. part i: Posterior distributions. Inverse Problems and Imaging, 6:215–266

  44. [44]

    LeCun, Y ., Cortes, C., and Burges, C. (2010). MNIST handwritten digit database. http: //yann.lecun.com/exdb/mnist. ATT Labs [Online]

  45. [45]

    Lee, S., Lin, Z., and Fanti, G. (2024). Improving the training of rectified flows.Advances in Neural Information Processing Systems, 37:63082–63109

  46. [46]

    Li, S., Kallidromitis, K., Gokul, A., Liao, Z., Kato, Y ., Kozuka, K., and Grover, A. (2024). Om- niflow: Any-to-any generation with multi-modal rectified flows.arXiv preprint arXiv:2412.01169

  47. [47]

    B., Azizzadenesheli, K., Liu, B., Bhattacharya, K., Stuart, A., and Anandkumar, A

    Li, Z., Liu-Schiaffini, M., Kovachki, N. B., Azizzadenesheli, K., Liu, B., Bhattacharya, K., Stuart, A., and Anandkumar, A. (2022). Learning chaotic dynamics in dissipative systems. In Oh, A. H., Agarwal, A., Belgrave, D., and Cho, K., editors,Advances in Neural Information Processing Systems

  48. [48]

    H., Kovachki, N

    Lim, J. H., Kovachki, N. B., Baptista, R., Beckham, C., Azizzadenesheli, K., Kossaifi, J., V oleti, V ., Song, J., Kreis, K., Kautz, J., et al. (2023a). Score-based diffusion models in function space. arXiv preprint arXiv:2302.07400

  49. [49]

    Lim, S., Yoon, E., Byun, T., Kang, T., Kim, S., Lee, K., and Choi, S. (2023b). Score-based generative modeling through stochastic evolution equations in hilbert spaces. InThirty-seventh Conference on Neural Information Processing Systems

  50. [50]

    Lipman, Y ., Chen, R. T. Q., Ben-Hamu, H., Nickel, M., and Le, M. (2023). Flow matching for generative modeling. InThe Eleventh International Conference on Learning Representations

  51. [51]

    L., Mei, X., Ni, Z., Kumar, A., Nagaraja, V ., Wang, W., Plumbley, M

    Liu, H., Lan, G. L., Mei, X., Ni, Z., Kumar, A., Nagaraja, V ., Wang, W., Plumbley, M. D., Shi, Y ., and Chandra, V . (2024a). Syncflow: Toward temporally aligned joint audio-video generation from text

  52. [52]

    Liu, H., Wang, J., Huang, R., Liu, Y ., Lu, H., Xue, W., and Zhao, Z. (2024b). Flashaudio: Rectified flows for fast and high-fidelity text-to-audio generation.arXiv preprint arXiv:2410.12266

  53. [53]

    Liu, P., Dai, D., and Wu, Z. (2024c). Rfwave: Multi-band rectified flow for audio waveform reconstruction.arXiv preprint arXiv:2403.05010

  54. [54]

    Liu, Q. (2022). Rectified flow: A marginal preserving approach to optimal transport.arXiv preprint arXiv:2209.14577

  55. [55]

    and Wang, D

    Liu, Q. and Wang, D. (2016). Stein variational gradient descent: a general purpose bayesian inference algorithm. InProceedings of the 30th International Conference on Neural Information Processing Systems, NIPS’16, page 2378–2386, Red Hook, NY , USA. Curran Associates Inc. 12

  56. [56]

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

  57. [57]

    Liu, Z., Luo, P., Wang, X., and Tang, X. (2015). Deep Learning Face Attributes in the Wild . In2015 IEEE International Conference on Computer Vision (ICCV), pages 3730–3738, Los Alamitos, CA, USA. IEEE Computer Society

  58. [58]

    and Hutter, F

    Loshchilov, I. and Hutter, F. (2019). Decoupled weight decay regularization

  59. [59]

    Ma, Y ., Liu, X., Chen, X., Liu, W., Wu, C., Wu, Z., Pan, Z., Xie, Z., Zhang, H., Zhao, L., et al. (2024). Janusflow: Harmonizing autoregression and rectified flow for unified multimodal understanding and generation.arXiv preprint arXiv:2411.07975

  60. [60]

    Mittal, S., Lajoie, G., Bauer, S., and Mehrjou, A. (2022). From points to functions: Infinite- dimensional representations in diffusion models. InICLR Workshop on Deep Generative Models for Highly Structured Data

  61. [61]

    Miyato, T., Kataoka, T., Koyama, M., and Yoshida, Y . (2018). Spectral normalization for generative adversarial networks. InInternational Conference on Learning Representations

  62. [62]

    Na, K., Lee, J., Yun, S.-Y ., and Lim, S. (2025). Probability-flow ODE in infinite-dimensional function spaces. InICLR 2025 Workshop on Deep Generative Model in Machine Learning: Theory, Principle and Efficacy

  63. [63]

    (2019).Fixed Point Theorems and Applications

    Pata, V . (2019).Fixed Point Theorems and Applications. UNITEXT. Springer Cham, 1 edition

  64. [64]

    Peng, W., Qin, S., Yang, S., Wang, J., Liu, X., and Wang, L. L. (2024). Fourier neural operator for real-time simulation of 3d dynamic urban microclimate.Building and Environment, 248:111063

  65. [65]

    Pidstrigach, J., Marzouk, Y ., Reich, S., and Wang, S. (2024). Infinite-dimensional diffusion models.J. Mach. Learn. Res., 25(1)

  66. [66]

    Preechakul, K., Chatthee, N., Wizadwongsa, S., and Suwajanakorn, S. (2022). Diffusion autoencoders: Toward a meaningful and decodable representation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10619–10629

  67. [67]

    Qin, S., Zhan, D., Geng, D., Peng, W., Tian, G., Shi, Y ., Gao, N., Liu, X., and Wang, L. L. (2025). Modeling multivariable high-resolution 3d urban microclimate using localized fourier neural operator.Building and Environment, 273:112668

  68. [68]

    A., Florez, M

    Rahman, M. A., Florez, M. A., Anandkumar, A., Ross, Z. E., and Azizzadenesheli, K. (2022). Generative adversarial neural operators.Transactions on Machine Learning Research

  69. [69]

    Rout, L., Chen, Y ., Ruiz, N., Caramanis, C., Shakkottai, S., and Chu, W.-S. (2024). Semantic image inversion and editing using rectified stochastic differential equations.arXiv preprint arXiv:2410.10792

  70. [70]

    Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E., Ghasemipour, S. K. S., Gontijo- Lopes, R., Ayan, B. K., Salimans, T., Ho, J., Fleet, D. J., and Norouzi, M. (2022). Photorealistic text-to-image diffusion models with deep language understanding. In Oh, A. H., Agarwal, A., Belgrave, D., and Cho, K., editors,Advances in Neural Information Pr...

  71. [71]

    Sedghi, H., Gupta, V ., and Long, P. M. (2019). The singular values of convolutional layers. In International Conference on Learning Representations

  72. [72]

    Sitzmann, V ., Martel, J., Bergman, A., Lindell, D., and Wetzstein, G. (2020). Implicit neural representations with periodic activation functions.Advances in neural information processing systems, 33:7462–7473

  73. [73]

    and Ermon, S

    Song, Y . and Ermon, S. (2020). Improved techniques for training score-based generative models. Advances in neural information processing systems, 33:12438–12448. 13

  74. [74]

    P., Kumar, A., Ermon, S., and Poole, B

    Song, Y ., Sohl-Dickstein, J., Kingma, D. P., Kumar, A., Ermon, S., and Poole, B. (2021). Score- based generative modeling through stochastic differential equations. InInternational Conference on Learning Representations

  75. [75]

    and Trevisan, D

    Stepanov, E. and Trevisan, D. (2017). Three superposition principles: Currents, continuity equations and curves of measures.Journal of Functional Analysis, 272(3):1044–1103

  76. [76]

    Sun, H., Yu, L., Dai, B., Schuurmans, D., and Dai, H. (2023). Score-based continuous-time discrete diffusion models. InThe Eleventh International Conference on Learning Representations

  77. [77]

    N., Kaiser, L

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L. u., and Polosukhin, I. (2017). Attention is all you need. In Guyon, I., Luxburg, U. V ., Bengio, S., Wallach, H., Fergus, R., Vishwanathan, S., and Garnett, R., editors,Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc

  78. [78]

    Wu, Z., Siarohin, A., Menapace, W., Skorokhodov, I., Fang, Y ., Chordia, V ., Gilitschenski, I., and Tulyakov, S. (2024). Mind the time: Temporally-controlled multi-event video generation. arXiv preprint arXiv:2412.05263

  79. [79]

    Yang, X., Chen, C., Yang, X., Liu, F., and Lin, G. (2024). Text-to-image rectified flow as plug-and-play priors.arXiv preprint arXiv:2406.03293

  80. [80]

    Yin, H., Wang, K., Yang, H., Huang, H., and Silamu, W. (2025). Robust and efficient text- based speech editing using noise conditioning and rectified flow. InICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5. IEEE

Showing first 80 references.

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.