Hierarchical Rectified Flow Matching with Mini-Batch Couplings

Alex Schwing; Yichi Zhang; Yici Yan; Zhizhen Zhao

REVIEW 3 major objections 6 minor 1 cited by

Hierarchical Rectified Flow Matching with Mini-Batch Couplings

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read By coupling source and target samples within mini-batches, hierarchical rectified flow can simplify the multi-modal velocity distribution it must model and, with an additional velocity coupling, generate useful samples in as few as one…

desk verdict Solid incremental extension of HRF; data coupling is a clear win, velocity coupling is regime-dependent and rests on an idealized theorem. read the letter →

arxiv 2507.13350 v1 pith:R2UB2FJW submitted 2025-07-17 cs.CV cs.LG

Yichi Zhang , Yici Yan , Alex Schwing , Zhizhen Zhao This is my paper

classification cs.CVcs.LG

keywords flowmatchinghierarchicalrectifiedmini-batchoptimaltransportvelocitycouplingdatagenerativemodelingfew-stepgenerationmarginalpreservation

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Hierarchical rectified flow models the multi-modal distribution of velocities that appear when interpolating between source and data samples, but until now every level of the hierarchy faced the same complexity. This paper claims that choosing how source and target samples are paired inside mini-batches controls that complexity: optimal-transport couplings in data space collapse the multi-modal velocity distribution toward a single mode, with batch size as the dial. A second coupling, applied to velocity samples generated by a pretrained model, does not simplify the velocity distribution but straightens its integration paths, so fewer solver steps are needed. The combined two-stage method, data coupling followed by velocity coupling, is reported to improve generation quality consistently on synthetic and image benchmarks and to keep quality high even at one neural function evaluation.

What carries the argument

The load-bearing object is the coupling-induced velocity distribution, expressed by Theorem 3.1: at a fixed interpolation point, $\pi_1(v; x_t, t) = \gamma(x_t - tv, x_t + (1-t)v) / \rho_t(x_t)$. This identity converts the question "how hard is the velocity distribution?" into a question about the joint coupling $\gamma$. Mini-batch optimal transport supplies the coupling in data space, and a second mini-batch OT on simulated velocity samples supplies the coupling in velocity space; the two-stage training procedure alternates by learning acceleration on data-coupled pairs first, then generating velocity pairs from that model and learning on velocity-coupled pairs. Theorem 3.2, proved by a characteristic-function calculation, shows that the data-coupled sampling ODE preserves the marginals of the linear interpolation.

What would settle it

On a fixed synthetic dataset, measure the empirical velocity distribution at t=0 for a range of OT batch sizes; if increasing the batch size does not drive the distribution toward a single mode, the paper's central simplification effect is refuted. Alternatively, corrupt the base model used for velocity coupling (or use a deliberately undertrained one) and check whether HRF2-D&V still beats HRF2-D at low NFE; the paper's stated dependence on base-model quality predicts it will not.

Watch

Extended reading notes

Core claim

The central discovery is a distribution-level identity and its consequences. For any joint coupling $\gamma$ of source and target samples, the velocity distribution at a space-time location is $\pi_1(v; x_t, t) = \gamma(x_t - tv, x_t + (1-t)v) / \rho_t(x_t)$; when source and target are independent this reduces to a shifted copy of the data distribution, so the velocity layer has to model exactly the complexity of the data. The paper shows that mini-batch optimal transport replaces independence with a coupling that concentrates $\gamma$ and therefore simplifies $\pi_1$: larger batches make the velocity distribution increasingly unimodal. Coupling velocities, obtained by simulating the pretrained depth-two model, does not change that distribution but cuts the number of integration steps required to sample it. The authors further prove that sampling with data-coupled mini-batches preserves the correct marginals, and demonstrate on MNIST, CIFAR-10, and CelebA-HQ that data coupling consistently improves FID over hierarchical rectified flow and OT-coupled flow matching, with additional gains at low NFE from joint velocity coupling.

Load-bearing premise

The load-bearing premise is that the pretrained model used to produce velocity samples for coupling is accurate enough that its outputs faithfully represent the true velocity distribution, since the paper's velocity-coupling gains depend on the quality of that base model.

Editorial extensions

If this is right

Data coupling alone makes HRF2 consistently beat both vanilla HRF2 and OT-coupled flow matching on the reported image benchmarks, at both low and high total NFE.
Joint data and velocity coupling yields usable generation at total NFE = 1, an extreme where data coupling alone collapses.
Batch size in the mini-batch OT acts as a control knob: larger batches push the velocity distribution toward unimodal and simplify every lower level of the hierarchy.
Velocity coupling reduces the integration steps needed to model the velocity ODE without altering the velocity distribution, so it complements rather than replaces data coupling.
The marginal-preservation theorem means the data-coupled sampling process produces the correct data distribution in the limit of an exactly learned acceleration field.

Reading between the lines

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

Nothing in the coupling construction is specific to depth two, so the same data/velocity coupling recipe should transfer to deeper hierarchies; the paper only tests HRF2.
Because velocity coupling inherits the pretrained model's errors, a better base model should yield better velocity pairs; one testable extension is iterating the two-stage procedure reflow-style and checking whether gains compound.
The paper uses batch size to trade multimodality against fidelity, but a natural and untested consequence is that very large batches could reduce sample diversity, so diversity metrics would be needed to find the sweet spot.
Theorem 3.2 assumes an exactly learned acceleration field; in practice finite-step and model-error bias remain, so the reported FID comparisons carry an approximation caveat that the paper does not quantify.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Short version: this is a clean, incremental extension of hierarchical rectified flow. The data-coupling variant is a consistent win across all their datasets and is worth building on. The velocity-coupling variant is more conditional: it helps only at low NFE, depends on a pretrained base model, and the marginal-preservation theorem doesn't cover the actual sampling procedure with finite steps and a learned inner ODE. I'd still send it out.

What's new: Theorem 3.1 gives the velocity density for arbitrary joint data couplings, a natural but useful generalization. Theorem 3.2 shows the marginal-preservation property for data coupling, which is the right thing to prove. The experiments are solid: HRF2-D beats HRF2 and OT-CFM on MNIST, CIFAR-10, and CelebA-HQ at essentially every NFE, and the tables include std over seeds. The paper also ships code, and the authors are upfront that velocity coupling needs simulated targets and depends on base-model quality.

Soft spots. The stress-test is right that Theorem 3.2 proves an idealized statement: it assumes velocity samples are drawn exactly from pi1(v; x_t,t), while in practice they are produced by numerically integrating a learned acceleration. This is the same idealization all flow matching theory makes, so I don't count it as a flaw unique to this paper. The more load-bearing issue is velocity coupling itself: Algorithm 3 uses a pretrained HRF2-D checkpoint to fabricate v1 targets, then re-couples them with OT. Any bias in that base model gets baked in and reinforced by two-stage training. The authors acknowledge this, but the paper never quantifies how sensitive the low-NFE gains are to the quality of that checkpoint. The tables show the joint coupling is worse than data coupling alone at high NFE, which hints the synthesized velocity targets are not neutral. Also, synthetic plots have no error bars, and Figure 3's low-NFE claims would be easier to trust with a few seeds.

Net: the data-coupling result is believable and reproducible; the velocity-coupling result is a provisional, regime-specific bonus. This is the kind of incremental contribution that is useful to the flow-matching community and warrants a normal review rather than a desk rejection. Main requests: error bars on synthetic curves, a sensitivity analysis over base-model checkpoints, and a statement that the theoretical guarantee is asymptotic/exact while the empirical procedure uses the learned inner ODE.

Referee Report

3 major / 6 minor

Summary. This paper studies hierarchical rectified flow matching (HRF2) and proposes to control the complexity of the multi-modal velocity distribution by coupling source and target samples within mini-batches via optimal transport. Three variants are presented: data coupling only (HRF2-D), velocity coupling only, and joint data-and-velocity coupling (HRF2-D&V). The authors prove a change-of-variables formula for the conditional velocity distribution under a general joint distribution (Theorem 3.1) and a marginal-preservation result for the hierarchical ODE under exact conditional velocity draws (Theorem 3.2). Empirically, HRF2-D consistently improves FID over HRF2 and OT-CFM on MNIST, CIFAR-10, and CelebA-HQ, and HRF2-D&V gives further gains at low numbers of function evaluations. The manuscript includes code and detailed experimental settings.

Significance. If the empirical results are robust, the paper offers a simple training-time modification with clear practical benefits, especially in the low-NFE regime, and the theoretical change-of-variables formula is a useful formalization. The experiments are broad (synthetic and three image datasets), and the computational-cost tables are helpful. However, the main theoretical result does not cover the velocity-coupling algorithm proposed in the paper, because Theorem 3.2 assumes exact draws from the conditional velocity distribution while Algorithm 3 uses a pretrained approximate model. Since velocity coupling is the component responsible for the headline low-NFE gains, this gap is load-bearing; the paper's current framing overstates the guarantee behind HRF2-D&V.

major comments (3)

[Appendix B; Section 3.3 (Algorithm 3) and Tables 1-3] Theorem 3.2 is proven for an idealized process in which, at each outer location (z_t,t), the velocity is drawn exactly from π1(·;z_t,t). In the proposed velocity-coupling procedure, v1 is obtained by numerically integrating Eq. (4) with a pretrained acceleration network aθ (Algorithm 3), and Appendix D.1 explicitly states that 'performance depends on the quality of the base model.' Hence the theorem does not cover the actual HRF2-D&V procedure, and bias in aθ or finite-step Euler error can be written into the pseudo-targets and reinforced by two-stage training. This concern is mirrored by the empirical pattern: HRF2-D&V is better than HRF2-D at low NFE but clearly worse at high NFE (e.g., MNIST at NFE 100: FID 1.847 vs 3.423; CIFAR-10 at NFE 100: 4.301 vs 5.078; CelebA-HQ at NFE 100: 4.961 vs 6.248). Please add an explicit approximation-error statement or, preferably, a synthetic experiment in which velocity coupling uses ground-truth velocities to test whether the low-NFE gains survive when the base model is unbiased.
[Section 3.5] The paper asserts that 'the consistency of the velocity distribution with mini-batch velocity coupling directly follows prior works that use mini-batch coupling and reflow for data generation,' but no formal statement or proof is provided for the velocity-coupling case. Since HRF2-D&V is a central contribution, the manuscript should either state explicitly that no marginal-preservation guarantee is claimed for velocity coupling, or give a theorem with the exactness assumptions made explicit.
[Section 3.2 and Theorem 3.2] Theorem 3.2 assumes a fixed joint distribution γ with marginals ρ0 and ρ1, but Algorithm 1 re-samples the mini-batch and recomputes the OT coupling at every iteration, so the training objective is an expectation over random couplings. The text should clarify whether the marginal-preservation claim refers to the induced mixture coupling (the expectation of the random mini-batch coupling), and if so state that explicitly; otherwise the proof does not directly support the data-coupling algorithm as implemented.

minor comments (6)

[Appendix B] In the proof of Theorem 3.2, the sentence 'The equality a is valid due to Theorem 3.2' should refer to Theorem 3.1.
[Eq. (16)] In the characteristic-function computation, 'γ(xt − vt, xt + (1 − t)v)' should be typeset as 'γ(xt − t v, xt + (1 − t)v)' to avoid confusing the scalar t with the variable v_t.
[Section 4.2] The statement 'OT-CFM is essentially equivalent to HRF1-D' uses the symbol HRF1-D, which is never defined; either define it or rephrase the comparison.
[Figure 4 caption] The caption mentions 'HRF2-D&V-OT' without defining what the '-OT' suffix denotes; please clarify the naming or remove it.
[Algorithm 3] Algorithm 3 says 'Use OT to couple source and target points' but does not specify the cost function or the solver used; the main text should state that the 2-Wasserstein cost and the POT solver are used, and report the number of inner integration steps used to generate v1.
[Appendix D] The evaluation protocol is described differently in D.1 (three seeds, 27 runs) and D.2 (five evaluation runs); please clarify whether the protocols differ by dataset or are described inconsistently.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the coupling analyses are change-of-variable identities, Theorem 3.2 is a standard characteristic-function argument, and the velocity-coupling targets are an acknowledged reflow-style training dependency rather than a renamed prediction.

full rationale

The paper's theoretical results do not reduce to their inputs. Theorem 3.1 is a direct change-of-variables computation: given a joint coupling gamma and the linear interpolation, it derives the conditional velocity density pi_1(v; x_t,t) = gamma(x_t - t v, x_t + (1-t)v) / rho_t(x_t). This is a mathematical consequence of the definition, not an assumed conclusion. Theorem 3.2 is proved by a characteristic-function argument: it assumes the inner velocity ODE transports pi_0 to pi_1 (the standard marginal-preserving property of rectified flow at the velocity level) and then shows the outer Euler step Z_{t+Delta t} = Z_t + V Delta t has the same marginal as the interpolation process. The proof does not use its own conclusion; the phrase in Appendix B, 'The equality a is valid due to Theorem 3.2,' is evidently a typo for Theorem 3.1, since the substituted identity is exactly Theorem 3.1's density formula. Velocity coupling (Algorithm 3) does generate targets v_1 by numerically integrating a pretrained HRF2-D model, and the paper explicitly acknowledges that 'performance depends on the quality of the base model.' This is a reflow/stage-wise training dependency, not a circular prediction: the final HRF2-D&V model is evaluated on held-out Wasserstein/FID benchmarks against independent baselines (RF, OT-CFM, HRF2), so the reported gains are not forced by the training objective. Self-citations to Zhang et al. [2025] supply the HRF framework and baselines, but the new mini-batch coupling results are independently validated and do not rely on a self-citation chain to forbid alternatives or to define the results into existence. No fitted parameter is renamed as a prediction, and no equation equals its own input by construction.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central theory uses no fitted parameters. The free parameters are training hyperparameters that control the coupling strength. The main axioms are standard measure-theoretic tools (change of variables, characteristic functions), the marginal-preserving property of the inner flow from prior work, and the engineering assumption that the pretrained model used for velocity coupling approximates the true velocity distribution well enough.

free parameters (3)

mini-batch OT batch size (data coupling) = 100 (synthetic), 128 (MNIST/CIFAR-10), 256 (CelebA-HQ)
Controls the degree of distribution simplification; chosen by hand without sensitivity analysis.
number of integration steps for velocity sample generation = Not specified exactly; described as 'more integration steps required' for image data
Used in Algorithm 3 to generate target velocity samples from the pretrained model; affects quality of velocity coupling.
training batch size = 1000 (synthetic), 128 (MNIST/CIFAR-10), 256 (CelebA-HQ)
Separate from OT batch size; used for gradient computation and affects stability.

assumptions (4)

standard math Standard change-of-variables formula and characteristic function arguments used in Theorems 3.1 and 3.2.
The density transformation in Theorem 3.1 relies on a linear change of variables with unit Jacobian; the marginal-preservation proof uses characteristic functions to show equality of distributions.
domain assumption The inner acceleration model, when perfectly learned, generates samples from the velocity distribution π1(v; xt, t) at each location and time.
The proof of Theorem 3.2 assumes Zt evolves with velocity sampled from π1; this is the marginal-preserving property of rectified flow applied to the inner ODE, cited from prior work (Liu et al. 2023, Zhang et al. 2025).
domain assumption Mini-batch optimal transport provides a valid joint distribution γ with the correct marginals, up to the approximation error discussed in Fatras et al.
The training objectives (Eq. 11 and 13) treat the mini-batch OT coupling as a genuine joint distribution with marginals ρ0 and ρ1, which holds exactly within each batch.
ad hoc to paper The pretrained HRF2-D model used in Algorithm 3 approximates the true velocity distribution accurately enough for velocity coupling to be beneficial.
Velocity coupling depends on simulated target velocities from a pretrained checkpoint; the paper notes 'performance depends on the quality of the base model' and selects checkpoints via validation.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Hierarchical Rectified Flow Matching with Mini-Batch Couplings." pith.science (2026). https://pith.science/paper/R2UB2FJW

@misc{pith2026250713350,
  author       = {Pith},
  title        = {Pith review of: Hierarchical Rectified Flow Matching with Mini-Batch Couplings},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/R2UB2FJW}},
  note         = {Machine review of arXiv:2507.13350}
}

read the original abstract

Flow matching has emerged as a compelling generative modeling approach that is widely used across domains. To generate data via a flow matching model, an ordinary differential equation (ODE) is numerically solved via forward integration of the modeled velocity field. To better capture the multi-modality that is inherent in typical velocity fields, hierarchical flow matching was recently introduced. It uses a hierarchy of ODEs that are numerically integrated when generating data. This hierarchy of ODEs captures the multi-modal velocity distribution just like vanilla flow matching is capable of modeling a multi-modal data distribution. While this hierarchy enables to model multi-modal velocity distributions, the complexity of the modeled distribution remains identical across levels of the hierarchy. In this paper, we study how to gradually adjust the complexity of the distributions across different levels of the hierarchy via mini-batch couplings. We show the benefits of mini-batch couplings in hierarchical rectified flow matching via compelling results on synthetic and imaging data. Code is available at https://riccizz.github.io/HRF_coupling.

Figures

Figures reproduced from arXiv: 2507.13350 by the authors.

**Figure 1.** The generated velocity distributions at (xt, t) = (−1, 0) for the dataset 1D N → 2N using HRF2, (a) without couplings, (b)-(c) with data coupling (batch sizes: 5 and 100), (d) with velocity coupling (batch size: 100), (e)-(f) with velocity coupling (batch size: 100) and data coupling (batch sizes: 5 and 100). Data coupling simplifies the velocity distribution (cf. (a)-(c)), while velocity coupling reduces the number… view at source ↗

**Figure 2.** Results on 8N → moon dataset. Three rows are HRF2, HRF2 with data coupling, HRF2 with data & velocity coupling. (a) and (c) are trajectories (green) of sample particles flowing from source distribution (grey) to target distribution (blue) with total NFEs 1 and 100. (b) and (d) are velocity distributions at the center of the bottom left Gaussian mode at t = 0. Data coupling simplifies the velocity distribution and ve… view at source ↗

**Figure 3.** Results on synthetic datasets: (a) 1D N → 2N (b) 1D N → 5N (c) 2D N → 6N (d) 2D 8N → moon. Top row: HRF2-D&V generated data distributions. Bottom row: performance vs. total NFEs. We use Wasserstein and sliced 2-Wasserstein distances for 1D and 2D data, respectively. We compare our method to RF [Liu, 2022], HRF2 [Zhang et al., 2025] and OT-CFM [Tong et al., 2024] baselines. HRF2 is the base pre-trained model for our … view at source ↗

Figures from the paper (5 more)

**Figure 4.** Figure 4: Results on MNIST, CIFAR-10 and CelebA-HQ 256 datasets. (a) HRF2-D with total NFE=1. [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]

**Figure 5.** Figure 5: Results on 1D N → 2N data. The three rows correspond to HRF2, HRF2 with data coupling, HRF2 with data & velocity coupling. (a) and (b) are generated data distribution with total NFEs 1 and 100. (c) is velocity distribution at (xt, t) = (−1, 0). Total NFEs RF (1.08M) OT…

**Figure 6.** Figure 6: Results on 1D N → 5N data. The three rows correspond to HRF2, HRF2 with data coupling, HRF2 with data & velocity coupling. (a) and (b) are generated data distribution with total NFEs 1 and 100. (c) is velocity distribution at (xt, t) = (0, 0). Total NFEs RF (35.75M) OT…

**Figure 7.** Figure 7: Results on 2D N → 6N data. The three rows correspond to HRF2, HRF2 with data coupling, HRF2 with data & velocity coupling. (a) and (c) are trajectories (green) of sample particles flowing from source distribution (grey) to target distribution (blue) with total NFEs 1 a…

**Figure 8.** Figure 8: Examples of the generated images for CelebA-HQ starting from the same noise for all [PITH_FULL_IMAGE:figures/full_fig_p016_8.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

Deep Neural Networks Inspired by Differential Equations
cs.LG 2025-10 unverdicted

A review of differential-equation-inspired neural networks that compiles known results into a taxonomy, with no new experiments or theory.

Reference graph

Works this paper leans on

34 extracted references · 24 canonical work pages · cited by 1 Pith paper

[1]

Albergo and E

M. Albergo and E. Vanden-Eijnden. Building normalizing flows with stochastic interpolants. In Proc. ICLR, 2023

work page 2023
[2]

Albergo, N

M. Albergo, N. Boffi, and E. Vanden-Eijnden. Stochastic interpolants: A unifying framework for flows and diffusions. arXiv preprint arXiv:2303.08797, 2023

arXiv 2023
[3]

R. T. Chen and Y. Lipman. Flow matching on general geometries. In ICLR, 2024

work page 2024
[4]

Y. Chen, M. Goldstein, M. Hua, M. S. Albergo, N. M. Boffi, and E. Vanden-Eijnden. Probabilistic forecasting with stochastic interpolants and f " ollmer processes. In ICML, 2024

work page 2024
[5]

H. K. Cheng and A. Schwing. The Curse of Conditions: Analyzing and Improving Optimal Transport for Conditional Flow-Based Generation . In Proc. ICCV, 2025

work page 2025
[6]

Q. Dao, H. Phung, B. Nguyen, and A. Tran. Flow matching in latent space, 2023. URL https://arxiv.org/abs/2307.08698

arXiv 2023
[7]

Deshpande, Z

I. Deshpande, Z. Zhang, and A. G. Schwing. Generative Modeling using the Sliced Wasserstein Distance . In Proc. CVPR, 2018

work page 2018
[8]

Deshpande, Y.-T

I. Deshpande, Y.-T. Hu, R. Sun, A. Pyrros, N. Siddiqui, S. Koyejo, Z. Zhao, D. Forsyth, and A. G. Schwing. Max-Sliced Wasserstein Distance and its use for GANs . In Proc. CVPR, 2019

work page 2019

Show all 34 references

[9]

Esser, S

P. Esser, S. Kulal, A. Blattmann, R. Entezari, J. M \"u ller, H. Saini, Y. Levi, D. Lorenz, A. Sauer, F. Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first International Conference on Machine Learning, 2024

2024
[10]

Fatras, Y

K. Fatras, Y. Zine, R. Flamary, R. Gribonval, and N. Courty. Learning with minibatch wasserstein: asymptotic and gradient properties. In AISTATS 2020-23nd International Conference on Artificial Intelligence and Statistics, volume 108, pages 1--20, 2020

2020
[11]

Fatras, Y

K. Fatras, Y. Zine, S. Majewski, R. Flamary, R. Gribonval, and N. Courty. Minibatch optimal transport distances; analysis and applications. arXiv preprint arXiv:2101.01792, 2021

2021 arXiv
[12]

Flamary, N

R. Flamary, N. Courty, A. Gramfort, M. Z. Alaya, A. Boisbunon, S. Chambon, L. Chapel, A. Corenflos, K. Fatras, N. Fournier, et al. POT : Python optimal transport. Journal of Machine Learning Research, 22 0 (78): 0 1--8, 2021

2021
[13]

I. Gat, T. Remez, N. Shaul, F. Kreuk, R. T. Chen, G. Synnaeve, Y. Adi, and Y. Lipman. Discrete flow matching. In NeurIPS, 2024

2024
[14]

Guo and A

P. Guo and A. G. Schwing. Variational rectified flow matching. arXiv preprint arXiv:2502.09616, 2025

2025 arXiv
[15]

B. Jing, B. Berger, and T. Jaakkola. Alphafold meets flow matching for generating protein ensembles. arXiv preprint arXiv:2402.04845, 2024

2024 arXiv
[16]

Karras, T

T. Karras, T. Aila, S. Laine, and J. Lehtinen. Progressive growing of GAN s for improved quality, stability, and variation. In International Conference on Learning Representations, 2018. URL https://openreview.net/forum?id=Hk99zCeAb

2018
[17]

Krizhevsky

A. Krizhevsky. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009

2009
[18]

LeCun, L

Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 1998

1998
[19]

Lipman, R

Y. Lipman, R. Chen, H. Ben-Hamu, M. Nickel, and M. Le. Flow Matching for Generative Modeling . In Proc. ICLR, 2023

2023
[20]

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

2022 arXiv
[21]

X. Liu, C. Gong, and Q. Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. In Proc. ICLR, 2023 a

2023
[22]

X. Liu, X. Zhang, J. Ma, and J. Peng. Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. In ICLR, 2023 b

2023
[23]

D. Park, S. Lee, S. Kim, T. Lee, Y. Hong, and H. J. Kim. Constant acceleration flow. In NeurIPS, 2024

2024
[24]

Peebles and S

W. Peebles and S. Xie. Scalable diffusion models with transformers, 2023. URL https://arxiv.org/abs/2212.09748

2023 arXiv
[25]

Peyr \'e , M

G. Peyr \'e , M. Cuturi, et al. Computational optimal transport: With applications to data science. Foundations and Trends in Machine Learning , 11 0 (5-6): 0 355--607, 2019

2019
[26]

Pooladian, H

A.-A. Pooladian, H. Ben-Hamu, C. Domingo-Enrich, B. Amos, Y. Lipman, and R. Chen. Multisample flow matching: Straightening flows with minibatch couplings. In Proc. ICML, 2023

2023
[27]

Rombach, A

R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer. High-resolution image synthesis with latent diffusion models. In Proc. CVPR, 2022

2022
[28]

Stark, B

H. Stark, B. Jing, C. Wang, G. Corso, B. Berger, R. Barzilay, and T. Jaakkola. Dirichlet flow matching with applications to dna sequence design. In ICML, 2024

2024
[29]

A. Tong, N. Malkin, G. Huguet, Y. Zhang, J. Rector-Brooks, K. Fatras, G. Wolf, and Y. Bengio. Improving and generalizing flow-based generative models with minibatch optimal transport. TMLR, 2024

2024
[30]

C. Villani. Optimal transport: old and new, volume 338. Springer, 2009

2009
[31]

J. Yim, A. Campbell, A. Y. Foong, M. Gastegger, J. Jim \'e nez-Luna, S. Lewis, V. G. Satorras, B. S. Veeling, R. Barzilay, T. Jaakkola, and F. Noe. Fast protein backbone generation with SE (3) flow matching. In NeurIPS Workshop: Machine Learning in Structural Biology, 2023

2023
[32]

Zhang and M

F. Zhang and M. Gienger. Affordance-based robot manipulation with flow matching. arXiv preprint arXiv:2409.01083, 2024

2024
[33]

Zhang, Y

X. Zhang, Y. Pu, Y. Kawamura, A. Loza, Y. Bengio, D. L. Shung, and A. Tong. Trajectory flow matching with applications to clinical time series modeling. In NeurIPS, 2024

2024
[34]

Zhang, Y

Y. Zhang, Y. Yan, A. Schwing, and Z. Zhao. Towards hierarchical rectified flow. In The Thirteenth International Conference on Learning Representations, 2025

2025

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Albergo and E

M. Albergo and E. Vanden-Eijnden. Building normalizing flows with stochastic interpolants. In Proc. ICLR, 2023

work page 2023

[2] [2]

Albergo, N

M. Albergo, N. Boffi, and E. Vanden-Eijnden. Stochastic interpolants: A unifying framework for flows and diffusions. arXiv preprint arXiv:2303.08797, 2023

arXiv 2023

[3] [3]

R. T. Chen and Y. Lipman. Flow matching on general geometries. In ICLR, 2024

work page 2024

[4] [4]

Y. Chen, M. Goldstein, M. Hua, M. S. Albergo, N. M. Boffi, and E. Vanden-Eijnden. Probabilistic forecasting with stochastic interpolants and f " ollmer processes. In ICML, 2024

work page 2024

[5] [5]

H. K. Cheng and A. Schwing. The Curse of Conditions: Analyzing and Improving Optimal Transport for Conditional Flow-Based Generation . In Proc. ICCV, 2025

work page 2025

[6] [6]

Q. Dao, H. Phung, B. Nguyen, and A. Tran. Flow matching in latent space, 2023. URL https://arxiv.org/abs/2307.08698

arXiv 2023

[7] [7]

Deshpande, Z

I. Deshpande, Z. Zhang, and A. G. Schwing. Generative Modeling using the Sliced Wasserstein Distance . In Proc. CVPR, 2018

work page 2018

[8] [8]

Deshpande, Y.-T

I. Deshpande, Y.-T. Hu, R. Sun, A. Pyrros, N. Siddiqui, S. Koyejo, Z. Zhao, D. Forsyth, and A. G. Schwing. Max-Sliced Wasserstein Distance and its use for GANs . In Proc. CVPR, 2019

work page 2019