acceptodds
Under review as a conference paper at ICLR 2027

Concentrating Velocities and Accelerations in Hierarchical Flow Matching

Abstract

Hierarchical rectified flow models the multimodal velocity distribution of a flow model by learning an acceleration field, replacing a single ODE with a nested pair. Nothing in this construction forces the distribution modeled at one level to be simpler than the one above. Specifically, at , when source and target samples are paired independently, the velocity distribution there is exactly a shifted copy of the data distribution. We address this by applying mini-batch optimal transport couplings within each level, giving two methods at different operating points. Coupling data pairs concentrates the velocity distribution and straightens the outer trajectories. Coupling velocity pairs concentrates the acceleration distribution and straightens the inner trajectories. Neither alone suffices, because the sampling cost is the product of the steps taken at each level: data coupling alone leaves the inner solve expensive, while velocity coupling alone leaves the outer trajectories curved. Applying both makes every level cheap at once, which is what enables generation with small number of neural function evaluations. We prove that data and velocity couplings hierarchically simplify the velocity and acceleration distributions. We also demonstrate that the couplings improve generation quality and efficiency on synthetic data, MNIST, CIFAR-10, CelebA-HQ 256, and ImageNet 256.

Primary area:
generative models
Keywords:
Generative Model, Flow Matching, Rectified Flow

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_SHabOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper aims to study how can we reduce conditional velocity and acceleration variability in Hierarchical Rectified Flow (HRF2). It proposes two levels of optimal transport coupling: (1) minibatch OT between the source and the target data; (2) conditional OT between the initail and target velocities of the inner flow model. This paper further analyzes the effects of these couplings on conditional variance and evaluate generation quality on extensive experiments, showing promising results.

Critical strengths

S1: Conditional velocity coupling is a clear methodological contribution for HRF2

Compared to previous methods that models conditional velocity distributions with an inner flow, this paper models OT coupling in data space and couples the endpoints of the inner velocity flow, which allows for directly acting on the acceleration matching objective. The methodoloical is grounded in several theoretical foundations. In summary, it provides a concrete framework for studying coupling at multiple levels of a hierarchical generative model.

S2: the experimental results are promising

The medthod shows great achievements in few-step and high-resolution image generation.

Critical weaknesses

W1: The contribution of velocity OT to the main empirical gains is unclear

It is observed that on CIFAR-10 at 5 NFE, second-stage training with teacher-generated velocity pairs achieves 6.632 FID without velocity OT. It shows 6.315 with OT (Table 13) and 22.817 for HRF2-D. Most of the reported improvement over HRF2-D seems to occurs before velocity OT is added. I believe some similar controlled ablation is needed for the headline CelebA-HQ and ImageNet results.

W2: Theorem 3.2 is stronger than its proof supports

This theorem claims marginal preservation for the continuous-time process in Eq. (6), but Appendix D seems to prove it for discrete updates that sample a conditional velocity at each step. Hence It does not establish the claim for a continuous coupled ODE with a fixed inner random seed. It is suggest that the authors should specify the intended process and revise the theorem or its proof accordingly.

Questions for authors

Q1: Please refer to Eq. (6), is the inner random seed reused throughout the continuous outer trajectory, or is a new seed drawn at each outer step? The distinction appears relevant to the claim in Theorem 3.2.

Q2: Does the HRF2-D&V training loss of 0.010 in Table 9 estimate conditional dispersion for the teacher-generated velocity distribution or for the conditional velocity distribution induced by the data coupling? How should this distinction affect the interpretation of the result?

Q3: Equations (8) and (22) express the data and velocity couplings as joint densities. How are these formulas interpreted for exact OT couplings, which may be supported on a transport map and have no joint density?

4 comments

We thank the reviewer for the careful assessment and for recognizing velocity coupling as a clear methodological contribution grounded in theory, the framework for studying coupling at multiple levels of a hierarchical model, and the promising few-step and high-resolution results.

QC1: The contribution of velocity OT to the main empirical gains is unclear. On CIFAR-10 at 5 NFE, second-stage training with teacher-generated velocity pairs reaches 6.632 FID without velocity OT, 6.315 with OT, and 22.817 for HRF2-D, so most of the improvement occurs before velocity OT is added. A similar controlled ablation is needed for the headline CelebA-HQ and ImageNet results.

Velocity coupling means replacing the independent pairing of (v0,v1)(v_0, v_1) at a fixed (xt,t)(x_t, t) with a dependent one; OT is one instance. Both second-stage rows of Table 13 are velocity couplings: the "w/o OT" row is reflow in velocity space, which pairs each v0v_0 with its inner-ODE endpoint v1=Φxt,t(v0)v_1 = \Phi_{x_t,t}(v_0), a deterministic coupling. The large gain over the uncoupled HRF2-D is therefore the effect of coupling itself: under independent pairing, a one-step inner solve returns the conditional mean velocity (Prop. 3.5), and any deterministic coupling removes this obstruction. We relabeled the row "v-space reflow" to make this explicit.

OT is the preferable coupling for three reasons. (1) It improves on v-space reflow at every NFE on CIFAR-10 (Table 13) and matches or improves on it at every NFE on ImageNet-256 (below). (2) It makes the acceleration a deterministic function of (xt,vτ)(x_t, v_\tau) for every τ\tau (Thm. 3.6), whereas v-space reflow pairs can cross for τ>0\tau > 0. (3) It depends only on the velocity distribution, so it can be built from data-coupled paths without simulation (Alg. 4, Table 3), whereas v-space reflow requires integrating the inner ODE.

Following the reviewer's suggestion, we ran the same ablation on ImageNet-256 (new Table 14), adding reflow in data space applied to HRF2-D for completeness (FID):

Total NFE48163264128
HRF2-D9.313.712.642.452.402.40
HRF2-D&Reflow (x-space)3.683.492.812.632.582.56
HRF2-D&V (v-space reflow)3.653.162.832.622.602.60
HRF2-D&V (OT)3.572.942.762.602.582.60

OT matches or outperforms v-space reflow at every NFE, with the largest gap at 8 NFE (2.94 vs. 3.16). We observed the same ordering on MNIST and CelebA-HQ (Appendix G.2.3).

QC2: Theorem 3.2 claims marginal preservation for the continuous-time process in Eq. (6), but Appendix D proves it for discrete updates that sample a conditional velocity at each step. The authors should specify the intended process and revise the theorem or its proof.

The intended process is the outer sampler with a fresh inner draw at every outer step (QC3), which is the sampler used in all our experiments. We restated Thm. 3.2 for exactly this process: on any grid 0=t0<⋯<tN=10=t_0<\dots<t_N=1, with the inner ODE solved exactly from a fresh u0∼π0u_0\sim\pi_0 at each outer step and linear motion in between, Law(Zt)=ρt\mathrm{Law}(Z_t)=\rho_t for every t∈[0,1]t\in[0,1] and every NN.

We also replaced the proof in Appendix D with a shorter one that does not use densities. Lemma D.1 shows that if (Z,V)(Z,V) has the same joint law as (Xs,X1−X0)(X_s, X_1-X_0), then Z+ΔV∼ρs+ΔZ+\Delta V\sim\rho_{s+\Delta} for every Δ∈[0,1−s]\Delta\in[0,1-s], since Xs+Δ=Xs+Δ(X1−X0)X_{s+\Delta}=X_s+\Delta(X_1-X_0); the theorem follows by induction over the outer steps. We do not claim marginal preservation for a coupled ODE that reuses one inner seed.

QC3: In Eq. (6), is the inner random seed reused throughout the outer trajectory, or is a new seed drawn at each outer step?

A new seed u0∼π0u_0\sim\pi_0 is drawn at every outer step, in all our experiments. We now state this in Sec. 2 and in Thm. 3.2. A new remark in Appendix D explains why it matters: if one seed were reused, the position ZtkZ_{t_k} would depend on it for k≥1k\ge 1, its conditional law given ZtkZ_{t_k} would in general differ from π0\pi_0, and the marginal argument would not apply.

QC4: Does the HRF2-D&V training loss of 0.010 in Table 9 estimate conditional dispersion for the teacher-generated velocity distribution or for the one induced by the data coupling? How should this affect the interpretation?

It estimates C2\mathcal C_2 for the velocity distribution π~1\tilde\pi_1 generated by HRF2-D, since HRF2-D&V is trained on pairs with v1∼π~1v_1\sim\tilde\pi_1. The HRF2 and HRF2-D losses use x1−x0x_1-x_0, i.e., the distribution π1\pi_1 induced by the data coupling. This matters in two ways.

(1) Thm. 3.6 holds for any target velocity distribution with a finite second moment, so the drop to 0.010 still tests it: under OT coupling, the acceleration toward π~1\tilde\pi_1 is nearly deterministic, and a single inner step suffices.

(2) The loss does not measure how well π~1\tilde\pi_1 matches π1\pi_1. That gap is a separate error, bounded through εD(M~)\varepsilon_D(\tilde M) and bD(M~)b_D(\tilde M) in Cor. E.13, and it is consistent with the higher FID of HRF2-D&V at large NFE (Remark E.14).

We now state this distinction in Sec. 3.5 and Appendix G.2, and no longer present Table 9 as a pure coupling comparison.

QC5: Eqs. (8) and (22) express the couplings as joint densities. How are these formulas interpreted for exact OT couplings, which may be supported on a transport map and have no joint density?

As in [1, 2], γ\gamma denotes a joint distribution with marginals ρ0\rho_0 and ρ1\rho_1, π1(⋅ ;xt,t)\pi_1(\cdot\,; x_t, t) is the conditional distribution of v1v_1 given xtx_t, and Eq. (8) is its density form. For the exact OT coupling, Eq. (8) holds in the distributional sense: substituting γOT(x0,x1)=ρ0(x0) δ(x1−T(x0))\gamma^{\mathrm{OT}}(x_0,x_1)=\rho_0(x_0)\,\delta(x_1-T(x_0)) and applying the change-of-variables rule for the delta gives π1(v;xt,t)=δ(v−v⋆(xt,t))\pi_1(v;x_t,t)=\delta(v-v^\star(x_t,t)), in agreement with Thm. 3.3. The same calculation applied to Eq. (22) gives the Dirac mass of Thm. 3.6.

We added a remark after Thm. 3.1 stating this interpretation and a remark after Cor. B.2 with the full derivation. The derivation is formal, since the OT map can be discontinuous for multimodal targets. All results involving exact OT (Thm. 3.3, Thm. 3.6, Lemma E.8, Cor. E.9) are therefore proved with conditional laws only, and the new proof of Thm. 3.2 does not use densities either.

References:
[1] Pooladian et al. Multisample flow matching: Straightening flows with minibatch couplings. ICML, 2023.
[2] Tong et al. Improving and generalizing flow-based generative models with minibatch optimal transport. TMLR, 2024.

Thank you for your detailed reply. Your response helps me understand the inner mechanisms of the method especially for the theory part. I will maintain my postive score.

Thank you for supporting the acceptance of our paper. We are glad our response addressed your concerns. Your comments have helped us strengthen the paper. We sincerely appreciate your time and effort.

Reviewer_mkyAOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes mini-batch optimal transport couplings in data and velocity spaces to simplify the velocity and acceleration distributions in Hierarchical Rectified Flow (HRF2). The authors establish marginal preservation and hierarchical simplification properties and demonstrate improvements on synthetic datasets and image generation benchmarks, including CelebA-HQ 256 and ImageNet 256.

Critical strengths
  1. Clear motivation. The observation that each level of HRF2 inherits the full multimodality of the data distribution is simple and well supported. The paper also explains why both couplings are needed: the total NFE is the product of the steps at the two levels.

  2. Solid theoretical analysis. The conditional dispersion equals the irreducible loss of each level, which gives a clear measure of difficulty. The exact decomposition across the two levels and the one-step exactness under velocity OT are clean results. The analysis also explains why HRF2-D&V is better at low NFEs but worse at large NFEs.

Critical weaknesses
  1. Incomplete evaluation of practical generation efficiency.
  • Sampling cost: The paper compares FID against total NFEs but does not report wall-clock sampling time. HRF2 uses nested ODEs with different network components, and on ImageNet 256 the HRF2 models (322M) are larger than RF and OT-CFM (204M) (Table 18). Total NFE alone therefore may not reflect actual sampling efficiency.
  • Training cost: HRF2-D&V requires a pre-trained HRF2-D model, offline velocity-pair generation, and additional fine-tuning. Tables 14–18 only report per-iteration training costs. Without the total training cost, it is hard to judge whether the FID gains at low NFEs are worth the extra training.
  1. Missing comparison with strong few-step generation methods. HRF2-D&V mainly targets the low-NFE regime. However, the paper does not compare against important few-step methods, such as MeanFlow [1] and Consistency Models [2].

[1]. Mean Flows for One-step Generative Modeling, NeurIPS 2025

[2]. Consistency Models, ICML 2023

Questions for authors

See weaknesses

3 comments

We thank the reviewer for the careful assessment and for recognizing the motivation, including why both couplings are needed given the product structure of the total NFE, and the theoretical analysis, including the two-level decomposition, one-step exactness under velocity OT, and the explanation of the low/high-NFE crossover.

QB1: The paper does not report wall-clock sampling time. HRF2 uses nested ODEs with different network components, and on ImageNet-256 the HRF2 models (322M) are larger than RF and OT-CFM (204M), so total NFE alone may not reflect actual sampling efficiency.

We measured wall-clock sampling time on ImageNet-256 (one L40S, batch size 64, classifier-free guidance; new Table 24). The large outer backbone runs once per outer step, and each inner step calls only the small inner head, so one HRF2 step costs about 1.3× an RF/OT-CFM step (3.8 vs. 2.9 ms/img), less than the 1.6× parameter ratio. HRF2-D&V uses one inner step, so its total NFE equals its number of outer steps.

At matched wall-clock time, HRF2-D&V still wins. At 4 NFE it reaches FID 3.57 in 15.2 ms/img, faster and better than OT-CFM at 8 NFE (23.3 ms, FID 4.26). At 8 NFE (30.3 ms) it reaches 2.94, below OT-CFM's best FID of 3.02, which takes about 3× as long (32 NFE, 93.3 ms).

MethodParamsNFETime (ms/img)FID
OT-CFM204M412.78.75
OT-CFM204M823.34.26
OT-CFM204M1646.63.20
OT-CFM204M3293.33.02
HRF2-D&V322M415.23.57
HRF2-D&V322M830.32.94
HRF2-D322M32122.82.45

QB2: HRF2-D&V requires a pretrained HRF2-D model, offline velocity-pair generation, and additional fine-tuning, but only per-iteration training costs are reported. Without the total training cost, it is hard to judge whether the low-NFE gains are worth the extra training.

Velocity coupling is a lightweight fine-tuning stage on top of a pretrained HRF2-D, not a second training run: it freezes the outer backbone, fine-tunes only the 118M inner head, and uses velocity pairs generated once, offline and in parallel. Training costs on ImageNet-256 (NVIDIA L40S GPU-hours; new Table 23):

MethodTrainable paramsBase trainingPair generationFine-tuningTotal GPU-h
RF204M800——800
OT-CFM204M800——800
HRF2322M1,040——1,040
HRF2-D322M1,040——1,040
HRF2-D&V118M—24130154
HRF2-D&Reflow (x-space)322M—24206230

The first four rows are full training runs; the last two are costs added to the shared HRF2-D checkpoint. Both use 5.12M pairs (20,000 space-time locations × 256 velocity pairs for velocity coupling, and the same number of (x0,x1)(x_0, x_1) pairs for data-space reflow). Velocity coupling adds about 15% to the cost of HRF2-D and improves its 4-NFE FID from 9.31 to 3.57, with less compute than data-space reflow (154 vs. 230 GPU-hours).

QB3: HRF2-D&V mainly targets the low-NFE regime, but the paper does not compare against important few-step methods such as MeanFlow and Consistency Models.

Table 8 includes IMM, which generalizes consistency models [1]. Dedicated few-step methods such as MeanFlow [2] and Improved MeanFlow (iMF) [3] attain lower FID at 1–2 NFE than HRF2-D&V at 4 NFE. Our goal is not a new one-step generator, but to remove the cost of the inner solve in hierarchical flow matching.

To examine how MeanFlow-style objectives relate to our framework, we trained HRF2-D&iMF, which replaces the inner acceleration-matching objective of HRF2-D with the iMF objective, under the same budget as HRF2-D&V (new Appendix H). On ImageNet-256 (FID):

Total NFEHRF2-DHRF2-D&VHRF2-D&iMF
49.313.574.66
83.712.943.39
162.642.763.36

HRF2-D&iMF improves substantially over HRF2-D at low NFE, so MeanFlow-style objectives are compatible with the hierarchy, but under the same budget velocity coupling is the more effective choice for the inner level. On the synthetic moon data, we also compare with standalone iMF (SWD, lower is better; Table 16):

NFE12100
iMF0.1750.1380.160
HRF2-D&iMF0.1180.1020.078
HRF2-D&V0.0580.0380.031
In this setting, HRF2-D&V outperforms both at every NFE.

References: [1] Zhou et al. Inductive moment matching. ICML, 2025. [2] Geng et al. Mean flows for one-step generative modeling. NeurIPS, 2025. [3] Geng et al. Improved mean flows: On the challenges of fastforward generative models. CVPR, 2026.

Thanks to the authors' effort and responses. My main concerns are addressed, and I will raise my rating.

Thank you for supporting the acceptance of our paper. We are glad our response addressed your concerns. Your comments have helped us strengthen the paper. We sincerely appreciate your time and effort.

Reviewer_CA8xOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes data and velocity coupling during training hierarchical rectified flow (HRF) model, a recent model for a conditional distribution of possible velocities rather than predicting average velocity as normal FM/RF models. The authors apply minibatch OT coupling at both data and velocity levels, and analyze its effects on synthetic data, to image-scale ones (MNIST, CIFAR-10, CelebA-HQ, ImageNet-256). The data-coupling theory largely adapts established OT flow-matching results to HRF, while the velocity-coupling analysis adds explanation of conditional dispersion and the cost of inner integration. The main empirical result is observed at lower NFE regime with better FID than baselines.

Critical strengths

First, the idea is simple. Applying minibatch OT coupling to data pairs is already well established and here also applied that to velocity pairs as well. Authors provide image-scale empirical evidence that coupling improves HRF, also the two-stage HRF2-D&V achieves FID improvements at lower NFE steps across multiple datasets. Also, authores provide the theoretical analysis to help explain why HRF requires multiple inner integration steps. They show that independent velocity pairing makes a single inner step return the conditional mean, and explains how velocity coupling can remove this limitation.

Critical weaknesses

Unfortunately, there is a major problem here, which is the fact that the original HRF paper already did show incorportating OT coupling to HRF ("HOTCFM2") in their own paper which uses the same data coupling construction as HRF2-D in this paper. They already showed improvements over OT-CFM on toy examples. Authors should've checked this and built further on it. Thus this substantially narrows the novelty of the paper. Likewise, much of the theory regarding DC restates properties already established in the OT-FM literature, e.g., marginal preservation, vanishing velocity dispersion under exact OT, etc.

Now for the paper's own novelty: which is VC, i am not convinced at all with the empirical results in the manuscript. Even on the synthetic examples, the additional improvement is usually minimal; for instance, the HRF2-D and HRF2-D&V curves nearly overlap in Fig2(d). CelebA, ImageNet results in Fig3 also show no clear additional benefit (even worse) from VC at standard NFE budgets.

The main positive empirical result is its few-step performance (ex. "Notably, HRF2-D&V achieves a 78% FID reduction on CelebA-HQ at five neural function evaluations (NFEs) and a 67% FID reduction on ImageNet 256 at 4 NFEs compared with the original HRF2 models." in the Introduction). However, it is difficult to attribute these gains specifically to velocity coupling. Adding VC enables reflow-style retraining, which seems like a major part of the improvement, similar to how RF-2/RF-3 improve few-step sampling by straightening flows. Indeed, Table13 in Appendix shows the improvement from reflow alone is a lot larger than the additional improvement from VC. In sum, DC has a direct related/existing work, while VC fail to show the empirical value.

(minor) presentation issues: some legend markers are missing Fig2 and remaining typos should be corrected.

Questions for authors

See weakness for details.

3 comments

We thank the reviewer for their careful assessment and for recognizing the simplicity of the idea, the image-scale evidence, and our analysis of the inner solve.

QA1: The original HRF paper already showed incorporating OT coupling to HRF ("HOTCFM2"), using the same data-coupling construction as HRF2-D. This substantially narrows the novelty of the paper.

We agree that HRF2-D coincides with HOTCFM2 [1], and we now credit [1] in the introduction, Sec. 3.2, and the related work. [1] reports HOTCFM2 as an empirical variant on synthetic data (its Appendix H.2), without analysis. Our paper adds four aspects.

(1) Theory for coupled HRF. [1] analyzes HRF under independent pairing of source and target samples. We derive the velocity and acceleration distributions for arbitrary couplings, including exact OT couplings without a density (Thm. 3.1, Cor. B.2), and prove marginal preservation for coupled HRF2 (Thm. 3.2). Thm. 3.2 is exact for any number of outer steps, through a mechanism different from marginal preservation in OT flow matching (see QA2).

(2) Where data coupling stops helping. The level-one dispersion C1\mathcal C_1 enters the level-two dispersion C2\mathcal C_2 additively (Thm. 3.4), and under independent velocity pairing C2\mathcal C_2 is strictly positive whenever C1(γ)>0\mathcal C_1(\gamma)>0 (Cor. E.9). Data coupling alone therefore cannot make the inner solve cheap. On ImageNet-256, the converged acceleration loss barely moves under data coupling (1.134 → 1.120, Table 9).

(3) Velocity coupling. This limitation motivates velocity coupling. Exact OT velocity coupling drives C2\mathcal C_2 to zero for every data coupling γ\gamma (Thm. 3.6), and the converged loss drops to 0.010 (measured on the HRF2-D-generated velocity targets).

(4) Scale. HOTCFM2 was evaluated only on low-dimensional data. We evaluate data-coupled HRF on MNIST, CIFAR-10, CelebA-HQ 256, and ImageNet-256 (Tables 4–7).

QA2: Much of the theory regarding DC restates properties already established in the OT-FM literature, e.g., marginal preservation and vanishing velocity dispersion under exact OT.

We agree that Thm. 3.3 adapts known results from OT flow matching [2, 3] to coupled HRF. Thm. 3.2 extends the marginal-preservation result of [1] from independent pairing to arbitrary couplings. However, note that it rests on a different mechanism from marginal preservation in flow matching [2–4]. There, the sampler follows the conditional-mean velocity field, and the marginals are preserved through the continuity equation in continuous time; an Euler discretization with finitely many steps is not marginal-preserving in general. Thm. 3.2 shows that HRF2 with couplings preserves marginal distributions at every t∈[0,1]t \in [0,1] and for any number of outer steps NN, with no discretization error and no regularity assumption on the outer velocity field.

Beyond these two results, our theoretical contribution is the analysis of how the two levels interact, which has no counterpart in single-level OT flow matching: an exact decomposition linking the dispersions of the two levels (Thm. 3.4); a characterization of the one-step inner solve (Prop. 3.5); a strictly positive lower bound on the level-two dispersion under independent velocity pairing (Cor. E.9); optimality of OT among velocity couplings (Lemma E.8) and exactness of a single inner step under it for every data coupling (Thm. 3.6); and error bounds for the simulated velocity sampler that explain the low/high-NFE crossover (Prop. E.12, Cor. E.13).

References:
[1] Zhang et al. Towards hierarchical rectified flow. ICLR, 2025.
[2] Tong et al. Improving and generalizing flow-based generative models with minibatch optimal transport. TMLR, 2024.
[3] Pooladian et al. Multisample flow matching: Straightening flows with minibatch couplings. ICML, 2023.
[4] Liu et al. Flow straight and fast: Learning to generate and transfer data with rectified flow. ICLR, 2023.

QA3: For VC, the additional improvement is usually minimal even on synthetic data (HRF2-D and HRF2-D&V nearly overlap in Fig. 2(d)), and CelebA-HQ and ImageNet show no clear benefit (even worse) at standard NFE budgets.

At low NFEs, the improvement is large, and our analysis explains the high-NFE behavior.

(1) Low NFE. At 4–5 total NFE, VC improves over HRF2-D on every image dataset (FID):

  • MNIST: 10.17 → 5.52
  • CIFAR-10: 22.82 → 6.32
  • CelebA-HQ: 32.92 → 7.61
  • ImageNet-256: 9.31 → 3.57 The gain persists at 8 NFE on ImageNet-256 (3.71 → 2.94) and up to 20 NFE on CIFAR-10 (6.86 → 5.33) and CelebA-HQ (7.05 → 6.56). On the moon data of Fig. 2(d), the gap is concentrated in the first few NFEs, which the plot compresses: SWD 0.574 → 0.058 at 1 NFE and 0.328 → 0.038 at 2 NFE (Table 16).

(2) High NFE. HRF2-D&V is trained on velocity pairs generated once by a fixed HRF2-D sampler. At large outer budgets, its error therefore has a floor bD(M~)+η(M)b_D(\tilde M)+\eta(M) that does not shrink with the inference budget, while HRF2-D keeps improving (Prop. E.12, Cor. E.13, Remark E.14). The floor comes from the simulated targets, not from the coupling: η(M)\eta(M) vanishes as the number of inner steps MM grows (Cor. E.13(b)), and bD(M~)b_D(\tilde M) is set by the accuracy of the base sampler.

(3) Practical use. The two models share the outer backbone, and on ImageNet-256 VC fine-tunes only the 118M inner head (154 GPU-hours, about 15% of the 1,040 for HRF2-D). Both operating points are therefore available at little extra cost.

QA4: It is difficult to attribute the few-step gains to velocity coupling. Adding VC enables reflow-style retraining, which seems to be a major part of the improvement; Table 13 shows the improvement from reflow alone is much larger than the additional improvement from VC.

The "w/o OT (reflow)" row of the original Table 13 is reflow in velocity space: it pairs each v0v_0 with its inner-ODE endpoint v1=Φxt,t(v0)v_1 = \Phi_{x_t,t}(v_0) at a fixed (xt,t)(x_t, t). This is itself a velocity coupling, so Table 13 compares two velocity couplings (v-space reflow vs. OT) against the uncoupled HRF2-D. The large gain over HRF2-D is the effect of velocity coupling: under independent pairing, a one-step inner solve returns the conditional mean velocity (Prop. 3.5), and any deterministic coupling removes this obstruction. OT then adds a further, consistent improvement. We relabeled the row "v-space reflow" and added the same comparison on ImageNet-256 (Table 14).

To test whether the gain is simply reflow-style retraining, we also applied reflow in data space to HRF2-D without velocity coupling (HRF2-D&Reflow (x-space)): we regenerate (x0,x1)(x_0, x_1) pairs with HRF2-D and fine-tune the full model. On ImageNet-256 (FID):

Total NFE48163264128
HRF2-D9.313.712.642.452.402.40
HRF2-D&Reflow (x-space)3.683.492.812.632.582.56
HRF2-D&V (v-space reflow)3.653.162.832.622.602.60
HRF2-D&V (OT)3.572.942.762.602.582.60

HRF2-D&V (OT) matches or outperforms data-space reflow up to 64 NFE, with the largest gap at 8 NFE (2.94 vs. 3.49). On CIFAR-10 the gap is larger (6.32 vs. 10.74 at 5 NFE, Table 13). Velocity coupling is also cheaper: it fine-tunes only the 118M inner head (154 GPU-hours), whereas data-space reflow fine-tunes the full 322M model (230 GPU-hours).

QA5: Some legend markers are missing in Fig. 2, and remaining typos should be corrected.

Thanks, the legend markers in Fig. 2 were too light to see. We enlarged and darkened them, and corrected the remaining typos.

Dear Reviewer CA8x,

Thank you again for your constructive feedback. As you suggested, we have credited HOTCFM2 throughout the paper and clarified our contributions beyond it. Our main theoretical contribution is the analysis of how the two levels interact, which has no counterpart in single-level OT flow matching. It motivates velocity coupling, which makes a single inner step exact and yields large gains at low NFE (e.g., FID 9.31 -> 3.57 at 4 NFE on ImageNet-256). We have also added a data-space reflow comparison on ImageNet-256, where velocity coupling outperforms reflow at lower training cost. Please feel free to reach out if any points need further explanation.

Best,
Authors of Submission 35665

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

We thank all reviewers and ACs for their time. The revised PDF highlights all changes in blue. The main revisions:

  1. Prior work (CA8x). We credit the HOTCFM2 variant of [1] in the introduction, Sec. 3.2, and the related work, and revised our contributions accordingly. Our additions are the velocity and acceleration distributions under general couplings (Thm. 3.1, Cor. B.2), marginal preservation for coupled HRF2 (Thm. 3.2), the two-level analysis (Thm. 3.4, Prop. 3.5, Cor. E.9), velocity coupling, and image-scale evidence.

  2. Attribution of the velocity-coupling gain (CA8x, SHab). We relabeled the “w/o OT” row of Table 13 as vv-space reflow, which is itself a velocity coupling. We added the same ablation on ImageNet-256 (Table 14), together with reflow in data space. Velocity coupling outperforms data-space reflow at low NFE at lower cost, and OT matches or improves on v-space reflow at every NFE.

  3. Practical cost (mkyA). We added wall-clock sampling time (Table 24) and total training cost (Table 23). At 4 NFE, HRF2-D&V reaches FID 3.57 in 15.2 ms/img, faster and better than OT-CFM at 8 NFE (4.26, 23.3 ms). Velocity coupling adds 154 GPU-hours, about 15% of HRF2-D’s training cost.

  4. Few-step methods (mkyA). We added comparisons with Improved MeanFlow, both standalone and as the inner objective of HRF2-D (Appendix H).

  5. Theory (SHab). We restated Thm. 3.2 for the sampler used in practice, with a fresh inner draw at every outer step, and gave a density-free proof. We added remarks on interpreting Eqs. (8) and (22) for exact OT couplings, and clarified what Table 9 measures.

References:
[1] Zhang et al. Towards hierarchical rectified flow. ICLR, 2025.

Area_Chair_1hzSOpenReview ↗

Engage in discussion

Reviewers: please review the authors’ rebuttal and consider revising your assessment/score.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.