acceptodds
Under review as a conference paper at ICLR 2027

Beyond Straightness: Non-Crossing Flow Matching via Quantile AlignTree Coupling

Abstract

The performance of Flow Matching largely depends on the quality of the coupling between the source and target distributions. However, independent coupling often leads to path crossings and local velocity ambiguity, while OT-based couplings typically incur high construction costs. To address this challenge, we propose Quantile AlignTree Flow Matching (QAT-FM), an efficient structured coupling strategy that constructs a hierarchical coupling between a Gaussian prior and the target data distribution via a quantile-aligned tree structure. QAT-FM constructs the coupling in time and supports per-pair source sampling with complexity, enabling scalable training for large-scale high-dimensional generative tasks. Theoretically, we prove that the QAT coupling satisfies marginal consistency, induces non-crossing linear interpolation paths, and consistently improves path separation at intermediate times compared with independent coupling, thereby alleviating local velocity ambiguity. QAT-FM further extends naturally to conditional generation, enabling structured conditional coupling while preserving global Gaussian alignment. Experiments across diverse benchmark datasets demonstrate that QAT-FM achieves competitive generative performance while substantially reducing coupling construction cost.

Primary area:
generative models
Keywords:
Flow Matching, Generative Models, Data Alignment, Binary Trees

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_A12MOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes Quantile AlignTree (QAT), a coupling for Flow Matching that splits the data with a KD-tree and the Gaussian at matching quantiles, so each training image is paired with its own region of the noise; the tree is built once offline and also supports conditional generation. The paper proves that the resulting paths do not cross and are more separated than under the independent coupling, and argues that this reduces velocity ambiguity. Experiments cover unconditional CIFAR-10, class-conditional ImageNet-32 and ImageNet-256 (latent), and text-conditional CelebA-64.

Critical strengths
  1. A cheap coupling over the whole dataset. The tree is built once, offline, in O(Nd log N) time, and the noise for a given image is drawn in O(d) time, so drawing a training pair costs about as much as with I-FM. Unlike mini-batch OT, the pairing is defined over the whole dataset rather than within a batch, and the construction scales to more than 2M samples on ImageNet and to VAE latents on ImageNet-256.
  2. Exact Gaussian prior, also per class. The quantile-aligned splits keep the noise distribution exactly N(0, I) by construction. For class-conditional generation, each class gets its own tree, so the noise paired with each class is also exactly N(0, I), which avoids the train-test gap of condition-agnostic OT (Cheng & Schwing, 2025) without solving any OT problem.
  3. Broad evaluation with gains over I-FM. The method is tested on unconditional, class-conditional, text-conditional, and latent-space generation, changing only the coupling. On all four benchmarks, QAT-FM improves over I-FM both with the fewest steps and with the most accurate solver (e.g., Euler-2 on CIFAR-10: 98.68 vs. 164.18; ImageNet-256 with 250 steps: 1.99 vs. 2.08).
Critical weaknesses
  1. The theory does not support the paper's central claim. The paper claims that non-crossing paths (Theorem 1) and larger path separation (Theorem 2) alleviate local velocity ambiguity and make the velocity field easier to fit. Both theorems are correct, but they do not establish this:

    • Theorem 2 is a statement about transport cost. For a given coupling, let S(t)=E∥xt−yt∥2S(t) = \mathbb{E}\lVert x_t - y_t \rVert^2 be the path separation for two independent pairs (x0,x1)(x_0, x_1) and (y0,y1)(y_0, y_1) from this coupling, and C=E∥x0−x1∥2C = \mathbb{E}\lVert x_0 - x_1 \rVert^2 the transport cost. Since xt−yt=(1−t)(x0−y0)+t(x1−y1)x_t - y_t = (1-t)(x_0 - y_0) + t(x_1 - y_1), S(t)=(1−t)2E∥x0−y0∥2+t2E∥x1−y1∥2+2t(1−t)E[(x0−y0)⊤(x1−y1)].S(t) = (1-t)^2\mathbb{E}\lVert x_0 - y_0 \rVert^2 + t^2\mathbb{E}\lVert x_1 - y_1 \rVert^2 + 2t(1-t)\mathbb{E}[(x_0 - y_0)^\top(x_1 - y_1)]. The first two terms do not depend on the coupling: x0,y0x_0, y_0 are two independent noise samples, and x1,y1x_1, y_1 are two independent data samples. In the last term, E[y0⊤y1]=E[x0⊤x1]\mathbb{E}[y_0^\top y_1] = \mathbb{E}[x_0^\top x_1], and the products across pairs vanish because the pairs are independent and the noise has zero mean (e.g. E[x0⊤y1]=E[x0]⊤E[y1]=0\mathbb{E}[x_0^\top y_1] = \mathbb{E}[x_0]^\top \mathbb{E}[y_1] = 0), so this term equals 4t(1−t)E[x0⊤x1]4t(1-t)\mathbb{E}[x_0^\top x_1]. Hence both quantities depend on the coupling only through E[x0⊤x1]\mathbb{E}[x_0^\top x_1]: S(t)=a(t)+4t(1−t)E[x0⊤x1],C=E∥x0∥2+E∥x1∥2−2E[x0⊤x1],S(t) = a(t) + 4t(1-t)\mathbb{E}[x_0^\top x_1], \qquad C = \mathbb{E}\lVert x_0 \rVert^2 + \mathbb{E}\lVert x_1 \rVert^2 - 2\mathbb{E}[x_0^\top x_1], where a(t)a(t) denotes the first two terms. QAT and the independent coupling have the same marginals, so a(t)a(t) and the norm terms are the same for both, and E[x0⊤x1]=0\mathbb{E}[x_0^\top x_1] = 0 under the independent coupling. Hence SQAT(t)−Sind(t)=4t(1−t)EQAT[x0⊤x1],Cind−CQAT=2EQAT[x0⊤x1].S_{\mathrm{QAT}}(t) - S_{\mathrm{ind}}(t) = 4t(1-t)\mathbb E_{\mathrm{QAT}}[x_0^\top x_1], \qquad C_{\mathrm{ind}} - C_{\mathrm{QAT}} = 2\mathbb E_{\mathrm{QAT}}[x_0^\top x_1]. So ΔQAT\Delta_{\mathrm{QAT}} in Theorem 2 equals EQAT[x0⊤x1]\mathbb E_{\mathrm{QAT}}[x_0^\top x_1], and the gain is 2t(1−t)(Cind−CQAT)2t(1-t)(C_{\mathrm{ind}} - C_{\mathrm{QAT}}), i.e. the theorem only says that QAT has a lower transport cost than the independent coupling. Any coupling with a lower cost has larger separation, and full OT maximizes it.

      This also contradicts Table 1, which marks mini-batch OT and RecFlow as not improving separation, although both lower the transport cost (for RecFlow, see Liu et al., 2023).

    • Path separation does not measure ambiguity. The paper says that ambiguity "arises locally", where paths cross or come close with different velocities, whereas path separation averages over all pairs of paths, most of which are far apart. So a coupling can remove all ambiguity and still have lower separation than the independent one. For example, let the data also be a 2D standard Gaussian and set x1=Rx0x_1 = R x_0, with RR a rotation by 120∘120^\circ. Then xt=Mtx0x_t = M_t x_0 with Mt=(1−t)I+tRM_t = (1-t)I + tR and det⁡Mt=1−3t(1−t)>0\det M_t = 1 - 3t(1-t) > 0, so xtx_t determines x0x_0: paths never cross and the velocity is unambiguous. Yet E[x0⊤x1]=tr⁡R=2cos⁡120∘=−1\mathbb{E}[x_0^\top x_1] = \operatorname{tr} R = 2\cos 120^\circ = -1, while it is 00 for the independent coupling, so by the identity above the separation is lower.

    • Non-crossing removes ambiguity only by making the target discontinuous. With one sample per leaf, the coupling is deterministic, so the ideal flow is exactly straight and one Euler step would suffice. But its velocity is piecewise, with one piece per training image and jumps at every split, so a network can only approximate it. The paper does not analyze whether this target is easier to learn than the smooth, averaged velocity of I-FM. In practice, the trained model is far from the ideal flow: the Euler-2 FID on CIFAR-10 is 98.68 (Table 2).

  2. The experiments only partially show the geometric benefit that motivates the method. Non-crossing and better-separated paths should make the ODE easier to integrate, i.e. require fewer function evaluations (NFE) and give better samples with few steps. The evidence for this is mixed:

    • NFE is not lower. With Dopri5, QAT-FM uses the most NFE on CIFAR-10 (145.6 vs. 137.3 for I-FM, Table 2), slightly more on ImageNet-32 (143.4 vs. 140.3, Table 4), and only slightly fewer on CelebA (144.1 vs. 150.0, Fig. 4).
    • The few-step gain is inconsistent. QAT-FM improves over I-FM at Euler-2 (e.g., 105.16 vs. 122.40 on ImageNet-32) and at 10 steps on ImageNet-256 (10.67 vs. 11.83, Table 5). However, on ImageNet-32 the gain shrinks to at most 0.16 FID at Euler-5 to Euler-20 and reverses at Euler-10 (Euler-5/10/20: 22.88/10.30/6.40 vs. 23.04/10.12/6.48); a gain of about 0.25 FID appears again only at Euler-50+ and Dopri5.
    • On CIFAR-10 and ImageNet-256, the gains with many steps are small. On CIFAR-10, QAT-FM is better than I-FM with Dopri5 (3.26 vs. 3.38) but worse at Euler-100 (3.91 vs. 3.79). On ImageNet-256, the gain is about 0.1 FID at 50 and 250 steps (Table 5, no std reported).

    Overall, the results show a modest quality improvement at low coupling cost, but do not show that it comes from better path geometry; the paper reports neither transport cost nor path straightness.

  3. The conditional extension is justified for classes, but not for continuous conditions such as text.

    • Classes work. With w = infinity, each class gets its own tree that covers the whole Gaussian, so the noise paired with each class is N(0, I), the same as at inference.
    • Continuous conditions break this. In CQAT, a split on a condition coordinate leaves the noise region unchanged, while a split on an image coordinate also splits the noise region in two. With finite w, as used for text, image splits occur before the captions are fully separated, and nodes with fewer than 128 samples are split only on image coordinates; the paper chooses this rule to increase the number of noise splits and thus the gain of Theorem 2 (Section 3.4). As a result, during training the noise paired with a given caption covers only a small part of the Gaussian, while at inference it is drawn from the full N(0, I). Cheng & Schwing (2025) identify the same train-test gap as the cause of the poor results of condition-agnostic OT, also visible here on ImageNet-32 (FID 6.65 for OT-FM vs. 3.82 for I-FM, Table 4). The paper cites this work but does not discuss the gap for CQAT.
    • The theoretical claims do not cover this case. Table 1 states that Gaussian alignment and path separation are "also preserved in conditional generation". For finite w, Gaussian alignment holds only for the noise averaged over all conditions, not for a fixed condition. Proposition 2 averages path separation over independent pairs of samples, which mostly have different conditions; what matters is ambiguity for a fixed condition, since the network receives the condition as input.
    • For text, the ambiguity that CQAT targets is already absent. In real text-image data, almost every caption belongs to a single image. Then all paths with a given caption go from different noise samples to the same image and meet only at that image, so the target velocity is unambiguous even under the independent coupling. A gain from CQAT would therefore have to come from how the network shares information across similar captions, which the paper does not analyze.
    • The CelebA experiment does not test this setting. CelebA-Dialog captions are generated from templates over five facial attributes with six levels each (Jiang et al., 2021), i.e. at most 15552 attribute combinations for about 200K images, so the task is close to class-conditional generation. In addition, caption alignment is measured only by a CLIP score reported to two decimals (0.10 for both QAT-FM and I-FM), which is too coarse to detect small changes.
Questions for authors
  1. Mini-batch OT with multiple GPUs. Appendix C lists the batch size per GPU (128 for ImageNet-32, 64 for CelebA) and uses 4 GPUs, i.e. an effective batch of 512 and 256. Is the OT-FM coupling computed on each GPU's local batch or on the effective batch? For C2OT, the paper states that OT is solved per GPU, but for OT-FM this is not specified. This matters because a smaller OT batch gives a coupling closer to the independent one.

  2. Coupling cost of C2OT. Appendix C states that the oversampling in C2OT "substantially increases the coupling cost" and refers to Appendix D.1, but the runtime comparison there (Fig. 6) does not include C2OT. Since low coupling cost is the main practical advantage of QAT over C2OT, could the authors report the total training time of I-FM, C2OT, and QAT-FM in their setup, including the tree construction for QAT?

  3. Memorization. Each training image is paired with a fixed region of the noise, so copying may be easier to learn than with I-FM, and FID against training statistics does not penalize it. Could the authors compare nearest-neighbor distances to the training set for QAT-FM and I-FM, and report FID on held-out data (e.g., the CIFAR-10 test set)?

7 comments
AuthorsOpenReview ↗

Response to Reviewer A12M (Part 1 / 3)

Thank you for your thoughtful feedback. Below, we provide responses to your comments in the following section and we have modified the manuscript to address these issues. (Changes are highlighted in blue in the updated PDF.)

  • On W1.1: Theorem 2 and Transport Cost

    First, we would like to clarify that Theorem 2 is not intended to show that QAT outperforms OT-based methods. More importantly, it provides an explicit node-wise expression for the path-separation gain, showing that further non-degenerate splits along data coordinates continuously increase this gain. This provides theoretical guidance for constructing deeper QAT trees in practice.

    We acknowledge that exact Semi-Discrete OT between the full data distribution and the standard Gaussian provides an upper bound on path separation gain. We thank the reviewer for pointing this out and have added this observation to the revised manuscript (P16, Lines 825-831). However, solving the full OT problem is computationally challenging. As shown in Figure 6, Full OT incurs substantial coupling construction and sampling costs, limiting its practical applicability.

    Regarding Table 1, we thank the reviewer for pointing out this issue and have corrected it in the revised manuscript.

  • On W1.2: Path Separation and Velocity Ambiguity

    We thank the reviewer for providing this counterexample. We would like to clarify that we do not regard Path Separation as a strict measure of Velocity Ambiguity, but rather as a way to improve path geometry.

    Specifically, larger Path Separation means that intermediate states XtX_t are more spatially dispersed, potentially reducing the likelihood that paths with different target velocities overlap or come too close in local regions, thereby alleviating Velocity Ambiguity.

    We acknowledge that there are special cases where Velocity Ambiguity is absent despite smaller Path Separation, as illustrated by the reviewer's counterexample. However, this does not invalidate our motivation to alleviate local Velocity Ambiguity by increasing Path Separation.

  • On W1.3: Learnability of the Velocity Field

    In general, it is difficult to directly determine whether the velocity field induced by a given coupling is easier for a neural network to learn. Nevertheless, QAT-FM achieves a substantially better Euler-2 FID than I-FM (98.68 vs. 164.18), suggesting that its piecewise structure does not prevent effective learning in practice.

  • On W2.1: NFE and FID Results

    Dopri5 uses adaptive step sizes, so NFE depends on the learned velocity field and solver tolerances. Although QAT-FM requires slightly more NFE, it achieves the best FID. Moreover, under the same Euler-100 sampling budget, QAT-FM outperforms OT-FM and SD-FM on CIFAR-10, and all baselines on ImageNet-32. These results support the effectiveness of our method.

  • On W2.2: Euler-10 Performance on ImageNet-32 On ImageNet-32, although we observe a slight fluctuation at Euler-10, QAT-FM outperforms I-FM under all other settings (Euler-2/5/20/50/100 and Dopri5), providing strong empirical evidence of its effectiveness.

  • On W2.3: The FID Gains

    On CIFAR-10, many coupling-based methods yield limited improvements. For instance, OT-FM achieves an FID of 3.49, while QAT-FM improves it to 3.26, demonstrating the effectiveness of our method.

    For ImageNet-256, we repeated the 250-step evaluation across three random seeds, obtaining FIDs of 2.088 ± 0.009 for I-FM and 2.005 ± 0.023 for QAT-FM. These results demonstrate that our improvement is consistent across different seeds.

  • On W3.1 & W3.2: Train–Inference Consistency for Continuous Conditions (Part 1/2)

    We thank the reviewer for highlighting this issue. First, train–inference mismatch is not unique to CQAT but is a general concern in conditional OT-based methods (e.g., C2OT [1]). We further provide a rigorous theoretical analysis showing that, for augmented OT with a finite condition weight, conditional Gaussian consistency holds if and only if the source and target conditions match almost surely (cs=ctc_s=c_t, π⋆\pi^\star-a.s.). This highlights a general challenge in continuous conditional coupling.

    An advantage of CQAT is that it can explicitly characterize and sample from the conditional source distribution. Specifically, CQAT represents this distribution as γc=(proj⁡x0)♯πCQAT(⋅∣C=c)\gamma _ c=(\operatorname{proj} _ {x_0}) _ {\sharp}\pi_{\mathrm{CQAT}}(\cdot\mid C=c). Since γc\gamma_c is a mixture of truncated Gaussians, we can directly sample from it at inference for conditions observed during training, ensuring consistency with the training source distribution.

AuthorsOpenReview ↗

Response to Reviewer A12M (Part 2 / 3)

  • On W3.1 & W3.2: Train–Inference Consistency for Continuous Conditions (Part 2/2)

    To further validate this strategy, we compare the FID and CLIP scores obtained using the standard Gaussian prior γ\gamma and the conditional prior γc\gamma_c.

    • CelebA Caption-Conditional Generation (Euler-100):

      PriorFID (Train)FID (Test)CLIP (Train)CLIP (Test)
      γ\gamma3.174.280.1040.104
      γc\gamma_c3.144.260.1050.105
    • ImageNet-256 Caption-Conditional Generation (MeanFlow, NFE = 1):

      PriorFID (Train)FID (Test)CLIP (Train)CLIP (Test)
      γ\gamma24.1322.400.3130.313
      γc\gamma_c23.6521.850.3140.314

    These results show that the conditional prior further improves generation quality and text–image alignment on both training and held-out data. Additional results are provided in Figure 4 (P9, Lines 466-474) and Section D.3 (P25, Lines 1329-1403).

  • On W3.3: Theoretical Claims

    • Gaussian Alignment. We would like to clarify that CQAT constructs a valid coupling between the continuous Gaussian γ\gamma and the joint data-condition distribution μ^DC\widehat{\mu} _ {D_C}, i.e., πCQAT∈Π(γ,μ^DC)\pi_{\mathrm{CQAT}}\in\Pi(\gamma,\widehat{\mu}_{D_C}). Thus, our claim of Gaussian Alignment refers to global marginal consistency, rather than conditional Gaussian consistency for each fixed condition.

    • Proposition 2. We would like to emphasize that Proposition 2 extends Theorem 2 to conditional generation, aiming to quantify the global Path Separation Gain of CQAT. More importantly, it provides an explicit node-wise expression showing that further non-degenerate splits along data coordinates increase the global Path Separation Gain, thereby providing theoretical guidance for constructing deeper trees in conditional generation.

  • On W3.4: The Condition–Image Pairs

    We would like to clarify that one-to-one text-image pairs in the training set do not imply a one-to-one mapping in conditional generation. In fact, Flow Matching learns to transport a noise distribution to the conditional image distribution, rather than learning a deterministic mapping from text to a single image. A well-trained conditional generative model should generate diverse images consistent with the same caption using different initial noise samples. Therefore, one-to-one text–Image pairs in a finite training set neither imply a deterministic underlying conditional distribution nor invalidate the motivation for optimizing conditional couplings.

    We also provide CelebA generation examples in Figure 12 (P30). For each fixed caption, our model generates diverse images consistent with the condition from different initial noise samples, even though each caption corresponds to only one image in the training set.

  • On W3.5: The CelebA Settings

    We would like to clarify that, in the actual caption-generation code of CelebA-Dialog, the same attribute combination can correspond to multiple different captions. Specifically, among approximately 200K samples, we identify 181,956 unique captions, substantially exceeding 15,552. Moreover, we use continuous text embeddings as conditioning inputs, making it appropriate to treat this task as continuous-condition generation.

    Regarding the CLIP scores, we have updated the reported values to three decimal places in the revised manuscript (P9, Lines 466-474).

  • On Q1: Mini-batch OT with Multiple GPUs

    We thank the reviewer for pointing this out. We clarify that OT-FM computes the coupling independently on each GPU's local batch, with batch sizes of 128 for ImageNet-32 and 64 for CelebA-64, rather than using the aggregated batch across four GPUs. We have clarified this implementation detail in Appendix C.

  • On Q2: C2OT Runtime Comparison

    We additionally report the training throughput and end-to-end runtime of I-FM, OT-FM, C2OT, and QAT-FM, including offline tree construction (with PCA) for QAT-FM.

    • ImageNet-32:
    MethodThroughput (img/s)Offline Build (h)E2E (h)
    I-FM1733.3—24.745
    OT-FM1727.0Online24.802
    C2OT1696.3Online25.251
    QAT-FM1730.70.06824.850
    • CelebA-64:
    MethodThroughput (img/s)Offline Build (h)E2E (h)
    I-FM624.0—11.525
    OT-FM620.3Online11.592
    C2OT590.2Online12.189
    QAT-FM621.00.01411.593

    These results show that QAT-FM achieves higher training throughput and lower end-to-end runtime than C2OT, with minimal offline construction overhead.

AuthorsOpenReview ↗

Response to Reviewer A12M (Part 3 / 3)

  • On Q3: Memorization

    We additionally evaluate the 1-NN MSE between generated samples and the training set, as well as FID against the held-out CIFAR-10 test set. The results are reported below.

    • 1-NN MSE:

      MethodEuler-2Euler-100Dopri5
      Test Set0.02920.02920.0292
      I-FM0.00340.02970.0314
      OT-FM0.00930.03020.0314
      SD-FM0.00350.02850.0300
      QAT-FM0.01070.02940.0306
    • CIFAR-10 Test-Set FID:

      MethodEuler-2Euler-100Dopri5
      I-FM167.97±0.667.83±0.077.37±0.07
      OT-FM81.71±0.447.99±0.077.49±0.05
      SD-FM166.37±0.778.17±0.117.69±0.08
      QAT-FM103.06±0.537.94±0.087.28±0.05

    QAT-FM achieves the highest 1-NN MSE among generative methods at Euler-2. Notably, at Euler-100 and Dopri5, its 1-NN MSE is even higher than that of real test samples. Moreover, it maintains competitive test-set FID, achieving the best result under Dopri5. These results suggest no clear evidence of excessive training-set memorization. Full experimental details are provided in Appendix D.3 (P25, Lines 1329-1347).

In summary, we hope the above response resolves your concerns and helps you re-evaluate our work. Feel free to contact us if you have any other questions.

Reference

[1] The Curse of Conditions: Analyzing and Improving Optimal Transport for Conditional Flow-Based Generation. ICCV 2025.

Reviewer_agWhOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Reducing the crossing of training paths alleviates the local ambiguity of the regression velocity target. Prior works try to find an global optimal coupling of all noise-data pairs to reduce such crossings: full OT coupling is great but computationally prohibitive; SD-FM aligns a continuous Gaussian to the full dataset but needs iterative optimization. This paper proposes QAT coupling: build an axis-aligned binary-space-partition tree over the data, then transfer the topology to Gaussian by matching the mass fraction at each split, and eventually pair each data point with its truncated-Gaussian box. The offline construction is O(Ndlog⁡N)\mathcal O(Nd\log N), sampling is O(d)\mathcal O(d) per pair, and source marginal is exactly Gaussian. The theoretical results yield non-crossing and path-seperation gain over independent coupling. The authors also include a practical variant that can leverage the condition information. Empirically, it shows competitive generation quality over synthetic 2D, CIFAR-10, class-conditioned ImageNet-32, text-conditioned CelebA-64, and latent-space ImageNet-256 datasets with engineering tricks for high-dimensional data.

Critical strengths
  • The coupling is built once, offline, in a single greedy pass. Training needs only a bound lookup and truncated-normal draws and inference remains the same because the source distribution tracks exactly to isotropic Gaussian.
  • The high-dimension experiments keep the source exactly Gaussian via orthogonal maps only, which is a interesting idea to restrict the coupling to a subspace.
  • Empirical experiments have sufficient coverage, including synthetic vs real, class-conditioned vs text-conditioned, pixel-space vs latent-space, CNN backbone vs DiT backbone, and many-step vs few-step, showing the generality of the method.
Critical weaknesses
  • The construction resembles Nurbekyan et al. (2019) who build a binary-space-partitioning tree over source and target simultaneously splitting both along the same coordinate with equal mass on each side, pair the resulting leaf cells, prove the straight-line paths never cross, and has the complexity of O(Nlog⁡N)\mathcal O(N\log N) with no optimization but only a median search. QAT coupling is this construction with their 50th-percentile split replaced by a data-dependent quantile (e.g., fraction of a node's data below its mean) and a continuous Gaussian source. In addition, Kong et al. (2025) also scales semi-discrete OT coupling to high-dimensional data, but remains undiscussed.
  • It's odd to claim superiority against independent coupling as main theoretical results because every non-trivial coupling beats independent coupling. An more informative comparison shall be against semi-discrete OT which shares QAT's marginals.
  • It seems there is a train-inference mismatch for the conditional case. At inference time, every condition is paired with the full isotropic Gaussian, so training must pair every condition to full Gaussian too, but Definition 3 only guarantees that the noise is Gaussian when pooled over all conditions.

References

  • Nurbekyan et al. (2019) No-collision Transportation Maps
  • Kong et al. (2025) AlignFlow: Improving Flow-based Generative Models with Semi-Discrete Optimal Transport
Questions for authors

Please prioritize the critical weaknesses above, while there is a minor point:

  • Could you report the offline build time, memory requirement, hardware configuration and training throughput for the latent-space ImageNet-256 and CelabA-64 experiments (i.e., top-2 highest ambient dimension), with numbers comparing with their respective baselines?
5 comments
AuthorsOpenReview ↗

Response to Reviewer agWh (Part 1 / 2)

We thank the reviewer for the positive assessment of our method and experimental results. Detailed responses to your comments are provided below, and the manuscript has been revised accordingly. (Changes are highlighted in blue in the updated PDF.)

  • On W1: Discussion of Related Work We thank the reviewer for pointing out the insufficient discussion of related work. We have added both studies to the revised manuscript and clarify their differences from our work below.

    • Nurbekyan et al. [1]. This work primarily studies Non-Crossing Transportation Maps by constructing BSP trees over both source and target distributions. We acknowledge its connection to QAT through BSP-based mass alignment, but the objectives and implementations differ. QAT constructs trees solely from the target data and analytically aligns them with the continuous Gaussian prior through Gaussian quantiles. After construction, only the Gaussian bounds associated with each data point are stored, enabling repeated sampling in O(d)O(d) time without additional optimization or tree traversal. Moreover, Theorem 2 quantifies the path-separation gain over Independent Coupling, and we extend the framework to conditional FM and large-scale generation. Thus, our main contribution is not the Non-Crossing property itself, but an efficient Gaussian coupling construction tailored to FM, together with its theoretical analysis.

    • AlignFlow [2]. We have added a discussion of AlignFlow in the Introduction (P2, Line 66). Given the time and computational constraints, we conduct experiments using MeanFlow (SiT-B/4) under two settings: class-conditional (discrete) and caption-conditional (continuous) generation. The results are presented below.

      • Class-conditional Generation:
      MethodFIDCoupling Time (h)Sampling Time (h)Total Time (h)
      AlignFlow13.540.910.2825.43
      QAT-MF13.670.080.0424.35
      • Caption-conditional Generation:
      MethodFIDCoupling Time (h)Sampling Time (h)Total Time (h)
      AlignFlow28.8215.251.4140.80
      QAT-MF24.130.010.0524.20

      For continuous conditions, SDOT cannot readily exploit class-wise partitioning to reduce the coupling problem size, leading to higher convergence and sampling costs. Full experimental settings and results are provided in Appendix D.2 (P24, Lines 1275-1328).

  • On W2: Role of Theoretical Results

    Theorem 2 is not merely intended to show that QAT improves path separation over Independent Coupling. More importantly, it provides an explicit node-wise expression for the path-separation gain, showing that further non-degenerate splits along data coordinates continuously increase this gain. This provides theoretical guidance for constructing deeper QAT trees in practice.

    Moreover, we note that exact quadratic-cost Semi-Discrete OT provides a theoretical upper bound on the path-separation gain. However, computing this bound is computationally expensive for large-scale datasets. In contrast, QAT allows the path-separation gain to be explicitly computed from the tree structure without additional optimization, while maintaining efficient coupling construction and sampling.

AuthorsOpenReview ↗

Response to Reviewer agWh (Part 2 / 2)

  • On W3: Conditional Train–Inference Consistency

    We thank the reviewer for highlighting this important issue. In fact, train–inference mismatch is not unique to CQAT but is a general concern in conditional OT-based coupling, as also discussed in prior work such as C2OT [3].

    We further provide a rigorous theoretical analysis (Detailed in P17, Lines 809-938) showing that, for augmented OT, conditional Gaussian consistency holds if and only if the source and target conditions match almost surely (cs=ctc_s=c_t, π⋆\pi^\star-a.s.). For discrete conditions, this can be ensured by constructing separate couplings for each class. However, for continuous conditions, a finite condition weight generally cannot guarantee exact matching, potentially leading to a train–inference mismatch.

    An advantage of CQAT is that it can explicitly characterize and sample from the conditional source distribution. Specifically, CQAT represents the conditional source distribution as γc=(proj⁡x0)♯πCQAT(⋅∣C=c)\gamma _ c=(\operatorname{proj} _ {x_0}) _ {\sharp}\pi_{\mathrm{CQAT}}(\cdot\mid C=c). Since γc\gamma_c is a mixture of truncated Gaussians, we can directly sample from it at inference for conditions observed during training, ensuring consistency with the training source distribution.

    To further validate this analysis, we compare the training-set and and held-out set FIDs and CLIP Scores obtained using the standard Gaussian prior γ\gamma and the CQAT-induced conditional prior γc\gamma_c.

    • CelebA Caption-Conditional Generation (Euler-100):
    PriorFID (Train)FID (Test)CLIP (Train)CLIP (Test)
    γ\gamma3.174.280.1040.104
    γc\gamma_c3.144.260.1050.105
    • ImageNet-256 Caption-Conditional Generation (MeanFlow, NFE = 1):
    PriorFID (Train)FID (Val)CLIP (Train)CLIP (Val)
    γ\gamma24.1322.400.3130.313
    γc\gamma_c23.6521.850.3140.314

    These results show that the conditional prior improves both generation quality and text–image alignment. Additional experimental results are provided in Figure 4 (P9, Lines 466-474) and Section D.3 (P25, Lines 1329-1403).

  • On Q1: Computational Cost

    We report the offline construction time (including PCA), training throughput, GPU memory usage, and end-to-end runtime as follows.

    • CelebA-64 (4 × NVIDIA RTX 4090 GPUs, 24 GB each):
    MethodThroughput (img/s)VRAM (GB)Offline build (h)E2E (h)
    I-FM624.021.057—11.525
    OT-FM620.321.069Online11.592
    C2OT590.220.201Online12.189
    QAT-FM621.021.1560.01411.593
    • ImageNet-256 (8 × NVIDIA RTX PRO 6000 GPUs, 96 GB each):
    MethodThroughput (img/s)VRAM (GB)Offline build (h)E2E (h)
    I-FM1998.37056.922—11.387
    QAT-FM1943.91557.3360.04211.754

In summary, we hope the above response successfully addresses your concerns and strengthens your confidence in supporting our work during the discussion and decision phases. Thank you for your consideration.

Reference

[1] No-collision transportation maps.Journal of Scientific Computing 2020.

[2] AlignFlow: Improving Flow-based Generative Models with Semi-Discrete Optimal Transport. ICLR 2026.

[3] The Curse of Conditions: Analyzing and Improving Optimal Transport for Conditional Flow-Based Generation. ICCV 2025.

AuthorsOpenReview ↗

Gentle Reminder

Dear Reviewer agWh,

As the discussion period is drawing to a close, please let us know whether our responses and additional experiments have addressed your concerns and strengthened your support for our work.

Thank you again for your time and constructive feedback.

Reviewer_xag8OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a novel method to couple noise and data points in training flow matching models. The method construct a quantile tree for data. Based one the data tree, the quantile tree for noise space is constructed. The data tree and the noise tree are aligned. Using the constructed trees, the authors proposed the method: Quantile AlignTree Flow Matching (QAT-FM) to train a flow matching model. The authors conducted experiments on the CIFAR-10, ImageNet, CelebA datasets.

Critical strengths

The idea of constructing the quantile tree for data space and quantile tree for the noise space and align the two trees is novel. The method for constructing the quantile tree for the data space is simple: selecting a dimension, choose a threshold and partition the space. For noise space tree construction, the same dimension corresponding to the data quantile tree is chosen, and the threshold is chosen so that the CDF is the same as the data tree for that dimension. This has a good property that the paired noise and data points from different tree leaves are not crossing (Theorem 1), an important property for training Flow Matching models.

The authors also proposed the Conditional QAT, which considers the conditional variables in training FM models. The tree construction strategy is similar to non-conditional one, owing to the simplicity of the tree construction method.

Critical weaknesses

I'm concerned about the results in Table 2. The authors reported the FIDs at 100k training steps. The OT-FM using Dopri5 is around 3.49. However, after checking the original paper: "Improving and Generalizing Flow-Based Generative Models with Minibatch Optimal Transport" (Figure 3), I found that for OT-FM should be over 5.5 at 100k training steps. The authors need to clarify this.

I also have similar concern in Table 5. In the ImageNet256 dataset, the vanilla Flow Matching achieves the FIDs as low as 11.83 with 10 steps, 2.57 with 50 steps and 2.08 with 250 steps. These numbers are too low to be true. The authors need to check their code to make sure they got the correct FID numbers.

Since this paper is claimed to align noise and data for training flow matching models, the authors need to compare with strong baselines: Mousavi-Hosseini et al., 2026 and Kong et al., 2026 (AlignFlow: Improving Flow-based Generative Models with Semi-Discrete Optimal Transport) on the ImageNet256 dataset.

Questions for authors

How to decide when to stop constructing the trees? What kind of granularity is enough? The authors need to have experiments for this.

5 comments
AuthorsOpenReview ↗

Response to Reviewer xag8

We thank the reviewer for the careful review and for recognizing the novelty and feasibility of our method. We have responded to your specific points below and updated the manuscript to reflect these changes. (Changes are highlighted in blue in the updated PDF.)

  • On W1: CIFAR-10 Settings

    We would like to clarify that the FID reported in our Table 2 was obtained under a different experimental configuration from that in Figure 3 of OT-FM [1]:

    • OT-FM [1]: channels: 128, batch size: 128, training steps: 400k.
    • Our setting: channels: 192, batch size: 256, training steps: 100k.

    Our experiments adopt the batch size and training schedule of C2OT [2], with a larger number of channels. Thus, we achieve a better FID at 100k training steps than that reported in the original OT-FM paper [1]. All methods are evaluated under the same experimental settings. These settings are detailed in the Appendix C (P22, Lines 1134–1150).

  • On W2: FID on ImageNet-256

    For ImageNet-256, we strictly follow the experimental settings of LightningDiT [3], which are also adopted by C2OT. As reported in Table 3 of LightningDiT, an FID of 2.11 is achieved with 250 sampling steps after 64 training epochs, consistent with the I-FM result reported in our paper.

  • On W3: Comparison with Semi-Discrete OT Methods

    Since both SD-FM [4] and AlignFlow [5] are based on Semi-Discrete OT, and SD-FM has not released its conditional generation code, we primarily compare with AlignFlow.

    Given the time and computational constraints, we conduct experiments using MeanFlow (SiT-B/4) under two settings: class-conditional (discrete) and caption-conditional (continuous) generation. The class-conditional experiments reproduce the official AlignFlow implementation, while the caption-conditional setting follows C2OT and SD-FM. Full experimental details are provided in Appendix D.2 (P24, Lines 1275-1328). We report the FID scores (NFE = 1) and computational costs of different methods below.

    • Class-conditional Generation:
    MethodFIDCoupling Time (h)Sampling Time (h)Total Time (h)
    I-MF14.72--24.23
    AlignFlow13.540.910.2825.43
    QAT-MF13.670.080.0424.35
    • Caption-conditional Generation:
    MethodFIDCoupling Time (h)Sampling Time (h)Total Time (h)
    I-MF24.82--24.14
    AlignFlow28.8215.251.4140.80
    QAT-MF24.130.010.0524.20

    The results demonstrate that:

    • Class-conditional generation: QAT-MF achieves comparable FID to AlignFlow with lower computational costs. Since ImageNet-256 contains 1,000 classes, AlignFlow can solve SDOT separately for each class, substantially reducing the size of each subproblem.
    • Caption-conditional generation: QAT-MF achieves better FID while reducing the total time. For continuous conditions, SDOT cannot directly exploit class-wise partitioning to reduce problem size, requiring substantially more computation to ensure convergence, while coupling sampling also becomes more expensive.

    These results demonstrate that QAT-MF not only achieves competitive generation quality compared to AlignFlow but also supports large-scale data coupling with better scalability and computational efficiency.

  • On Q1: Tree Stopping Criterion

    We construct trees until each leaf contains a single sample or no further non-degenerate splits are possible. Theorem 2 shows that deeper trees increase the path-separation gain ΔQAT\Delta_{\mathrm{QAT}}, motivating this choice. We further validate it through a leaf-size ablation on CIFAR-10:

    Leaf SizeEuler-2Euler-100Dopri5
    198.683.913.26
    4105.323.933.31
    16108.943.983.37

    The results confirm that the smallest leaf size achieves the best FID, consistent with Theorem 2. Moreover, tree depth does not affect sampling complexity, as only the Gaussian bounds are stored, enabling O(d)O(d) sampling without tree traversal.

In summary, we hope the above response addresses your concerns and assists you in re-evaluating our work. Please feel free to contact us if you have any further questions.

Reference

[1] Improving and Generalizing Flow-Based Generative Models with Minibatch Optimal Transport. TMLR 2024.

[2] The Curse of Conditions: Analyzing and Improving Optimal Transport for Conditional Flow-Based Generation. ICCV 2025.

[3] Reconstruction vs. Generation: Taming Optimization Dilemma in Latent Diffusion Models. CVPR 2025.

[4] Flow Matching with Semidiscrete Couplings. ICLR 2026.

[5] AlignFlow: Improving Flow-based Generative Models with Semi-Discrete Optimal Transport. ICLR 2026.

Reviewer_xag8OpenReview ↗

Response

Thanks the authors for providing the response.

For the CIFAR-10 settings, I am curious why the authors choose a different experimental configuration of using a different network architecture? Usually, we follow the standard experimental configuration and compare with baselines.

It's good to see that in the class-conditioned setting, the proposed method achieves similar performance as AlignFlow.

Under caption-conditional Generation, the improvement of QAT-MF is small compared to I-MF (24.13 vs 24.82).

AuthorsOpenReview ↗

Response to Reviewer xag8

We thank the reviewer for the prompt response. We address your two additional concerns below.

  • On CIFAR-10 Settings

    We initially conducted experiments using the original OT-FM [1] configuration. The results are as follows:

    MethodEuler-2Euler-100Dopri5NFE
    I-FM170.40 ± 0.184.49 ± 0.053.66 ± 0.04143.30 ± 1.50
    OT-FM90.86 ± 0.124.74 ± 0.073.79 ± 0.05142.27 ± 1.51
    QAT-FM107.38 ± 0.314.69 ± 0.053.68 ± 0.04137.11 ± 0.55

    From these results, we find that:

    • We were unable to reproduce the FID of approximately 3.58 reported in the original OT-FM paper, obtaining 3.79 instead. A similar discrepancy appears in AlignFlow [5], which reports an OT-FM FID of 3.82, close to our result.

    • Neither OT-FM nor QAT-FM outperforms I-FM under Euler-100 or Dopri5. A similar situation may also occur in AlignFlow, as its paper does not report I-FM results.

    • Nevertheless, QAT-FM still outperforms OT-FM under Euler-100 or Dopri5, consistent with the results reported in our paper.

    This motivated us to explore a wider network architecture, under which QAT-FM achieves a better Dopri5 FID than I-FM.

  • On Caption-Conditional Generation

    First, we emphasize that QAT-MF already outperforms AlignFlow in this setting, reducing FID from 28.82 to 24.13 while decreasing the total runtime from 40.80h to 24.20h.

    Furthermore, beyond standard Gaussian sampling, CQAT supports sampling from a condition-dependent source distribution, as newly described in Section 3.3 (P6, Lines 319-334) and Section D.3 (P26, Lines 1373-1403). The full results are reported below:

    MethodTrain FIDTrain CLIPVal FIDVal CLIP
    I-MF24.820.31323.080.313
    AlignFlow28.820.30426.600.304
    QAT-MF24.130.31322.400.313
    QAT-MF (conditional prior)23.650.31421.850.314

    With conditional prior sampling, QAT-MF further improves FID on both training and validation sets, while slightly improving text–image alignment.

Thank you again for your feedback. We hope these additional results address your concerns and support your reconsideration of our work.

Comments on the paper

2 comments
Area_Chair_x8RQOpenReview ↗

Reminder: Please Respond to the Authors’ Rebuttal

Dear Reviewers,

Thank you for your time and effort in reviewing this submission. Please take a moment to read the authors’ responses and follow up with any remaining questions or concerns. If the responses have changed your assessment, please update your review and score accordingly.

Your engagement is greatly appreciated and will help us reach a well-informed decision.

Best,

AC

AuthorsOpenReview ↗

Summary of the Rebuttal and Discussion

Dear ACs, SACs, and PCs,

We sincerely thank you for handling our paper, and we greatly appreciate the reviewers' positive feedback on our work. Given the length of the discussion session, we have summarized the key points below for your convenience.

  • Summary of Our Responses

    During the discussion phase, we have addressed the reviewers' questions point by point. Our main contributions and additional results are summarized as follows:

    1. We clarified the practical role of Theorem 2 in guiding tree construction and further validated the benefits of sufficiently fine tree partitioning through ablation studies. We also provided an additional theoretical upper-bound analysis for Theorem 2.

    2. We provided an additional comparison with AlignFlow using MeanFlow (SiT-B/4). The results demonstrate that QAT-FM achieves comparable or better performance than AlignFlow, with only minimal computational overhead beyond I-FM.

    3. We investigated training–inference consistency under continuous conditions, showing that general augmented OT formulations do not ensure such consistency. To address this issue, we proposed a conditional sampling prior for CQAT, which achieves training–inference consistency and improves experimental results.

    4. We provided additional analyses of generalization, including nearest-neighbor image comparisons, test-set FID, and CLIP scores. The results provide further evidence of the strong generalization capability of QAT-FM.

    Overall, we believe that our responses and additional experiments have substantially strengthened the paper and addressed the reviewers' key concerns.

  • Summary of Reviewer Feedback

    We summarize the reviewers' detailed assessments of our paper and their feedback throughout the discussion phase as follows.

    1. Reviewer xag8 recognized the novelty of our approach and its potential for extension to conditional generation. We addressed the reviewer's concerns regarding parameter settings and numerical results, provided additional comparisons with AlignFlow, and conducted further ablation studies on tree partitioning. The reviewer stated that the experimental results were convincing and raised their score.

    2. Reviewer agWh recognized the computational efficiency and theoretical soundness of QAT-FM, as well as its broad applicability across diverse generative tasks and settings. The reviewer raised questions regarding the relationship to related work, the connection between Theorem 2 and OT, training–inference consistency under continuous conditions, and computational costs. Our responses and additional results further clarified these aspects and strengthened the theoretical and empirical support for QAT-FM. Following our responses, Reviewer agWh stated that all the weaknesses and questions had been resolved and raised their score from 3 to 4.

    3. Reviewer A12M recognized the simplicity and efficiency of our method, the comprehensive experimental evaluation, and its improvements over I-FM in most settings. The reviewer raised several concerns regarding the upper bound and practical implications of the path separation theorem, the learnability of the velocity field, experimental results and robustness across random seeds, training–inference consistency, Gaussian alignment, velocity field ambiguity under conditional generation, baseline configurations, computational costs, and potential memorization. We addressed these concerns point by point and resolved most of them. In the subsequent discussion, the reviewer requested additional 1-NN experiments in the DINOv2 feature space and raised further questions regarding velocity field learnability, the applicability of QAT, and CQAT sampling. We provided detailed responses to all these follow-up questions and additional experimental results. Although the reviewer has not provided further feedback, we believe that the remaining concerns have also been sufficiently addressed.

    We sincerely appreciate the constructive feedback from the reviewers throughout the discussion. These comments have helped us further clarify the theoretical positioning, applicability, and practical advantages of QAT-FM, while also improving the paper's theoretical analysis, experimental coverage, and presentation.

We once again thank the ACs, SACs, PCs, and reviewers for their time and effort, as well as their careful evaluation and constructive feedback on our work.

Best regards,

Authors (Submission ID: 7921)

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.