acceptodds
Under review as a conference paper at ICLR 2027

Generalization Bounds for Flow-matching Generative Models for Intrinsically Low-dimensional Data

Abstract

Despite its remarkable empirical performance, the theoretical properties of flow-matching in terms of generalization accuracy remain largely unexplored. The current state-of-the-art generalization analyses not only impose stringent assumptions on the estimated velocity field, but also scale poorly with the ambient dimension—precisely the high-dimensional setting where flow matching is most commonly applied. In particular, the data distributions targeted by flow matching models, such as natural images, molecular geometries, etc., are commonly hypothesized to concentrate near a low-dimensional structures embedded in the high-dimensional ambient data space. However, existing analyses fail to reflect this geometry in their derived convergence rates. In this paper, we bridge the gap between the theory and practice of flow matching by establishing generalization bounds on the Wasserstein- distance between the learned distribution and the target measure, for any . Specifically, given independent and identically distributed samples from a target distribution and for every , we show that for appropriate network and hyperparameter choices, with probability at least , where denotes the Wasserstein- dimension of the target measure. Our results demonstrate that flow matching naturally adapts to the intrinsic geometry of the data, effectively mitigating the curse of dimensionality: the convergence exponent depends only on the intrinsic dimension, , rather than the ambient dimension of the feature space, yielding convergence rates that remain meaningful in high-dimensional regimes and offering a principled theoretical explanation for the strong empirical performance of flow matching on structured data distributions with significantly relaxed assumptions, compared to the existing literature.

Primary area:
learning theory
Keywords:
Flow Matching, Generalization Bounds, Intrinsic Dimension

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_mqXMOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

This paper presents convergence guarantees for the distribution learned by flow matching involving only the intrinsic dimension of the data. The authors work with a notion of data dimension arising from optimal transport, the pp-upper Wasserstein dimension, and show that the convergence rate of the learned distribution in Wasserstein space depends only on this dimension. This contribution is interesting and a potentially significant step in our understanding of flow models, and gives a result analogous to the geometric guarantees we have for classical manifold learning. The presentation of the problem is well done and the authors identify a clear gap in the literature. However, the paper needs some significant changes to both its mathematical presentation and experiments to be accepted. The theory is overly dense and difficult to follow, and the experiments are limited and do not address the main claims.

The paper is potentially a significant theoretic contribution, but needs much clearer presentation. Given a sufficient author response, I would be willing to increase my score.

Critical strengths
  1. The introduction clearly establishes the need for the work. I really like this motivation and the use of the optimal-transport-based distance to characterise the dimension is very interesting.
Critical weaknesses
  1. For a mostly theoretical paper, I think the results should be much more clearly presented and explained. While I don't doubt your AI use statement, I think the proliferation of AI based results makes it especially important to clarify the point or spirit of a technical assumption or results. For example, I have no idea how to interpret your 'design choices' for α,β\alpha,\beta, and the standard examples are mentioned both before and after their statement. The main theorem is also nearly impossible to read, both in its formatting on the page and its content. Why should the network assumptions and hyperparameters satisfy these laws in nn, and what do they mean in practice? For example, a fuller and more intuitive explanation of the result and its practical consequences would be more useful than the Proof Sketch section, which could be in the appendix.
  2. Since the proof of concept experiment is an illustration on synthetic data, it would be useful to use an example with a more explicit metric than FID. FID is not a perfect metric for image quality, and is confounded by the choice of inception network - it may, for example, respond differently to the different dimensions you construct. It would be clearer to have synthetic data from a deliberately chosen manifold or stratified space where you could estimate the Wasserstein distance explictly? This would also allow you to fix the data distribution while lifting to higher ambient dimenions - I would find this to be more compelling evidence that the intrinsic dimension alone (or mostly) controls the convergence rate. As stated, the experiments do not support your claims, and the fact that lower-dimensional data is easier to learn is not a new result.
Questions for authors
  1. You say there's no Lipschitz assumption but the bounded-weight part of the RN(W,L,B,K)\mathcal{RN}(W,L,B,K) assumption, together with Lipschitz activations, gives only Lipschitz functions. Does this affect your analysis?
  2. Could you discuss the relationship to methods that explicitly exploit low dimensional structure in flow matching, e.g. [1]?

[1] Bamberger, J., Jones, I., Duncan, D., Bronstein, M., Vandergheynst, P., & Gosztolai, A. (2026, April). Carré du champ flow matching: better quality-generalisation tradeoff in generative models.

4 comments
AuthorsOpenReview ↗

Response to Reviewer mqXM

We sincerely thank the reviewer for their careful reading of our manuscript and thoughtful, constructive feedback. We particularly appreciate the reviewer's recognition that the work is well-motivated and potentially significant. Below, we respond to the reviewer's critiques in details.

Presentation We agree with the reviewer that the paper would benefit by a more detailed presentation and explanations of the main theorem. The network size and the hyperparameters are user-defined choices and the main theorem specifies that with suitable choices of network size, step size, and stopping time, the generated distribution is, with high probability, the Wasserstein error rate depends on the intrinsic data dimension of the target probability distribution.

The standard examples are mentioned both before and after not by mistake, but the latter emphasizes the choice of κ1\kappa_1 and κ2\kappa_2 for these popular choices. We will mention this only once (before D1) for better clarity in the revised draft.

To illustrate how these hyperparameters and network sizes scale in practice we will demonstrate with the linear and VP interpolation scheme and remove the proof-sketch to the appendix, following the reviewer's suggestion.

New Experiments We thank the reviewer for suggesting a more direct way to study the role of ambient dimension. The FID results in Figure 1 illustrate how performance varies across our synthetic image datasets. To complement them, we will conduct experiments with a uniform distribution on a line segment embedded isometrically in spaces of different ambient dimensions. For each embedding, we will vary the training sample size and use the SDE sampler from the paper. This will allow us to examine how the observed error varies with sample size and ambient dimension while keeping the intrinsic geometry fixed. We thank the reviewer for the discussion and will add the experiments in the final version of the paper.

Lipschitz assumption. We thank the reviewer for raising this point. Our statement refers to the true velocity field v⋆v^\star. Under bounded support and the conditions on the interpolation schedule, we show that, for each δ>0\delta>0,

Lip⁡(v⋆(⋅,t))≤Cδ,0≤t≤1−δ\operatorname{Lip}(v^\star(\cdot,t)) \le C_\delta, \quad 0\le t\le 1-\delta

Thus, Lipschitz continuity of v⋆v^\star on the early-stopped interval follows from the analysis, and CδC_\delta may grow as δ\delta decreases. The reviewer is correct that each fitted ReLU network is Lipschitz. Our proof allows its Lipschitz constant to grow with the network size. We will clarify this distinction in the manuscript.

Relation to Carré du champ flow matching. For linear interpolation, our conditional path toward a training point xix_i is txi+(1−t)Zt x_i+(1-t)Z, where Z∼N(0,I)Z\sim\mathcal N(0,I). Bamberger et al. [1] add the term tΓ^i1/2Zt\widehat\Gamma_i^{1/2}Z, so setting their local covariance Γ^i\widehat\Gamma_i to zero recovers our path exactly. To see what this changes, consider data lying on a curve. If Γ^i\widehat\Gamma_i is concentrated along the tangent at xix_i, the noise perpendicular to the curve vanishes as t→1t\to1, while some variation along the tangent remains. Their method thus uses estimated geometry to spread mass locally around each observation. Our theorem, on the other hand, shows how the data's intrinsic geometry can govern the Wasserstein convergence rate even when the interpolation uses no such estimate. It would be interesting to see if the use of such geometric intuitions can benefit flow-matching, theoretically. We thank the reviewer for this interesting discussion and will add this to the final draft of the manuscript.

Thank you for your response and clarifications. I can see I misread the statement about design choices, so thank you for the correction.

While I can see that these amendments would address my concerns about presentation, you have not resubmitted a draft with these changes implemented. Also, while I agree that the proposed line segment experiment would go some way toward improving the empirical evidence for dimension-dependence, I think more substantial synthetic data evidence (or even real data) would be required for ICLR, and these experiments have also not been run. As such, I will maintain my current scores.

AuthorsOpenReview ↗

Revised Version

Dear Reviewer,

We thank you for your follow-up. We understand your concern that the proposed changes and experiments were insufficient without an updated manuscript and completed results.

We are providing the revised manuscript with this response. It incorporates the presentation changes discussed above, including a clearer explanation of the main theorem, the interpolation design choices, and the scope and limitations of the guarantee.

We have also added synthetic experiments beyond the line-segment example. In particular, we study a uniform five-dimensional torus embedded in ten dimensions, training a neural velocity field by conditional flow matching and sampling with the SDE. We evaluate six training sample sizes from 64 to 2,048, with five training seeds and error bars. The estimated Wasserstein-2 error has an observed log–log slope of −0.189, compared with the reference slope −1/5. We explicitly acknowledge that this finite-sample comparison does not establish the asymptotic rate. The line-segment experiments additionally examine changes in ambient dimension while preserving the target geometry.

The full BigGAN SDE evaluation remains underway. We appreciate your assessment and would be grateful if you could consider the revised manuscript and the additional completed experiments.

Thank you for providing these updated results. I appreciate the effort to improve the empirical evidence for the paper and think that it has improved.

However, I feel that the presentation remains a significant issue. The core of the paper is the main theorem and design choices, which remain completely impenetrable to a human reader. The very large number of technical assumptions and details of each constant provide no intuition. For example, the exponent for n in the rate for W can be enormous when D is high - what does this mean in practice? It seems that this has important consequences that are not discussed.

The same can be said for the new experiments, which show positive results but are not explained clearly enough for the reader. The details of the torus experiment, while providing important empirical evidence for the result, focus on bookkeeping not intuition. We don't need another explanation of flow matching and every hyperparameter choice (which can be in the appendix), but do need an explanation of what a 'Sobol reference' is, why it's appropriate, etc. This section should also appear after the theorem it is testing.

In light of these issues, I don't think the paper is yet ready for ICLR and will maintain my score.

Reviewer_2s5NOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the statistical accuracy of flow-matching generative models when the data distribution is intrinsically low-dimensional. Assuming i.i.d. samples supported on [0,1]D[0,1]^D, the authors train a ReLU velocity network with the time-discretized conditional flow-matching objective. Samples are generated with an equivalent SDE, discretized by Euler–Maruyama and stopped early at time 1−δ1-\delta. The main result (Theorem 6) gives a high-probability bound on the Wasserstein-pp distance, for every p≥1p \ge 1, between the (truncated) generated distribution and the data distribution. The bound is of order n−1/d+n−1/(2p)log⁡1/(2p)(1/ξ)n^{-1/d} + n^{-1/(2p)}\log^{1/(2p)}(1/\xi) for any dd larger than the upper Wasserstein dimension dp∗(Pdata)d_p^*(P_{\text{data}}). The rate exponent therefore depends on the intrinsic rather than the ambient dimension. For d~\tilde d-regular supports, the authors argue this matches the minimax rate up to an arbitrarily small slack in the exponent. A synthetic BigGAN experiment comparing intrinsic dimensions d∈{10,100}d \in \{10, 100\} is provided as empirical support.

Critical strengths
  1. Weak distributional assumptions. Beyond bounded support, the analysis requires no density, smoothness or manifold assumption on the data distribution. It also requires no Lipschitz assumption on the learned velocity field, which is avoided by analyzing an SDE sampler via Girsanov’s theorem. The regularity of the target velocity field is derived rather than assumed.
  2. A strong form of guarantee. Prior flow-matching analyses are largely restricted to W1W_1 or W2W_2 and in-expectation bounds. In contrast, this paper provides high-probability Wasserstein-pp guarantees for every p≥1p \ge 1.
  3. End-to-end analysis. Discretization, early stopping, approximation and Monte Carlo error are all accounted for. The paper gives explicit prescriptions for δ\delta, hh, MM, WW, LL, BB, KK and mm, and the main text acknowledges that the exponent of the network size grows linearly in DD.
Critical weaknesses
  1. Unclear novelty relative to Chakraborty et al. (2026). The paper cites Chakraborty, Berthet & Bartlett (2026), “Generalization Properties of Score-matching Diffusion Models for Intrinsically Low-dimensional Data,” but never explains what is technically new relative to that work. The error decomposition, rate, minimax discussion and experimental setup closely follow it. The key rate step (Lemma 20) relies on its Theorem 10. The genuinely new components appear to be the velocity-to-SDE conversion (Theorem 13), the Girsanov/Euler–Maruyama analysis and the high-probability bound, but the paper does not identify them as such. Moreover, lines 368–370 claim to relax the compact-support assumption of that line of work. This is inaccurate: the diffusion paper requires only a finite moment, whereas Assumption 1 here requires support in [0,1]D[0,1]^D. An explicit comparison is needed.
  2. The curse of dimensionality is moved into the network size. The exponent of nn in the prescribed width grows linearly in DD. The proof of Theorem 6 also contains a factor (Dlog⁡n)D/2(D\log n)^{D/2} (line 1376) that is omitted from the theorem statement. Further DD-dependent factors, such as (8D)D(8\sqrt D)^D from the choice of ε\varepsilon in Lemma 9, are absorbed into constants on the grounds that DD is fixed. Statements such as “effectively mitigating the curse of dimensionality” in the abstract are therefore misleading. They should be qualified to say that only the exponent of the statistical rate is intrinsic.
  3. The experiment does not test the theory. The analysis requires an SDE sampler with σ>0\sigma > 0, whereas the experiment uses the deterministic probability-flow ODE. The experiment also varies the intrinsic dimension dd at fixed DD, so it cannot show that the rate is independent of DD. Varying DD at fixed dd would be the relevant test. In addition, for d=100d = 100 the theory predicts essentially no improvement between n=1,000n = 1{,}000 and n=10,000n = 10{,}000 (since 10−1/100≈0.9810^{-1/100} \approx 0.98), yet the reported FID drops from roughly 350 to 200.
Questions for authors
  1. Could you please address the novelty relative to the diffusion literature.
  2. Please make the dependence on DD explicit in all constants.
  3. Can the experiment be rerun with the analyzed SDE sampler?
4 comments
AuthorsOpenReview ↗

Response to Reviewer 2s5N (Part 1)

We sincerely thank the reviewer for their careful reading of our manuscript and thoughtful, constructive feedback. We particularly appreciate the reviewer's recognition that the work provides a strong end-to-end analysis under minimal assumptions. Below, we respond to the reviewer's critiques in details.

Novelty Chakraborty et al. (2026) establish an intrinsic-dimensional Wasserstein rate for a reverse Ornstein–Uhlenbeck diffusion trained by denoising score matching. We share their broad proof strategy of deriving an oracle inequality and controlling its terms. Our contribution is to show that a velocity network trained by conditional flow matching attains the same statistical benchmark when used in the discretized stochastic sampler of Eq. (9).

The distinction is already visible for the linear interpolant It=tX+(1−t)ZI_t=tX+(1-t)Z. Conditional on XX, a single Gaussian draw ZZ determines the entire interpolation path. The forward Ornstein–Uhlenbeck process, by comparison, receives independent Brownian increments over disjoint time intervals. Theorem 13 establishes that, for the exact velocity v⋆v^\star, the flow ODE and our SDE sampler have the same one-time marginal distributions. The next step is to quantify what happens when the velocity is estimated and evaluated only on a time grid. For linear interpolation, replacing v⋆v^\star by v^\widehat v at a fixed (x,t)(x,t) changes the SDE drift by $\left(1+\frac{\sigma^2 t}{2(1-t)}\right) \bigl(\widehat v(x,t)-v^\star(x,t)\bigr).$ Thus, although the exact score and velocity are algebraically related, the sampler amplifies velocity estimation error near t=1t=1. The proof must control this amplification up to the stopping time.

Lemma 7 applies Girsanov’s theorem to compare the ideal SDE with the discretized learned sampler, separating the weighted velocity regression error from the discretization error. Lemma 8 bounds the discretization terms, and Lemma 14 controls the additional regression error from Monte Carlo approximation of the training loss. The conditions on (αt,βt)(\alpha_t,\beta_t) control the growth and time variation of the drift coefficients before stopping. We derive the required regularity of v⋆v^\star, and the analysis allows the fitted network’s Lipschitz constant to grow with nn.

Together, these estimates yield choices of stopping time, grid spacing, network size, and Monte Carlo sample size that preserve the n−1/dn^{-1/d} Wasserstein rate at fixed confidence for every d>dp∗(Pdata)d>d_p^\ast(P_{\mathrm{data}}). For distributions with a density on a kk-dimensional regular support with k>2pk>2p, this matches the minimax exponent up to arbitrarily small slack. We thank the reviewer for prompting this comparison and will make the flow-matching-specific steps explicit in the revision.

Network Size We agree with the reviewer that the required network size can be large as DD increases. Theorem 6 displays the terms as a function of nn only as is standard in the theoretical error bounds for generative models (Tang and Yang, 2025; Chakraborty and Bartlett, 2025) etc.) Following the reviewer's suggestion, we will make the book-keeping of these nn-independent terms more explicit. The curse of dimensionality is mitigated in the behavior of the sample size for both the final rate (O(n−1/d\mathcal{O}(n^{-1/d}) and the network size (depending only polynomial in nn). In comparison to SOTA, our SDE analysis also avoids the exponential dependence on the fitted velocity’s Lipschitz constant LxL_x that appears in ODE stability bounds such as Gao et al. (2024, Lemma 4.2). If LxL_x grows linearly with nn, that factor can lead to an exponential sufficient parameter budget. Our bound remains polynomial in nn without requiring uniform Lipschitz control of the fitted velocity. We thank the reviewer for raising this point and will make the dimension-dependent factors explicit in the revision.

AuthorsOpenReview ↗

Response to Reviewer 2s5N (Part 2)

Experiments We ran Eq. (9) on a controlled target, PD=L(UuD)P_D=\mathcal L(Uu_D) with UU uniform on [0,1][0,1] and uD=(1,…,1)/Du_D=(1,\ldots,1)/\sqrt D. This keeps all target distances fixed as DD varies. We fitted a two-hidden-layer ReLU velocity and sampled the same fitted weights with both the SDE and the ODE. At n=512n=512, the results below give W2W_2 from 2,048 generated points to the continuous target, averaged over five training seeds (mean ± sample SD).

Ambient DDSDE, σ=0.25\sigma=0.25ODE
20.092 ± 0.0010.093 ± 0.003
80.252 ± 0.0050.259 ± 0.006
160.503 ± 0.0120.528 ± 0.012

Both samplers used 96 steps to t=0.96t=0.96. Error increased with DD at this fixed training and sampling budget. We have started an SDE rerun of the BigGAN experiment. It should be noted that Theorem 6 controls the Wasserstein distance between the distribution generated by the stochastic sampler in Eq. (9) and the target distribution in image space. Figure 1 uses the deterministic ODE sampler and reports FID, which compares Gaussian fits to the Inception features of generated and reference images. The decrease in FID from roughly 350 to 200 therefore shows improved agreement in feature space. We will revise the text to present this as a qualitative observation and remove the claim that Figure 1 validates the theoretical rate.

AuthorsOpenReview ↗

Revised Version

Dear Reviewer,

We have revised the manuscript to incorporate most of your suggestions. The remaining experiments are currently underway. We sincerely thank you for your thoughtful and constructive feedback, which has helped us improve the overall quality of the paper.

Thanks for the updated results. I think they greatly improve the paper. I couldn't find the results for the remaining experiments, therefore, I will maintain my current scores for now.

Reviewer_rob2OpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The manuscript discusses the Wasserstein distance between a given probability distribution μ\mu and the approximation by a flow matching algorithm with SDE sampling. In the training it is assumed that nn samples are available, which they assume are i.i.e. generated. Furthermore they assume that their data distribution μ\mu has intrinsic dimension dd which is smaller than the ambient dimension DD. Their main result is a statement about W2(μ,μθ)W_2(\mu,\mu_\theta) where μθ\mu_\theta is the measure generated by the velocity field trained on nn training points. They then show for any pp, that with probability at least 1−ξ1-\xi it holds that W2(μ,μθn)≲n−1/d+n−1/2plog⁡1/2p(1/ξ)W_2(\mu,\mu_\theta^n)\lesssim n^{−1/d} + n^{− 1/2p} \log^{1/2p} (1/\xi). Here μθn\mu_\theta^n depends on a network with parameter count that is itself dd depended.

Critical strengths

The results of the paper seem to be correct and the paper is in general understandable. Furthermore the mathematics is solid.

Critical weaknesses

Major Weaknesses The claims in the abstract and introduction are too strong for what the paper in its current form shows. In particular:

  • "In this paper, we bridge the gap between the theory and practice of flow matching by establishing generalization bounds on the Wasserstein-p distance between the learned distribution and the target measure, for any p ≥ 1." It is not clear if the bounds in the paper have are relevant in practice, since the number of parameters of their theoretical network scales with n(5(D+2)/d)n^(5(D+2)/d) at least. Also for their practical implications, the number of noise samples they need scales similarly. Both seem to be extremely out of reach for any practical implications.
  • This also makes the claim "offering a principled theoretical explanation for the strong empirical performance of flow matching on structured data distributions" questionable.

They cite the paper of Chakraborty et al. "Generalization Properties of Score-matching Diffusion Models for Intrinsically Low-dimensional Data", but it seems they do not point out enough the similarities and differences. Section 3 seems to be exactly the same experiment as Section 2 in the Chakraborty et al. paper, just with flow matching instead of diffusion. Also Chakraborty et al. uses the same intrinsic dimension, the same network structure and W,L,B for a network. Their results look very similiar, just that Chakraborty et al. use the expectation instead of a "with probability at least 1−ξ1-\xi".

Given that amount of similarity, the novelty and technical difficulty seem to be not sufficiently strong for ICRL, when compared to the existing literature.

Questions for authors

Questions and minor weaknesses

  • Theorem 4 is known under the name "closed form" velocity field, and the corresponding literature should be cited
  • There is some notation overload at the beginning of page 4, XtX_t is used for both, the process Xt=(1−t)X0+tX1X_t=(1-t)X_0+t X_1 as well as the process driven by the vector field vtv_t.
  • in line 370 they state "while relaxing the compact-support and manifold-structure assumptions common to both lines of work". but Assumption 1 states Pdata([0,1]D)=1P_{data}([0,1]^D)=1.
  • The paper is partly hard to read due to notation overload. E.g. they in line 133 two notations for the cal(O)cal(O) notation are introduced. Definition 2 is overly complicated and parts could be put into the appendix. D.1 as well as Theorem 6 are also overloaded with notation, parts of it could be done in the appendix.
  • It should be made clear that SDE sampling is used already in the introduction or even the abstract.
  • Also in the introduction it should be made clear that the term "generalization" is not used in contrast to memorization, because perfect memorization fulfills their main theorem
  • in Section 3, what is the FID when resampling from the BigGAN? E.g. for dimension d=10d=10 one samples to distinct sets with the same number of elements used in Figure 1, and then computes the FID, and the same for d=100d=100. Then one could compare which effect might come just from the difference of the ground truth distributions, and which is related to the difficulty in learning.
  • Can you estimate the constants and network sizes for Theorem 6 for a realistic example like ImageNet generation?
6 comments
AuthorsOpenReview ↗

Response to Reviewer rob2 (Part 1)

We sincerely thank the reviewer for their careful reading of our manuscript and thoughtful, constructive feedback. We particularly appreciate the reviewer's recognition that the work is mathematically solid. Below, we respond to the reviewer's critiques in details.

Network Size We agree with the reviewer and, as acknowledged in our manuscript, that the derived theoretical network size and the number of noise samples scale unfavorably with DD. However, this is a sufficient bound on the network size that covers all unfavorable and irregular distributions on [0,1]D[0,1]^D and is not a necessary condition. It remains open as to whether this bound can be improved for better-behaved distributions. This unfavorable scaling primarily occurs due to the velocity-field approximation, and since we are not imposing any regularity conditions (e.g., uniform Lipschitzness) on the true velocity field, such large network size becomes unavoidable to cover all irregular distributions. Similar results in the generative model literature, such as those by Huang et al. (2022), Chakraborty and Bartlett (2025), Chakraborty et al. (2026) all have a similar drawback. While such unfavorable scaling limits the practical utility of the derived bounds for worst-case distributions, we believe our work will act as a stepping stone for refined bounds in the future.

Based on the reviewer's suggestions, we will change those phrases and highlight that the derived bounds under minimal assumptions might render future refined analysis with better practical implications.

Experiments We agree that this experimental setup appears in Chakraborty et al. (2026), Chakraborty and Bartlett (2024) etc. The experimental results presented in the paper are to motivate how the error rates should behave with the intrinsic data dimension: lower intrinsic dimension implies faster generalization, which the theory corroborates. Exact experimental verification of the theory is infeasible for two reasons: (i) Estimating the Wasserstein distance empirically in a high-dimensional image space is a difficult problem, and (ii) estimating the Wasserstein dimension for image distributions becomes difficult (see Block et al. (2022)). For this reason, we resort to proxies such as the FID score and latent space dimensions for computational ease.

Novelty We would respectfully disagree with the reviewer that the novelty and the technical difficulty are not strong enough. As highlighted in the paper, our analysis gives the fastest known error rates for flow matching under minimal assumptions. This rate should be similar to that by Chakraborty et al. (2026) as it is minimax optimal (see Theorem 29 of Chakraborty and Bartlett (2025)).

The proof philosophy broadly follows the standard prescription for deriving generalization bounds for generative models: (i) establish an oracle inequality and (ii) control the individual terms. For comparison, Chakraborty et al. (2026) establish an intrinsic-dimensional Wasserstein rate for a reverse Ornstein–Uhlenbeck diffusion trained by denoising score matching. Our contribution is to prove that a velocity network trained by conditional flow matching can attain the same minimax-optimal benchmark when used in the discretized stochastic sampler of Eq. (9). The distinction is already visible for the linear interpolant It=tX+(1−t)ZI_t=tX+(1-t)Z. Given XX, a single Gaussian draw ZZ determines the entire interpolation path. The forward Ornstein–Uhlenbeck process receives independent Brownian increments over disjoint time intervals. For the linear interpolant, substituting v^\widehat v for the exact velocity v⋆v^\star changes the sampler’s drift by (1+σ2t2(1−t))(v^(x,t)−v⋆(x,t)).\left(1+\frac{\sigma^2 t}{2(1-t)}\right)\bigl(\widehat v(x,t)-v^\star(x,t)\bigr).

Although the exact score and velocity are algebraically related, this factor amplifies velocity estimation error as t→1t\to1. Transferring the statistical rate therefore requires controlling the amplified error up to the stopping time.

Lemmas 7–8 quantify how velocity estimation error and time discretization affect the endpoint law. Lemma 14 controls the additional regression error from approximating the conditional flow-matching loss by Monte Carlo. Together, these bounds give choices of stopping time, grid spacing, network size, and Monte Carlo sample size that preserve the n−1/dn^{-1/d} Wasserstein rate at fixed confidence for every d>dp∗(Pdata)d>d_p^\ast(P_{\mathrm{data}}). We derive the required regularity of v⋆v^\star and allow the fitted network’s Lipschitz constant to grow with nn. For distributions with a density on a kk-dimensional regular support, k>2pk>2p, the resulting rate matches the minimax exponent up to arbitrarily small slack.

We thank the reviewer for this discussion and will add this discussion in the final version.

References

Block, A., Jia, Z., Polyanskiy, Y., & Rakhlin, A. (2022). Intrinsic dimension estimation using Wasserstein distance. Journal of Machine Learning Research, 23(313), 1-37.

AuthorsOpenReview ↗

Response to Reviewer rob2 (Part 2)

Answers to Raised Questions

  1. Theorem 4 goes beyond merely recognizing the form of the velocity field as done in the existing literature, e.g. Kumar et al. (2026), Appendix F.1, Lemma 3.1 of Gao et al. (2024), who do not provide a rigorous derivation; Theorem 4 establishes the existence and well-definedness of the IVP in equation (4) for any initialization. Following the reviewer's suggestion, we will make this distinction in the final version. We will also cite Albergo et al. (2025), whose analysis is performed for stochastic interpolators.

  2. We apologize for the notational overload. We will use a different notation, e.g., X\mathcal{X} for the process driven by the velocity field.

  3. We apologize for this oversight. We will correct this in the revised version.

  4. To mitigate the notational overload, we will add a table of important notations at the beginning of the appendix. The two standard notations O\mathcal{O} and O~\tilde{\mathcal{O}} are necessary, as the latter ignores the poly-log factors in nn. Having a table of notations will make such distinctions clearer to the reader. We will also trim down Definition 2 and Theorem 6 as suggested by the reviewer.

  5. We will cite the relevant literature as cited in lines 233-234 in the introduction to clarify the use of SDE sampling for flow matching.

  6. We thank the reviewer for suggesting this comparison. We will revise Figure 1 to include a BigGAN resampling baseline for each class and each latent dimension (d=10d=10 and d=100d=100). For each baseline, we will compute FID between two independently drawn sets of 1,000 BigGAN images, using the same preprocessing and ten repetitions as in Figure 1. We will plot the mean baseline FID and its variability alongside the corresponding flow-matching curve. This will show how much FID varies across the two target distributions even when both image sets come directly from BigGAN, providing a reference for interpreting the learned models’ scores.

  7. It is difficult to estimate a practical network size for ImageNet from Theorem 6 because the dimension-dependent constants are implicit. The bound is conservative because it requires uniform approximation of the velocity over the entire ambient cube. We will clarify this limitation in the revision.

AuthorsOpenReview ↗

Revised Version

Dear Reviewer,

We have revised the manuscript to incorporate most of your suggestions. The remaining experiments are currently underway. All changes in the manuscript are marked in blue. We sincerely thank you for your thoughtful and constructive feedback, which has helped us improve the overall quality of the paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.