acceptodds
Under review as a conference paper at ICLR 2027

RW-Flow: One-Step Generation on Compact Manifolds via Wasserstein Gradient Flows

10 views

Abstract

Manifold-valued data, and consequently the distributions they induce, are prevalent across many domains, ranging from the locations of geospatial events, such as earthquakes, to biomolecular torsion angles that encode information about three-dimensional structure. While diffusion and flow-based generative models have been successfully extended to compact manifolds, sampling typically requires tens or hundreds of sequential network evaluations. We introduce RW-FLOW, a theoretically grounded framework for learning one-step generative models on compact manifolds via Wasserstein gradient flows. The main challenge is identifiability: driving the velocity field to zero should guarantee that the model distribution matches the target distribution. We establish a necessary and sufficient condition for identifiability on compact, connected Riemannian manifolds. We specifically show that, for a symmetric, Lipschitz-continuous cost function, the velocity field induced by the Sinkhorn divergence is identifiable if and only if the associated Gibbs kernel is nondegenerate. This characterization provides a general principle for designing identifiable costs on compact manifolds. It also reveals that the squared geodesic distance, the natural manifold analogue of the squared Euclidean distance, does not always guarantee identifiability. Across benchmarks involving geospatial events, protein side chain torsion angles, RNA backbone torsion angles, and general manifolds discretized as triangular meshes, RW-FLOW outperforms existing one-step methods in nearly all settings under fair comparison conditions.

Primary area:
generative models
Keywords:
W-Flow, Wasserstein Gradient Flows, Generative Modeling

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_tKDMOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces RW-Flow, which extends W-Flow for learning one-step flow matching methods on Riemannian manifolds. The method is motivated by the question of when a vanishing Sinkhorn divergence induced velocity field shows that the model distribution can match the target. They find that this can be achieved by the nondegeneracy of the Gibbs kernel. They then use this to investigate spectral and distance costs of the squared geodesic cost.nts, showing the limitations.

They evaluate their method on modeling Riemannian manifolds, such as the sphere, protein/RNA angles on tori, and triangular meshes. RW-Flow seems competitive with other Riemannian Generative Model baselines.

Critical strengths
  1. I think this paper investigates an interesting question. Understanding when a vanishing velocity field matches the target distribution is an important problem, especially for modeling distributions on Riemannian spaces. The connection to the Gibbs kernel and the analysis of different cost families make the theoretical part substantial and useful.
  2. The experimental part covers a bunch of Riemannian datasets and supports the theoretical result. It shows that the proposed method is comparable to baselines.
Critical weaknesses
  1. I wonder if there could be more detailed comparisons of training overhead, inference latency, and geometry costs against other methods; for example, Algo 1 uses two minibatch Sinkhorn calculations, and the spectral variants additionally require evaluating spectral kernels and their gradients.
  2. I also suggest including more figures, e.g., some teaser figures to show the motivation and the overall framework of the method; this could help readers understand.
Questions for authors
  1. Could you include the per-update training time and GPU memory usage, comparing RW-Flow with other methods, such as GFM RMF baselines?
  2. For the spectral variants, especially on meshes, what preprocessing is required? How much time and memory does it use, and can it be reused across runs?
  3. Can you provide head-to-head comparisons on sampling latency with other few step generation baselines ?
2 comments

We thank the reviewer for their time and feedback. We have added per-update training time, GPU memory, sampling latency and mesh preprocessing cost, with comparisons to the other baselines. We answer each question below.

Q1, Q3. Head-to-head comparisons of training costs and sampling latency

In the table below, we include the per-update training time, GPU memory usage, sampling latency, along with other methods on the Earth Fire dataset. As shown, the training time for our RW-Flow is comparable to RMF and faster than the best performing GFM baseline (GFM Lagrangian). Sampling latencies of all methods are around the same at 1 NFE.

MethodTrain time / update (ms) ↓Peak train memory (MiB) ↓Sampling latency, 1 NFE (ms) ↓
RW-Flow-SG16.907640.72
RW-Flow-C16.987760.73
RW-Flow-G17.727640.72
RW-Flow-H19.287640.73
RW-Flow-M19.257640.72
RW-Flow-S19.157640.73
GFM-L29.746200.63
GFM-E31.396200.80
GFM-S18.225161.06
GFM-MF19.216580.60
RMF-Lx28.735301.56
RMF-Lv22.624121.44
RMF-Ex28.365261.72
RMF-Ev23.064201.26
RMF-Sx28.164741.33
RMF-Sv25.984440.88
KGD-GG17.375880.75
KGD-GL17.465880.70
KGD-M55.0813060.85
RCM-CT13.546140.95
RCM-sCT11.176140.94
MMDF-G13.617400.73
MMDF-L11.037400.72
MMDF-H11.417420.72
MMDF-M11.907420.73
RFM6.854760.71

Q2. Preprocessing costs

Our method requires an one-time mesh preprocessing step that computes the first 1,000 eigenfunctions. Once a mesh has been preprocessed, the result can be reused across all subsequent runs. For the Spot (cow) mesh, this step takes about 20 seconds and 500 MB of GPU memory. The other manifolds do not require such preprocessing since the eigenfunctions on those surfaces are explicitly available.

Since the eigenfunctions are piecewise linear on the mesh, their gradients are constant on each face and precomputed in closed form, so the kernel gradient is computed analytically, without automatic differentiation, and adds negligible compute to training.

Conclusion

We hope these additions address the reviewer's questions about cost. We plan to add a teaser figure in the camera-ready version. We are happy to discuss further.

I appreciate the authors’ response and their careful answers to my questions. I will keep my positive score.

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

The paper extends the W-Flow [1] to Riemannian manifolds M\mathcal{M} by taking the (distribution-dependent) drift vector field Vq,pV_{q,p} to be the intrinsic (manifold-restricted) Wasserstein gradient of the Sinkhorn divergence (Theorem 3.1). It further establishes conditions for when this drift field is identifiable, i.e. when Vq,p=0⇒p=qV_{q,p} = 0 \Rightarrow p = q, in terms of the integral operator corresponding to the Gibbs kernel k=exp⁡(−c/ϵ)k = \exp(-c/\epsilon). The main result (Theorem 3.2) states: Vq,pV_{q,p} is identifiable iff this integral operator is injective on signed measures on M\mathcal{M}. The following conditions are derived from this: When the integral operator is diagonalizable w.r.t. Laplace-Beltrami eigenbasis, then identifiability is equivalent to the non-zero eigenvalues w.r.t. this basis (Lemma 3.3). On the dd-sphere and dd-torus the Gibbs kernel corresponding to the squared geodesic distance cost implies an identifiable drift for almost all entropy regularization parameters ϵ>0\epsilon > 0 (Proposition 3.3.2). Further the authors introduce the notion of spectral kernels and further show identifiability for distance-based kernels with chordal distance ∣x−y∣2|x-y|^2 for any manifold and (non-squared) geodesic distance on the sphere via the notion of kernel universality. Finally the work provides some numerical experiments for the RW-flow and compares it with other one- and multi-step methods.

Critical strengths
  • The reviewer liked the if-and-only if connection between the injectivity of the integral operator and identifiability and connection to the spectral decomposition. Notably previous work has only established sufficiency for similar statements (e.g. [2, Proposition 3.7] or [3, Theorem 3.13] or [4, Proposition 5.1]). This work also does not require the kernel to be positive definite.
  • RNA experiments show significant improvements (i.e. in Table 4 kMMD is half of the best competitor). For the mesh experiment the method shows substantial speed advantage.
Critical weaknesses

There are multiple weaknesses that needs to be addressed:

  • As in other papers on drifting models and W-Flows, the question is why identifiability is relevant: Since the loss (in (2) or (5)) contains the stopgrad operator, a gradient descent on this loss does not correspond to the minimization this loss, so in particular is it unclear why we would care about its zeros. While it is true that qq such that Vq,p=0V_{q,p} = 0 are fixed points of the intended population dynamics, when we parametrize qq via a pushforward fθf_\theta to qθq_\theta, additional stationary points θ⋆\theta^\star can be introduced that do not imply Vqθ⋆,p=0V_{q_{\theta^\star},p} = 0. Conversely, if we want qθ=pq_\theta = p to imply a stationary point θ\theta, then we only need the direction q=p⇒Vq,pq = p \Rightarrow V_{q,p} which readily holds for many drifting fields. Thus identifiability is neither necessary, nor sufficient for stationary point θ\theta to imply qθ=pq_\theta = p.
  • The authors should place their work in the context of [2,3,4].
  • The current theory is disconnected from the numerical experiments: An identifiable Vq,pV_{q,p} is not directly related to the performance of RW-Flow. The authors should e.g. run experiments comparing identifiable and non-identifiable drift fields and show that non-identifiable drift fields yield empirically subpar performance or get stuck in a stationary point not corresponding to the true distribution pp.
  • The statistical evidence for the superiority of RW-Flow in some experiments is weak, e.g. in Table 3/Table 9 several gains are small relative to the reported run-to-run variability: For the kMMD the runs for Volcano, Earthquake, Flood, Fire are essentially tied to the strongest competitor. The authors should mention/explain that.

[1] Jiaqi Han, Puheng Li, Qiushan Guo, Renyuan Xu, Stefano Ermon, and Emmanuel J Cande`s. One-step generative modeling via Wasserstein gradient flows. arXiv preprint arXiv:2605.11755, 2026 [2] He, Ping, et al. "Sinkhorn-drifting generative models." arXiv preprint arXiv:2603.12366 (2026). [3] Lee, HakGeun, and Hyonho Chun. "Identifiability and Stability of Generative Drifting with Companion-Elliptic Kernel Families." arXiv preprint arXiv:2604.24196 (2026). [4] Turan, Erkan, Nicolas Dufour, and Maks Ovsjanikov. "Generative drifting is secretly score matching: a spectral and variational perspective." arXiv preprint arXiv:2603.09936 (2026).

Questions for authors

See weaknesses

7 comments
AuthorsOpenReview ↗

Rebuttal Part 1/2

We thank the reviewer for their time and effort in evaluating our manuscript. The point about stop-gradient stationarity under a finite parametrization is well taken. It has led us to state the scope of our identifiability result more carefully and to add an experiment that directly compares identifiable and non-identifiable costs (see also the General Response and App. C of the revision). We address each point below.

Q1. Why is identifiability relevant?

It is true that for a finite-dimensional parametrization fθf_\theta, stationarity of θ\theta only gives Ez[Jθ(z)⊤Vqθ,p(fθ(z))]=0\mathbb{E}_{z}[J_\theta(z)^\top V_{q_\theta,p}(f_\theta(z))]=0 (Lai et al. [2026]), which does not imply Vqθ,p=0V_{q_\theta,p}=0. Identifiability alone therefore cannot guarantee qθ∗=pq_{\theta^\ast}=p for a fixed architecture, and we now state this explicitly (App. C). However, we still believe it is a very relevant population-level condition, for two reasons.

1. Necessity. The direction q=p⇒Vq,p=0q=p\Rightarrow V_{q,p}=0 only says that pp is a fixed point, not that it is the only one; the trivial field V≡0V\equiv 0 satisfies it. If identifiability fails, there is q≠pq\neq p with Vq,p=0V_{q,p}=0, and every θ\theta with qθ=qq_\theta=q is stationary for every architecture that can represent qq (Lai et al. [2026]). These spurious solutions come from the field, not from the specific network parametrization.

2. Sufficiency in the nonparametric limit. The gap raised by the reviewer is due to the limited representation capability of neural networks. In the Euclidean case, if the capacity of the network ff is unlimited, stationarity is equivalent to Vqf,p=0V_{q_f,p}=0 qfq_f-a.e. (cf. Shen et al. [2026, Thm. 1]). The remaining gap lies in the representation capability of the neural network, and is shared by all deep learning frameworks including GANs, score matching and flow matching, whose guarantees are also stated at the population optimum. It does not depend on the choice of field.

In short, identifiability removes the spurious stationary points caused by the drift field and leaves only those caused by the parametrization. This is one part of a broader picture, which we give in App. C of the revised paper and summarize in the General Response. There, each step from a stationary point to qθ=pq_\theta = p is tied to the assumption it needs, which shows clearly where identifiability fits, along with the new experiment comparing identifiable and non-identifiable costs (App. C.2).

Q2. Related Works

We thank the reviewer for these pointers. While these works are not yet peer-reviewed, we have added a discussion of them to Section 4 (Related Work) of the revised paper, as they give readers useful context. Below, we summarize how RW-Flow relates to each. All these works consider Euclidean generative models. In contrast, RW-Flow operates on compact Riemannian manifolds, and we establish sufficient and necessary conditions for the Wasserstein gradient of the Sinkhorn divergence to be identifiable compared to prior works showing sufficiency only.

  • [2] He et al. propose using the Wasserstein gradient of the Sinkhorn divergence as the drift. In essence, their method is the Euclidean analogue of ours, restricted to the Sinkhorn divergence. As the reviewer noted, they establish sufficient conditions under which the Sinkhorn divergence with the squared Euclidean cost induces an identifiable field.
  • [3] Lee and Chun study the identifiability and stability of the original drifting field for companion-elliptic kernel families. These include the Laplacian kernel, which Deng et al. [2026] originally used.
  • [4] Turan et al. show that drifting with a Gaussian kernel is equivalent to score matching, and they prove identifiability for this kernel. They also show that the drifting field can be interpreted as an approximate Wasserstein gradient field of a smoothed KL divergence, and that the drifting formulation corresponds to a JKO discretization of the underlying Wasserstein gradient flow. Finally, they propose the Sinkhorn drift.

Q3. Empirical Experiment showing subpar performance of non-identifiable costs

For a clear comparison, we run an experiment on the circle S1S^1, comparing an identifiable cost $d_g $ and non-identifiable cost 13∥x−y∥3\frac13\|x-y\|^3 . The target distribution is an one-dimensional checkerboard on the circle. Please refer to App. C.2 for figures.

We report the proportion of generated points that lie on the support of the target distribution. The quantitative results are shown below.

CostMass on support
Geodesic0.9124 ± 0.01218
Chordal, p=3p=30.02352 ± 0.007006

As can be seen both qualitatively and quantitatively, we emphasize not all costs are equal, and identifiability is one principled way to choose one.

AuthorsOpenReview ↗

Rebuttal Part 2/2

Q4. Statistical tie in kMMD

On Earth datasets, RW-Flow and the strongest baseline, GFM-Lagrange, are statistically indistinguishable in terms of kMMD. However, as discussed around line 400, the qualitative comparison in Figure 2 shows that, under equal compute, RW-Flow captures fine-scale structure that GFM-Lagrange misses. We attribute this discrepancy to a limitation of kMMD. At the bandwidth commonly used, it is insensitive to fine-scale differences. To illustrate this, the table below reports kMMD and 1-NNA for the samples shown in Figure 2. At the standard bandwidth, GFM-Lagrange attains lower kMMD on the Volcano and Fire datasets, even though Figure 2 shows that it misses fine-scale detail.

DatasetMethodkMMD ↓1-NNA (→ 0.5)
VolcanoGFM-Lagrange0.124060.58537
RW-Flow subordinated0.131560.48780
EarthquakeGFM-Lagrange0.033050.58415
RW-Flow subordinated0.033280.53186
FireGFM-Lagrange0.026590.71797
RW-Flow subordinated0.028360.59258
FloodGFM-Lagrange0.048500.57598
RW-Flow subordinated0.044070.53799

This is one reason we report COV, 1-NNA, and MMD in addition to kMMD, whereas prior works (Davis et al. [2026], Woo et al. [2026], Esteban-Casadevall et al. [2026], Cheng et al. [2025]) take kMMD as their primary metrics in similar experiments.

We have revised the paper to explain the statistical tie on kMMD.

Conclusion

We hope these clarifications and the new experiment address the reviewer's concerns about the relevance of identifiability. We would be glad to discuss any remaining points during the rest of the rebuttal period.

References

  1. Shen, Max W., et al. "How I learned to stop worrying and love StopGrads: Stationarity, Convergence, and a case study on Flow Map Learning." arXiv preprint arXiv:2609.16222 (2026).
  2. He, Ping, et al. "Sinkhorn-drifting generative models." arXiv preprint arXiv:2603.12366 (2026).
  3. Lee, HakGeun, and Hyonho Chun. "Identifiability and Stability of Generative Drifting with Companion-Elliptic Kernel Families." arXiv preprint arXiv:2604.24196 (2026).
  4. Turan, Erkan, Nicolas Dufour, and Maks Ovsjanikov. "Generative drifting is secretly score matching: a spectral and variational perspective." arXiv preprint arXiv:2603.09936 (2026).
  5. Esteban-Casadevall, Maria, et al. "Kernel-gradient drifting models." arXiv preprint arXiv:2605.10727 (2026).
  6. Davis, Oscar, et al. "Generalised flow maps for few-step generative modelling on riemannian manifolds." ICLR. 2026.
  7. Cheng, Chaoran, et al. "Riemannian consistency model." NeurIPS 2026.
  8. Woo, Dongyeop, et al. "Riemannian MeanFlow." ICML 2026.
Reviewer_yNHNOpenReview ↗

Identifiability of the cubic cost

Thank you for the clarifications and the additional experiment. Most of my concerns were addressed.

However I still have a remaining concern about Appendix C.2, which states that c(x,y)=∥x−y∥3/3c(x,y)=\|x-y\|^3/3 on S1S^1 is non-identifiable for every ε>0\varepsilon>0. This appears inconsistent with Theorem 3.2 and Lemma 3.3: the cubic cost induces an identifiable field for all but at most countably many positive values of ε\varepsilon.

One way to see it, is to note that each Fourier coefficient of the Gibbs kernel is analytic in β=1/ε\beta=1/\varepsilon and its zeros are therefore isolated, and the union of these exceptional parameters over all Fourier modes is at most countable (see the same argument used in Appendix B.4). Note that Lemma 3.3 seems to explicitly permit negative, nonzero eigenvalues (so even though this "cubic Gibbs kernel" fails to be positive definite for every ε\varepsilon, it can still be identifiable).

Could you specify the ε\varepsilon used in the checkerboard experiment and identify a vanishing Fourier coefficient establishing non-identifiability at that value? Otherwise the experiment demonstrates different performance between the two costs, but does not establish that the poorer performance is due to non-identifiability, but rather due to a possibly different, unrelated reason (e.g. maybe positive definiteness of the Gibbs kernel?).

Reviewer_Yot3OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

RW-Flow extends W-Flow to one-step generation on compact Riemannian manifolds. During training, generated samples are transported along a velocity field derived from a Sinkhorn energy, and the generator is regressed toward these transported targets using the exponential map and squared geodesic distance.

The main theoretical contribution is a necessary-and-sufficient condition for global field identifiability: a globally vanishing velocity field identifies the target distribution if and only if the associated Gibbs kernel is nondegenerate. The paper uses this characterization to study spectral, chordal, and geodesic costs, including exceptional regularization values for squared geodesic costs.

Experiments cover geospatial distributions on S2S^2, protein backbone angles on T2T^2, RNA angles on T7T^7, and distributions on triangular meshes. The results show competitive one-step generation and substantial sampling speedups relative to the reported multi-step mesh baseline.

Critical strengths

Strengths

In general, the paper is presented and formatted well and is quite pleasant to read.

  1. The main contribution goes beyond replacing Euclidean operations with their manifold counterparts. Theorem 3.2 and the accompanying cost analysis are the strongest novelty relative to W-Flow. The necessary-and-sufficient characterization, including its necessity argument, provides a useful explanation of which kernels preserve global identifiability. The squared-geodesic analysis also identifies a limitation that would be easy to miss when transferring the Euclidean method.
  2. The empirical evidence includes improvements on independent sample-quality metrics. RW-Flow-G improves 1-NNA over GFM-L from 0.643±0.0090.643\pm0.009 to 0.583±0.0080.583\pm0.008 on Fire, and from 0.572±0.0170.572\pm0.017 to 0.508±0.0110.508\pm0.011 on RNA, where values closer to 0.50.5 are better (Tables 9–10). Table 11 also reports roughly 228228–375×375\times faster sampling than the 1,000-step RFM configuration. This comes with some loss of quality: RFM has better 1-NNA in all four mesh settings, with the largest gap on Spot (k=10k=10), where the values are 0.5220.522 for RFM and 0.6010.601 for RW-Flow-H. The one-step results still offer a useful speed–quality tradeoff.
Critical weaknesses

Weaknesses

Because I had limited time to study the mathematics in detail, my understanding is limited, so the following weaknesses may reflect confusion after reading the text.

  1. The practical interpretation of the identifiability theorem needs qualification.

Theorem 3.2 characterizes when a field that vanishes throughout the manifold identifies the target. Section 3.2 then treats nondegeneracy as making the construction suitable for learning. However, nondegeneracy alone does not ensure that the target minimizes the associated Sinkhorn energy as the standard positivity theorem assumes a positive-definite universal Gibbs kernel.

For example, on the unit circle S1S^1, take

q=\tfrac12(\delta_{\pi/2}+\delta_{3\pi/2}),\qquad c(x,y)=\tfrac12 d_g(x,y)^2.$$ Solving the two-point transport problems gives $$\frac{\pi^2}{8} + \varepsilon\log\left(\frac{1+e^{-\pi^2/(2\varepsilon)}}{2}\right).$$ At $\varepsilon=3$, this equals approximately $-0.3163$, whereas $S_{\varepsilon,c}(p,p)=0$. Negativity persists on an interval, and Proposition 3.3.2 implies that the kernel is nondegenerate for almost every $\varepsilon$ in that interval. Thus, nondegeneracy can hold even though the target does not minimize the energy. This concerns the squared-geodesic cost used by RW-Flow-SG. It does not contradict Theorem 3.2 or establish a failure of the reported experiments, and it does not apply to the exact positive-definite cost constructions. Please distinguish the conditions for global field identifiability from the stronger conditions that justify the energy as a discrepancy to minimize. The actual regularization values used for SG would help clarify the practical relevance. 2. Important settings are missing from the matched-compute comparison. Table 8 documents the common architecture and optimizer, but it gives no numerical wall-clock budgets. I also could not find the regularization $\varepsilon$, transport step $\eta$, Sinkhorn iterations or tolerance, spectral parameters, etc. These choices matter both for reproducing the comparisons and for understanding how the theoretical cost constructions are used. Please provide these settings and the completed training steps for the main comparisons. For spectral costs, explain how the truncation threshold is chosen, as finite truncation removes exact identifiability.
Questions for authors
  1. How should the recommendation in Section 3.2 be qualified in light of the negative-energy example? In particular, which conclusions require positive definiteness or other assumptions beyond nondegeneracy, and which regularization values were used for RW-Flow-SG?
  2. What assumptions connect zero loss in Eq. 5 to the globally vanishing field?
  3. Table 8 specifies k(x,y)=exp⁡[−dg(x,y)2]k(x,y)=\exp[-d_g(x,y)^2] for the primary MMD statistic. This kernel is not positive definite on the evaluated spheres and flat tori, although its negative spectral components can be very small at this bandwidth. Was this the kernel used in the scorer, and can you provide a comparison using a positive-definite characteristic kernel at a comparable scale? Please also state whether negative V-statistics occurred and how they were handled. I view this as a check of robustness, not evidence that the current rankings are wrong.
  4. Have you evaluated RFM at fewer integration steps, for example 50 or 100? The reported acceleration is clear relative to the 1,000-step configuration, but a smaller-step comparison would help establish the speed advantage at comparable sample quality.
3 comments
AuthorsOpenReview ↗

Rebuttal Part 1/2

We thank the reviewer for their time and feedback. We have added the full hyperparameters, RFM results at fewer integration steps, and a discussion on the exact assumptions and implications of identifiability. We also present an evaluation with positive definite kernel and explain why the negative-energy example is not a concern. We answer each question below.

W2. Comprehensive Hyperparameters

In the table below, we report the requested hyperparameters. Overall, each torus run was allocated a budget of approximately 615 seconds, while each sphere run was allocated approximately 3,000 seconds. The number of Sinkhorn iterations ranged from 10 to 40, and the entropic regularization parameter ε ranged from 0.002 to 0.05. For mesh datasets, about 60 hours of compute was allocated per run.

MethodManifoldDatasetsεηSinkhorn itersSpectral paramsBatch (n_source)Training stepsWall-clock budget (s)
RW-Flow-SGS2S^2Earth (Volcano, Earthquake, Flood, Fire)0.002110–2048179,8123000
RW-Flow-CS2S^2Earth0.01110–2048177,0013000
RW-Flow-GS2S^2Earth0.01110–2048170,0773000
RW-Flow-HS2S^2Earth0.01110t=0.01t = 0.012048155,6583000
RW-Flow-MS2S^2Earth0.002110ν=1.5\nu=1.5, κ=0.5\kappa=0.5, σ2=(2νκ2)ν+d/2\sigma^2=(\frac{2\nu}{\kappa^2})^{\nu+d/2}2048157,4553000
RW-Flow-SS2S^2Earth0.05110α\alpha = 0.5, t=0.03t = 0.032048156,8133000
RW-Flow-SGT2T^2Top500 (General, Glycine, Pre-Pro, Proline)0.002120–204827,608615
RW-Flow-CT2T^2Top5000.01120–204827,411615
RW-Flow-GT2T^2Top5000.01120–204826,726615
RW-Flow-HT2T^2Top5000.002120t=0.001t = 0.001204827,229615
RW-Flow-SGT7T^7RNA0.002140–204814,142615
RW-Flow-CT7T^7RNA0.01140–204813,525615
RW-Flow-GT7T^7RNA0.01140–204813,739615
RW-Flow-HT7T^7RNA0.002140t=0.001t = 0.001204813,434615
RW-Flow-HMeshBunny (k=10)0.011100t=0.005t = 0.0051024642,857–
RW-Flow-HMeshBunny (k=50)0.011100t=0.005t = 0.00510241,082,926–
RW-Flow-HMeshSpot (k=10)0.021100t=0.01t = 0.011024613,402–
RW-Flow-HMeshSpot (k=50)0.011100t=0.005t = 0.0051024476,578–

Indeed, truncation means exact identifiability no longer holds. To keep this effect negligible, we choose the truncation level so that the truncation error is below 10−1510^{-15}. All these details are added to the App. F.

W1, Q1.Negative-energy example and positive definiteness

It is true that the Sinkhorn divergence can be negative. However, the training dynamics are driven by the Wasserstein gradient of the Sinkhorn divergence and not the value of the Sinkhorn divergence itself. Thus, the training dynamics only require that the gradient vanishes, which does not require the minimization of the energy, as in the example given by the reviewer.

Accordingly, our main theorem (Thm. 3.2) does not assume positive definiteness. More generally, we use positive definiteness only to establish nondegeneracy (Prop. 3.3.1), which we then apply to the costs in Section 3.3.2.

For the regularization values used in the experiments, check the table above.

Q2. What assumptions connect zero loss in Eq. 5 to the globally vanishing field?

The key assumption linking zero loss in Eq. 5 to a globally vanishing field is that qθq_\theta has full support on the manifold. The loss is zero exactly when VV vanishes qθq_\theta-almost everywhere. A similar assumption is made in the Appendix C of Deng et al. [2026]. Other works on drifting models have similar assumptions. (Deng et al. [2026], He et al. [2026], Lee and Chun [2026], Turan et al. [2026], Esteban-Casadevall et al. [2026], Lai et al. [2026], and Gretton et al. [2026]) App. C now discusses these assumptions clearly (also see the general response).

AuthorsOpenReview ↗

Rebuttal Part 2/2

Q3. Squared Geodesic Gaussian is not positive definite

It is indeed true that, in general, squared geodesic kernel is not positive definite. We used it to quantify our results mainly because most of the relevant prior works do so (Davis et al. [2026], Woo et al. [2026], Esteban-Casadevall et al. [2026], Cheng et al. [2025]). Nevertheless, the table below reports MMD results on the Earth Flood dataset using the heat kernel, which is positive definite. We note that the overall rankings were not affected much.

MethodHeat kMMD ↓Gauss kMMD ↓
RW-Flow-S0.05535 ± 0.002040.03816 ± 0.01018
RW-Flow-G0.05565 ± 0.006540.04047 ± 0.01019
RW-Flow-C0.05799 ± 0.004090.04527 ± 0.00646
MMDF-L0.05822 ± 0.002030.03821 ± 0.00663
RW-Flow-SG0.05892 ± 0.011470.04789 ± 0.01436
RW-Flow-H0.05898 ± 0.004770.04552 ± 0.01512
RW-Flow-M0.06361 ± 0.005360.04559 ± 0.00256
GFM-MF0.06771 ± 0.008160.06663 ± 0.02300
MMDF-G0.06793 ± 0.007200.04164 ± 0.01071
MMDF-H0.06865 ± 0.003100.04082 ± 0.00852
RCM-CT0.07362 ± 0.003050.05185 ± 0.01874
RMF-Ev0.07389 ± 0.006780.04734 ± 0.01253
GFM-L0.07423 ± 0.027350.06581 ± 0.01752
RMF-Sv0.07556 ± 0.001940.04431 ± 0.00938
KGF-M0.07568 ± 0.009470.05746 ± 0.02376
RMF-Lv0.07858 ± 0.019180.05365 ± 0.01972
GFM-S0.09202 ± 0.022150.05189 ± 0.02397
KGD-GL0.09813 ± 0.055200.11352 ± 0.09515
KGD-GG0.10905 ± 0.069150.14239 ± 0.20187
RCM-sCT0.13548 ± 0.007810.11573 ± 0.02086
RFM0.15646 ± 0.006470.24647 ± 0.01872
GFM-E0.16922 ± 0.006220.17819 ± 0.01220

In the experiments, negative V-statistics did not occur.

Q4. RFM at fewer integration steps on meshes

We initially used 1000 steps following the official implementation and settings of RFM. Upon your request, we tested RFM on fewer sampling step settings. In the table below, we show the effects of different NFEs on quality and sampling time. We found that RFM perform well with 50 step setting performing comparable to 1000 step setting. However, our method still achieves about 20x speedup. We have revised our paper and now include this analysis.

DatasetMethodNFEkMMD ↓MMD ↓COV ↑1-NNA (→ 0.5)Sampling time (s) ↓
Bunny (k=10)Held-out (val vs. test)–0.02010.01400.58250.4928–
RFM100.04120.02080.47800.62680.460
RFM500.04320.01560.56450.53202.298
RFM1000.04510.01480.57000.52104.598
RFM10000.04830.01460.56850.513849.540
RW-Flow-H10.03040.01540.55850.53670.158
Bunny (k=50)Held-out (val vs. test)–0.02500.01450.58250.4980–
RFM100.05590.02850.48200.61650.460
RFM500.03420.01570.55900.52602.299
RFM1000.03000.01510.56650.50654.597
RFM10000.02780.01550.56500.516048.665
RW-Flow-H10.03400.01650.54450.54980.128
Spot (k=10)Held-out (val vs. test)–0.01180.01590.58700.5002–
RFM100.12590.03260.38050.70000.539
RFM500.04540.01760.54300.54172.697
RFM1000.02850.01680.55550.52825.394
RFM10000.01630.01660.56100.522054.825
RW-Flow-H10.01810.02130.49350.60080.240
Spot (k=50)Held-out (val vs. test)–0.01440.01570.58400.5152–
RFM100.08010.02820.41100.68450.540
RFM500.04680.01880.53650.56782.687
RFM1000.03130.01740.54250.54885.398
RFM10000.02340.01740.54050.549054.692
RW-Flow-H10.01470.01890.54900.55920.243

Conclusion

We appreciate the reviewer's time and effort. We hope our responses and revisions address the reviewer's questions. We are happy to clarify anything further.

AuthorsOpenReview ↗

Rebuttal References

References

  1. Deng, Mingyang, et al. "Generative modeling via drifting." arXiv preprint arXiv:2602.04770 (2026).
  2. He, Ping, et al. "Sinkhorn-drifting generative models." arXiv preprint arXiv:2603.12366 (2026).
  3. Lee, HakGeun, and Hyonho Chun. "Identifiability and Stability of Generative Drifting with Companion-Elliptic Kernel Families." arXiv preprint arXiv:2604.24196 (2026).
  4. Turan, Erkan, Nicolas Dufour, and Maks Ovsjanikov. "Generative drifting is secretly score matching: a spectral and variational perspective." arXiv preprint arXiv:2603.09936 (2026).
  5. Esteban-Casadevall, Maria, et al. "Kernel-gradient drifting models." arXiv preprint arXiv:2605.10727 (2026).
  6. Lai, Chieh-Hsin, et al. "A Unified View of Score-Based and Drifting Models." arXiv preprint arXiv:2603.07514 (2026).
  7. Gretton, Arthur, et al. "On the wasserstein gradient flow interpretation of drifting models." arXiv preprint arXiv:2605.05118 (2026).
  8. Franz, Leonard T., et al. "Drifting fields are not conservative." arXiv preprint arXiv:2604.06333 (2026).
  9. Davis, Oscar, et al. "Generalised flow maps for few-step generative modelling on riemannian manifolds." ICLR. 2026.
  10. Cheng, Chaoran, et al. "Riemannian consistency model." NeurIPS 2026.
  11. Woo, Dongyeop, et al. "Riemannian MeanFlow." ICML 2026.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

General Response

We thank all reviewers for their constructive feedback. We are glad that reviewers found the paper well presented (Yot3) and the question important (tKDM), and that they saw the necessary-and-sufficient identifiability result (Thm. 3.2) as the main contribution, going beyond the sufficiency-only results of prior work (Yot3, yNHN, tKDM). Reviewers also noted the strong RNA results (yNHN) and the gains on independent sample-quality metrics (Yot3). Below, we summarize the main revisions. Detailed answers are in the individual responses. All changes in the revised PDF are highlighted in blue.

Summary of revisions

  • Experiment on Identifiable vs. non-identifiable costs (yNHN). New experiment on S1S^1 with a one-dimensional checkerboard target, comparing the geodesic cost (identifiable) with the cubed chordal cost 13∥x−y∥3\frac13\|x-y\|^3 (non-identifiable). (App. C.2)
  • Identifiability and training dynamics (Yot3, yNHN). New discussion of stop-gradient stationarity under a finite parametrization, and of the full-support assumption linking zero loss to a globally vanishing field. (App. C.1)
  • Related work (yNHN). Expanded discussion of recent drifting models and their identifiability results, including Sinkhorn drifting, companion-elliptic kernels and drifting as score matching. (Sec. 4)
  • Statistical tie. We now explicitly mention the statistical tie between our method and GFM-Lagrange in terms of kMMD. (Sec. 5)
  • Hyperparameters (Yot3). Full per-run settings: entropic regularization ε\varepsilon, transport step η\eta, Sinkhorn iterations, spectral truncation, training steps and compute budgets. (App. F)
  • RFM with fewer steps (Yot3). RFM at fewer integration steps on meshes. The speedup claim is revised to about 20× against RFM at comparable quality. (Sec. 5, App. F)
  • Training and sampling costs (tKDM). Per-update training time, GPU memory and sampling latency vs. GFM and RMF baselines, plus the one-time mesh preprocessing cost. (App. F)

On the role of identifiability

Reviewer yNHN asked why identifiability is the relevant population-level property for our method, and Reviewer Yot3 asked which assumptions connect zero loss in Eq. 5 to a globally vanishing field. We added a new section (App. C) that addresses both questions. In brief:

The chain of implications. Consider the statements

  1. qθ=pq_\theta = p,
  2. Vqθ,p≡0V_{q_\theta,p} \equiv 0 on M\mathcal{M},
  3. L=0\mathcal{L} = 0, i.e., Vqθ,p≡0V_{q_\theta,p} \equiv 0 qθq_\theta-a.e.,
  4. ∇θL=0\nabla_\theta \mathcal{L} = 0, i.e., θ\theta is stationary.

For any well-behaved drift (e.g., the Sinkhorn drift), (1) ⇒ (2) ⇒ (3) ⇒ (4). Training only guarantees (4), but we want (1), so each implication has to be reversed:

  • (4) ⇒ (3): no spurious points from the parametrization (yNHN, Q1). This fails for general networks fθf_\theta. However, for an unrestricted ff, every stationary point of the stop-gradient loss is a global minimizer with zero loss (Prop. C.1, following Shen et al., 2026).
  • (3) ⇒ (2): full support (Yot3, Q2). This holds whenever qθq_\theta has full support on M\mathcal{M} and VV is continuous, e.g., after smoothing with the heat kernel for an arbitrarily short time.
  • (2) ⇒ (1): identifiability. This is exactly the identifiability property.

Takeaway. Identifiability is a necessary link in going from a stationary point to the target distribution. It also does not depend on the network parametrization, so it rules out spurious solutions caused by a poor choice of cost or drift. Under the full-support condition, the only remaining source of failure is limited network capacity.

Empirical evidence. On a checkerboard distribution on S1\mathbb{S}^1, we compare the geodesic cost (identifiable) with the cubed chordal cost (not identifiable for any ε\varepsilon). The geodesic cost recovers the target, with 91% of the learned mass on the support. The chordal cost collapses the mass into narrow spikes, with only 2% on the support. [App. C].

We hope these revisions address the reviewers' concerns, and we welcome further discussion.

AuthorsOpenReview ↗

General Response 2

Summary of the second revision

We thank the reviewers and the AC for the continued discussion. We especially thank Reviewer yNHN for finding an error in our first-round response. Below we summarize what changed in the second revision. All changes are highlighted in the revised PDF.

1. Correction: the cubic chordal cost is identifiable at the ε\varepsilon we used. In the first revision and in our first General Response, we described the cubic chordal cost 13∥x−y∥3\frac13\|x-y\|^3 as non-identifiable for every ε>0\varepsilon > 0. This is incorrect: at ε=0.5\varepsilon = 0.5, the value used in our experiment, the induced vector field is identifiable. The empirical results stand (91% vs. 2% of the learned mass on the target support), but our explanation of them was wrong. We have corrected it as described below.

2. New explanation: non-positive-definite kernels and negative Sinkhorn divergence (App. C.3). The Gibbs kernel of the cubic chordal cost is not positive definite for any ε\varepsilon. As Reviewer Yot3 pointed out, without positive definiteness the Sinkhorn divergence can take negative values. The Wasserstein gradient flow can only decrease the divergence. So a flow that starts with Sε(q0,p)<0S_\varepsilon(q_0, p) < 0 can never reach pp, where Sε(p,p)=0S_\varepsilon(p, p) = 0. App. C.3 now presents this argument together with an extended experiment on the checkerboard target on S1\mathbb{S}^1. The configurations that start from a negative divergence are exactly the ones that completely collapse.

3. New example of a non-identifiable cost (App. C.2). We give a cost that is non-identifiable for every ε>0\varepsilon > 0: the axial cost c(x,y)=1−⟨x,y⟩2c(x, y) = 1 - \langle x, y\rangle^2 on S1\mathbb{S}^1. Following the proof of Thm. 3.2, we construct a simple pair q≠pq \neq p with Vq,p≡0V_{q,p} \equiv 0. The flow is therefore stationary at the wrong distribution. We validate the stationarity empirically.

4. Positive definiteness can matter (Sec. 3.2). In Sec. 3.2, we added a remark explicitly stating positive definiteness can matter for optimization especially when the initial Sinkhorn divergence between the source and target distributions is negative and point the readers to App. C.3 for in-depth discussion.

5. A remark saying most of the costs used in the paper are positive definite (Sec. 3.3) We added a remark stating most of the costs have positive definite kernels with the exception of the squared geodesic cost on Sd\mathbb{S}^d and Td\mathbb{T}^d and the geodesic cost on Td(d≥2)\mathbb{T}^d (d\ge 2).

6. Revised speedup claim (Sec. 1, Sec. 5). We now report a speedup of up to 18× over RFM at comparable sample quality, replacing the "about 20×" stated in the first revision.

We are grateful for all the reviewers' careful reading, which has made the theoretical claims of the paper more precise. We welcome any further discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.