acceptodds
Under review as a conference paper at ICLR 2027

One-Step Generation via Riemannian Wasserstein Gradient Flows

13 views

Abstract

Recently, Drifting Models and Wasserstein Gradient Flows have attracted substantial attention because they move iterative distributional refinement to training and amortize it into a generator, enabling fast inference. However, existing formulations have been developed largely for continuous Euclidean domains, such as image spaces, where particles admit unconstrained additive updates. On constrained spaces, these updates can leave the valid domain or ignore its geometry, making them unsuitable targets for training. Recent work has adapted updates to these spaces, but has focused on particular fields or offered limited empirical comparison. We derive and compare several geometry-aware fields within a common training framework for one-step generators. We test the method on data with different structures and obtain competitive one-step results in each setting. The best-performing field varies by task, showing why the choice of objective matters in practice.

Primary area:
generative models
Keywords:
diffusion language models, categorical generation, Wasserstein gradient flows, Riemannian geometry

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_7gzgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Wasserstein Gradient Flows (WGF) into Drifting Models, extending them to Riemannian manifolds. The authors first formulate WGF on Riemannian manifolds. They then consider three different distributional objectives and derive the corresponding drift fields. Experiments are conducted on the Earth, Binary-MNIST, CIFAR-10, and LM1B datasets, achieving competitive results.

Critical strengths

The paper extends Drifting Models to Riemannian manifolds and derives the drift fields from explicit objectives rather than starting from kernel functions. The method can also work on general manifolds where geodesics are not available in closed form. It achieves competitive results on image and language generation tasks.

Critical weaknesses

The authors mention that they use relatively small datasets due to limited computational resources, which is understandable. However, the experiments cover too few types of manifolds. Based on existing Riemannian generative models, I suggest adding experiments on the torus, SO(3), and SE(3)^n. These datasets are also relatively small, so running additional experiments should be feasible.

The paper proposes three different drift fields. On the Earth benchmark, no single method consistently performs the best across all datasets. However, R-WGF-Sinkhorn is selected as the best-performing method for the remaining tasks. Is there any theoretical explanation or empirical evidence that can help determine which drift field will perform best for a given task before training?

Questions for authors

see weakness

2 comments
AuthorsOpenReview ↗

Answer to Reviewer 7gzg

Dear Reviewer 7gzg, thank you for your insightful comments and for taking the time to review our work. We appreciate your high confidence and constructive feedback. Please find below the answers to your questions and comments.

[W.1] […] the experiments cover too few types of manifolds. […] I suggest adding experiments on the torus, SO(3)\mathrm{SO}(3), and SE(3)n\mathrm{SE}(3)^n.

Thank you for suggesting these benchmarks. In the revised manuscript, we derived the R-WGF updates for flat tori, SO(3)\mathrm{SO}(3), and SE(3)n\mathrm{SE}(3)^n, with drift formulas in Table 1 and derivation details in Appendix A. We evaluated our one-step generation pipeline on experimental setups adapted from Riemannian MeanFlow (RMF) [1], which Reviewer iffN also suggested as a comparison.

We tested Proline and RNA torsion angles on T2\mathbb{T}^2 and T7\mathbb{T}^7, respectively, Fisher and Swiss Roll rotations on SO(3)\mathrm{SO}(3), and ACRONYM grasps on SE(3)\mathrm{SE}(3). We also added a synthetic benchmark for joint generation of four correlated rigid poses on SE(3)4\mathrm{SE}(3)^4. The table below shows one-step results.

MethodMMD (↓\downarrow) — ProlineMMD (↓\downarrow) — RNAMMD (↓\downarrow) — FisherMMD (↓\downarrow) — Swiss RollMMD$^2$ (×104\times10^4, ↓\downarrow) — SE(3)4\mathrm{SE}(3)^4Grasp success (%, ↑\uparrow) — ACRONYM
RFM [2]0.520.520.680.680.150.150.350.35—3.23.2
G-LSD [3]0.040.040.080.080.070.070.030.03—6060
RMF [1]0.090.090.200.200.080.080.080.084.3±0.114.3\pm0.1170.30±5.2570.30\pm5.25
RMF-MT [1]0.040.040.070.070.040.040.060.06—6060
R-WGF-MMD (Ours)0.04±0.010.04\pm0.010.02±0.010.02\pm0.010.07±0.010.07\pm0.010.01±0.01\mathbf{0.01\pm0.01}5.2±0.0955.2\pm0.09581.60±3.7081.60\pm3.70
R-WGF-Sinkhorn (Ours)0.03±0.01\mathbf{0.03\pm0.01}0.02±0.01\mathbf{0.02\pm0.01}0.03±0.01\mathbf{0.03\pm0.01}0.01±0.010.01\pm0.013.7±0.18\mathbf{3.7\pm0.18}84.10±2.63\mathbf{84.10\pm2.63}
R-WGF-KL (Ours)0.04±0.010.04\pm0.010.02±0.010.02\pm0.010.07±0.010.07\pm0.010.01±0.010.01\pm0.016.0±0.216.0\pm0.2182.30±1.9982.30\pm1.99

R-WGF results are mean ±\pm SD over five training seeds. The RMF results on SE(3)4\mathrm{SE}(3)^4 and ACRONYM use our adapted implementation. The other baseline values are from [1].

At one model evaluation, the best R-WGF variant achieves lower reported MMD than the published baselines on all four torus and rotation benchmarks. R-WGF-Sinkhorn leads on Proline, RNA, and Fisher, and R-WGF-MMD leads on Swiss Roll. On SE(3)4\mathrm{SE}(3)^4, R-WGF-Sinkhorn lowers scaled joint MMD$^2$ from 4.3 to 3.7 relative to the adapted RMF control. On ACRONYM, we reran RMF using the authors' released code and evaluated both methods on the same test split with the same evaluator. R-WGF-Sinkhorn achieves 84.10±2.63%84.10\pm2.63\% Success@20, compared with 70.30±5.25%70.30\pm5.25\% for our RMF rerun. For reference, the RMF paper [1] reports 65%65\% grasp success. Section 5.4 of the revised manuscript summarizes the new experiments, and Appendix B provides the evaluation protocols and full results. Also, Figure 6 shows examples of one-step generation.

[W.2] Is there any theoretical explanation or empirical evidence that can help determine which drift field will perform best […]?

Under the stated assumptions, all three objectives are minimized at the target distribution, and each exact population flow decreases its own objective. In practice, we estimate the drifts from finite samples and take discrete steps, so the relative performance depends on the task and training setup. Sinkhorn uses a transport cost to match generated and data samples, which is useful when the cost reflects meaningful differences between samples. MMD uses kernel averages, so it is easy to estimate from samples and depends on the kernel scale. KL moves samples according to the difference between the data and model scores, which must be estimated reliably.

Our experiments show that the best choice varies across tasks. The best drift varies across Earth datasets even on the same sphere, and KL achieves the lowest FID on Binary-MNIST. Sinkhorn was selected for full training after performing best in the CIFAR-10 and LM1B pilots. In Table 5, Sinkhorn leads on five of six benchmarks and MMD on Swiss Roll. For a new task, we suggest comparing the fields in short pilot runs with the same architecture and budget, then choosing based on a prespecified validation metric.

[1] Riemannian MeanFlow (RMF, https://arxiv.org/abs/2603.10718)

[2] RFM (https://arxiv.org/abs/2302.03660)

[3] G-LSD (https://arxiv.org/abs/2510.21608)

Concluding remarks. Thank you for your feedback. We derived the R-WGF updates for tori, rotations, and rigid poses and benchmarked our one-step generators against RMF on these geometries. We would be grateful if you could let us know whether these revisions address your concerns. If so, we kindly ask that you consider increasing your rating. We are also open to discussing any other questions you may have.

AuthorsOpenReview ↗

Kind Reminder

Dear Reviewer 7gzg,

Thank you again for your careful reading of our paper and for your thoughtful comments. We have responded in detail to each of the points you raised and hope that our rebuttal and revisions address your concerns. We would be grateful for your feedback on our response and we would welcome the opportunity to clarify any remaining points. If you feel that your concerns have been resolved, we hope you might consider a higher overall score.

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

The paper proposes a novel method for one-step constrained generation by making use of constraint geometry. The work extends the highly recent development of drift models to the case where data is strictly supported on a Riemannian manifold. The authors rigorously derive a one-step generative model that appears to be intrinsic to the Riemannian manifold geometry. Experiments cover a broad range of tasks.

Critical strengths
  1. The paper appears highly mathematically rigorous and makes use of known results in manifold geometry, which is increasingly relevant in efficient generative modeling.
  2. The benchmarks are extensive and the authors evaluate on a diverse suite of experimental tasks, suggesting that this paper will be useful for many fields.
Critical weaknesses
  1. The conclusions are extremely short, and limitations are outsourced to the appendix. It is unclear how the work leaves opportunities open for future work or precisely how this would be used by the scientific community.
  2. It is unclear why language modeling has connections to Riemannian manifold geometry. It is slightly more intuitive why Binary-MNIST has geometric structure, but the paper would benefit from more introductory exposition on the Riemannian structure of the given tasks. The sphere tasks is the most natural, but a clear explanation of the manifold in question (is it learned? analytically known in closed form?) would be highly useful.
Questions for authors
  1. Are all FLM parameters finetuned from the pretrained base model, or just a subset?
  2. What exactly is the manifold for Binary-MNIST and the language modeling task?
  3. The related work is further towards the end of the paper. Is there any reason why it couldn't go closer to the background?
  4. Are drifting models (either the original work or the proposed method in this paper) intrinsically a one-step technique, or are they trained to be one-step? The main question is whether there is any approximation error due to the one-step sampling that should be acknowledged.
  5. How is this an improvement upon Riemannian mean flow generation (https://arxiv.org/abs/2603.10718), or is it a different task entirely?
4 comments
AuthorsOpenReview ↗

Answer to Reviewer iffN

Dear Reviewer iffN, thank you for your comments. Please find below the answers to your questions and comments.

[W.1] The conclusions are extremely short, and limitations are outsourced to the appendix. […] opportunities open for future work […] used by the scientific community.

Thank you for raising these points. In Section 6 of the revised manuscript, we expanded the conclusion, moved the limitations into the main text, and added a discussion of open directions, including scaling and conditional generation.

In short, R-WGF extends the distributional refinement strategy of Drifting and W-Flow to Riemannian manifolds and generalize other methods. Our experiments show that it achieves competitive sample quality with a single forward pass of the generator across several applications, including image and text generation on products of simplices, molecular torsion-angle sampling on tori, and generation of rotations on SO(3)\mathrm{SO}(3) and rigid-body poses on SE(3)\mathrm{SE}(3). Therefore, for scientific community working with manifold-valued data, this enables fast sampling with a reasonably good quality of generation.

[W.2 / Q2] It is unclear why language modeling has connections to Riemannian manifold geometry. […] the manifold in question (is it learned? analytically known in closed form?) […] What exactly is the manifold for Binary-MNIST and the language modeling task?

We specify the manifold and metric analytically based on the structure of each dataset and keep both fixed throughout training, following standard approaches [1,2,3,4]. We represent Earth locations on the round sphere S2\mathbb{S}^2. For images, the generator predicts a categorical distribution over intensity levels for each pixel channel, with two levels for Binary-MNIST and 256 levels (0–255) for CIFAR-10. For text, it predicts a categorical distribution over the vocabulary at each token position. Each probability vector lies on a simplex, so the joint output space is a product of simplices. These spaces are (Δ1)28×28(\Delta^1)^{28\times28} for Binary-MNIST, (Δ255)32×32×3(\Delta^{255})^{32\times32\times3} for CIFAR-10, and (Δ29524)128(\Delta^{29524})^{128} for LM1B. We equip the interior of each such space with the product Fisher–Rao metric.

The revised manuscript now specifies the manifold for each dataset at the start of the corresponding experimental subsection (Sections 5.1–5.4), which is highlighted with blue color.

[Q1] Are all FLM parameters finetuned from the pretrained base model, or just a subset?

We fine-tune all parameters of the FLM generator. FMLM uses the same pretrained FLM checkpoint and also fine-tunes the full generator, so it is a fair one-step baseline with respect to initialization and trainable parameters.

Moreover, we refined the R-WGF-Sinkhorn training setup, new results in the following table.

LM1B, 1 NFEGenPPL ↓\downarrowEntropy ↑\uparrow
DFM (PSD)94.084.06
FMLM119.344.16
R-WGF-Sinkhorn (Ours)112.974.20

At 1 NFE on LM1B, the revised R-WGF-Sinkhorn outperforms FMLM on both metrics, lowering GenPPL (112.97 vs. 119.34). DFM (PSD) achieves the lowest GenPPL (94.08), but its lower entropy of 4.06 indicates a partial collapse to less diverse text, which also lowers GenPPL.

We have included these LM1B results in Section 5.3 (Table 4) of the revised manuscript. Appendix B.4 describes the fine-tuning protocol.

[Q3] The related work is further towards the end of the paper. Is there any reason why it couldn't go closer to the background?

Thank you for this suggestion. We have moved the related work section after the background section in the revised manuscript.

[Q4] Are drifting models […] intrinsically a one-step technique, or are they trained to be one-step? […] any approximation error due to the one-step sampling

Drifting models are one-step by construction. R-WGF generalizes this approach to Riemannian manifolds and retains a direct map from noise to data. Each sample requires a single forward pass, avoiding the numerical integration error associated with standard flow-matching and diffusion samplers during the inference.

AuthorsOpenReview ↗

Answer to Reviewer iffN

[Q5] How is this an improvement upon Riemannian mean flow generation […], or is it a different task entirely?

RMF and R-WGF are the different approaches to the same task of one-step generation on manifolds. RMF extends MeanFlow [5] by learning the average velocity over a time interval along a prescribed probability flow. This lets it advance a sample across the entire interval in a single model evaluation. R-WGF draws on Drifting and W-Flow to refine the generator’s output distribution during training. At each iteration, we take a Riemannian Wasserstein gradient step to move generated samples toward the data distribution and use the updated samples as training targets. The trained generator maps noise directly to samples in one forward pass.

Motivated by your question, we benchmarked R-WGF against RMF using experimental setups adapted from the original RMF study. The table below shows the one-step results.

MethodMMD (↓\downarrow) — ProlineMMD (↓\downarrow) — RNAMMD (↓\downarrow) — FisherMMD (↓\downarrow) — Swiss RollMMD$^2$ (×104\times10^4, ↓\downarrow) — SE(3)4\mathrm{SE}(3)^4Grasp success (%, ↑\uparrow) — ACRONYM
RMF0.090.090.200.200.080.080.080.084.3±0.114.3\pm0.1170.30±5.2570.30\pm5.25
R-WGF-MMD (Ours)0.04±0.010.04\pm0.010.02±0.010.02\pm0.010.07±0.010.07\pm0.010.01±0.01\mathbf{0.01\pm0.01}5.2±0.0955.2\pm0.09581.60±3.7081.60\pm3.70
R-WGF-Sinkhorn (Ours)0.03±0.01\mathbf{0.03\pm0.01}0.02±0.01\mathbf{0.02\pm0.01}0.03±0.01\mathbf{0.03\pm0.01}0.01±0.010.01\pm0.013.7±0.18\mathbf{3.7\pm0.18}84.10±2.63\mathbf{84.10\pm2.63}
R-WGF-KL (Ours)0.04±0.010.04\pm0.010.02±0.010.02\pm0.010.07±0.010.07\pm0.010.01±0.010.01\pm0.016.0±0.216.0\pm0.2182.30±1.9982.30\pm1.99

R-WGF-Sinkhorn achieves lower reported MMD than RMF on all four torus and rotation benchmarks and lowers joint MMD$^2$ on correlated rigid poses from 4.3 to 3.7 in the units of the table.

We have added these results to Section 5.4 and Table 5 of the revised manuscript. Appendix B provides the experimental details.

[1] PairFlow (https://arxiv.org/abs/2512.20063)

[2] Categorical Flow Maps (https://arxiv.org/abs/2602.12233)

[3] Flow Map Language Models (https://arxiv.org/abs/2602.16813)

[4] Kernel-Gradient Drifting Models (https://arxiv.org/abs/2605.10727)

[5] MeanFlow (https://arxiv.org/abs/2505.13447)

Concluding remarks. Thank you for your feedback. We clarified the method, expanded the conclusion, and added updated LM1B results and an RMF comparison. We would be grateful if you could let us know whether these revisions address your concerns. If so, we kindly ask that you consider increasing your rating. We are also open to discussing any other questions you may have.

AuthorsOpenReview ↗

Kind Reminder

Dear Reviewer iffN,

Thank you again for your careful reading of our paper and for your thoughtful comments. We have responded in detail to each of the points you raised and hope that our rebuttal and revisions address your concerns. We would be grateful for your feedback on our response and we would welcome the opportunity to clarify any remaining points. If you feel that your concerns have been resolved, we hope you might consider a higher overall score.

I thank the authors for their clear and detailed responses to my questions and for the substantial additional experiments! I have correspondingly raised my score.

Reviewer_6Eb7OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes R-WGF, a framework for training one-step generators on Riemannian spaces. It derives Sinkhorn, MMD, and KL update fields and constructs geometric training targets. Experiments are done on spherical geospatial data, Binary-MNIST, discretized CIFAR-10, and LM1B.

Critical strengths
  1. The paper's spherical and Fisher–Rao implementations are innovative. The paper also cleanly distinguishes its population guarantees from the approximations used in training.

  2. The method achieves good results under different data types. It improves Binary-MNIST FID from CatDrift’s 8.0 to 6.5 and discretized CIFAR-10 FID from 58 to 46.9.

Critical weaknesses
  1. The benefit of the geometric constraints for generator optimization needs clarification. Suppose we're in the ideal smooth setting, then an exponential-map target followed by squared geodesic regression will give a gradient at the current generator proportional to direct differentiation of the distributional objective. In that case, the metric in the drift cancels that in the regression loss. Therefore, a matched direct Sinkhorn/MMD optimization baseline is therefore important to establish what the construction adds.

  2. The image experiments lack a direct comparison with W-Flow under matched settings. Table 3 in the paper includes some drifting results from prior work, but these do not show how is it benefited from the proposed geometry. The motivation for representing each pixel or channel as a categorical probability distribution also needs further justification. Ordinary W-Flow can generate images directly, without introducing a probability representation. The motivation of the selected experiments needs more discussion.

Questions for authors

Could the authors address the weaknesses above, particularly what the geometric construction adds, and compare against ordinary W-Flow using both categorical and continuous image representations?

5 comments
AuthorsOpenReview ↗

Answer to Reviewer 6Eb7

Dear Reviewer 6Eb7, thank you for your insightful comments and for taking the time to review our work. We appreciate your high confidence and constructive feedback. Please find below the answers to your questions and comments.

[W.1] […] the metric in the drift cancels that in the regression loss. […] a matched direct Sinkhorn/MMD optimization baseline is therefore important to establish what the construction adds.

Thank you for pointing out this equivalence. In Section 4.4, we derive its Riemannian counterpart. In brief, the identity follows by differentiating the squared geodesic loss with the target held fixed. Let x=fθ(z)x=f_\theta(z) and let ψ\psi be the first variation of Fp\mathcal F_p at qθq_\theta. For an exponential-map step, define the stop-gradient target

x~=sg[Expx(−ηgradgψ(x))].\widetilde{x}=\mathrm{sg}[\mathrm{Exp} _ {x}(-\eta \mathrm{grad} _ {\mathsf{g}}\psi(x))].

The metric cancels in the chain rule. With the target held fixed, the tracking loss and its gradient at the current iterate are

\qquad \nabla_\theta\mathcal L(\theta)=2\eta\nabla_\theta\mathcal F_p(q_\theta).$$ However, in practice, this equivalence can break down. We summarize these points below. **Geometry in the distributional objective.** The Sinkhorn cost and the MMD and Parzen KL kernels determine which samples the functional considers close. This motivates choosing costs and kernels that reflect the underlying manifold metric, since ambient Euclidean distances can misrepresent how close points are on the manifold. **Geometry in the implemented update.** We construct finite targets, then fit them with geodesic loss. These choices can change the gradient at a finite step size. We also normalize the drift following Drifting Models [1] and W-Flow [2]. In our manifold experiments, the Riemannian RMS norm provides a natural measure of drift magnitude. To assess how the two choices discussed above affect training in practice, we conducted an ablation on synthetic categorical sequences for each of the three drifts. The synthetic construction lets us compute the target pairwise probabilities exactly, providing a common reference for all variants. It also makes a full comparison across vocabulary sizes and training seeds feasible within our available compute budget. **Dataset.** We model the joint distribution of eight categorical tokens. To generate a sequence, we first draw one of 64 latent topics uniformly. The selected topic specifies a pair of candidate tokens at each position. Given the topic, we sample one token from each pair independently with probabilities $0.7$ and $0.3$. **Comparison of geometry in the distributional objective.** We compare squared Hellinger distance, which reflects the Fisher–Rao geometry of the simplex, with squared Euclidean distance. These distances define the transport cost for Sinkhorn and the Gaussian kernels for MMD and the Parzen-smoothed KL drift. **Comparison of geometry in the implemented update.** For each cost or kernel, we compare direct gradient updates with retraction-based tracking. The direct update optimizes the distributional functional directly with respect to the generator parameters. The tracking update uses the drift to construct a regression target on the simplex.
AuthorsOpenReview ↗

Answer to Reviewer 6Eb7

We compute the total variation (TV) distance between the generated and exact target joint token distributions. The table below reports results, where dd denotes the number of tokens in the vocabulary. At d=2d=2, all variants achieve similarly low mean pairwise TV. For every tested d≥16d\geq16, retraction-based tracking improves on the corresponding direct update for all three drifts under both distance choices. Tracking with the geometry-aware Hellinger cost or kernel yields the lowest mean pairwise TV for each drift at these vocabulary sizes.

FunctionalCost or kernelUpdateVocabulary size dd = 2Vocabulary size dd = 16Vocabulary size dd = 128Vocabulary size dd = 1,024Vocabulary size dd = 4,096
SinkhornHellingerRetraction0.014 ± 0.0040.064 ± 0.0010.147 ± 0.0030.161 ± 0.0030.158 ± 0.001
SinkhornHellingerDirect0.007 ± 0.0010.138 ± 0.0200.308 ± 0.0070.214 ± 0.0050.215 ± 0.004
SinkhornEuclideanRetraction0.017 ± 0.0040.072 ± 0.0030.164 ± 0.0070.164 ± 0.0120.161 ± 0.004
SinkhornEuclideanDirect0.007 ± 0.0010.126 ± 0.0030.259 ± 0.0030.231 ± 0.0080.252 ± 0.002
MMDHellingerRetraction0.011 ± 0.0090.100 ± 0.0020.344 ± 0.0090.383 ± 0.0240.478 ± 0.031
MMDHellingerDirect0.018 ± 0.0120.313 ± 0.0490.889 ± 0.0200.945 ± 0.0050.938 ± 0.007
MMDEuclideanRetraction0.008 ± 0.0010.403 ± 0.0080.932 ± 0.0040.945 ± 0.0200.937 ± 0.025
MMDEuclideanDirect0.014 ± 0.0080.663 ± 0.0190.961 ± 0.0020.983 ± 0.0030.983 ± 0.002
Parzen-smoothed KLHellingerRetraction0.014 ± 0.0040.120 ± 0.0020.273 ± 0.0060.310 ± 0.0170.355 ± 0.009
Parzen-smoothed KLHellingerDirect0.009 ± 0.0010.279 ± 0.0060.742 ± 0.0210.928 ± 0.0110.951 ± 0.008
Parzen-smoothed KLEuclideanRetraction0.029 ± 0.0030.123 ± 0.0040.281 ± 0.0120.435 ± 0.0060.453 ± 0.002
Parzen-smoothed KLEuclideanDirect0.010 ± 0.0020.136 ± 0.0030.308 ± 0.0140.550 ± 0.0630.692 ± 0.025

Entries show mean TV (↓\downarrow) with ±\pm standard deviation over five held-out training seeds. Bold marks the lowest mean for each drift and vocabulary size.

Therefore, in our experiments, incorporating the underlying geometry into both the objective and the training update improves agreement with the target pairwise marginals at larger vocabulary sizes. In the revised manuscript, we summarize these findings in Section 4.4 and provide the experimental details and full results in Appendix A.7.

AuthorsOpenReview ↗

Answer to Reviewer 6Eb7

[W.2] The image experiments lack a direct comparison with W-Flow under matched settings. […] The motivation for representing each pixel or channel as a categorical probability distribution also needs further justification.

Categorical image representation. Our goal is to extend one-step Wasserstein gradient-flow training from Euclidean spaces to Riemannian manifolds. As discussed in Section 4, R-WGF with the Sinkhorn objective recovers the W-Flow [2] update in Euclidean space. Image generation is one experimental domain in which we study this extension. Continuous intensities and categorical pixel values define distinct modeling settings, so we compare training methods within a fixed data representation. We focus on categorical images because W-Flow already studies the continuous Euclidean formulation. Following standard works PairFlow [3] and Categorical Flow Maps [4], we represent each pixel channel by a probability vector over discrete intensity levels.

We test this extension on MNIST, which lets us study both categorical and continuous representations within the rebuttal timeline and compute budget.

Categorical. On Binary-MNIST, we compare R-WGF-Sinkhorn and W-Flow using the same categorical pixel representation and generator architecture. The W-Flow baseline applies Euclidean updates directly to the pixel probabilities, using a squared-Euclidean ground cost and MSE tracking. R-WGF-Sinkhorn uses Fisher–Rao updates and Hellinger tracking.

Continuous. We evaluate a W-Flow run trained on grayscale MNIST. We emphasize that in this case W-Flow matches with R-WGF-Sinkhorn.

For all runs, we compute FID against the 60,000 Binary-MNIST training images. We decode categorical outputs by argmax.

Training stepCategorical — R-WGF-Sinkhorn ↓\downarrowCategorical — W-Flow ↓\downarrowContinuous — R-WGF-Sinkhorn / W-Flow ↓\downarrow
4,000204.37230.31267.56
8,00089.4756.13131.58
12,00057.5832.2734.08
16,00028.5532.9116.64
20,00028.1383.7411.04
24,00023.39127.7310.11
28,00016.91150.588.00
32,00013.31132.546.40
36,00012.44101.285.58
40,00013.48110.155.17

R-WGF-Sinkhorn improves more steadily and achieves lower FID than the categorical W-Flow adaptation. The trajectory suggests that applying W-Flow directly to categorical image generation without accounting for the geometry of the simplex can lead to unstable performance.

The continuous R-WGF-Sinkhorn / W-Flow run achieves the lowest FID. As discussed above, the continuous and categorical runs are trained on grayscale MNIST and Binary-MNIST, respectively, so this comparison addresses the choice of image representation rather than comparison of the methods. For example, with the same underlying W-Flow update, the categorical run yields a substantially worse than the continuous run. These results support adapting training updates to the geometry of the data representation. In the revised manuscript, we briefly describe this W-Flow comparison at the end of Section 5.2 and report the full details in Appendix B.2.

[1] Drifting Models (https://arxiv.org/abs/2602.04770)

[2] W-Flow (https://arxiv.org/abs/2605.11755)

[3] PairFlow (https://arxiv.org/abs/2512.20063)

[4] Categorical Flow Maps (https://arxiv.org/abs/2602.12233)

Concluding remarks. Thank you for your thoughtful feedback. We clarified the connection between R-WGF and direct optimization, added an ablation of the cost and update on categorical sequences, and compared R-WGF-Sinkhorn with a dicrect W-Flow adaptation on MNIST. We would be grateful if you could let us know whether these revisions address your concerns. If so, we kindly ask that you consider increasing your rating. We are also open to discussing any other questions you may have.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response to Reviewers

We thank the reviewers for their constructive feedback, for highlighting the strengths of our mathematical formulation, and for recognizing the competitive empirical performance of R-WGF. In this revision, we have addressed your comments by adding the material summarized below, which is highlighted in blue in the revised manuscript.

[6Eb7] New comparison with direct optimization. Section 4.4 derives the exact gradient equivalence with direct optimization. The same paragraph explains that this exact identity relies on an idealized update and can break down with the finite retraction steps used in practice. Our ablation in Appendix A.7 shows their experimental difference.

[6Eb7] Comparison with W-Flow. We compare R-WGF-Sinkhorn with two W-Flow variants on categorical and continuous representations of MNIST. The revised manuscript now briefly describes the W-Flow comparison at the end of Section 5.2 and presents all experemntal detail in Appendix B.2.

[7gzg, iffN] New manifold benchmarks. We derive the updates for flat tori, SO(3)\mathrm{SO}(3), and SE(3)n\mathrm{SE}(3)^n (Table 1 and Appendices A.4–A.6) and add one-step benchmarks with Riemannian MeanFlow comparisons (Section 5.4 and Appendix B).

[iffN] Clarifications and revisions. We specify the fixed output manifolds and explain how per-position categorical probabilities for images and text define a product of simplices. We move Related Work immediately after Background, and expand the conclusion to discuss limitations and open directions (Section 6).

New LM1B results. We would like to draw the reviewers' attention to the updated one-step LM1B results in Table 4 (Section 5.3). After refining the R-WGF-Sinkhorn training setup, we reduced GenPPL and increased entropy. The final results are shown in the table below.

MethodsGenPPL ↓\downarrowEntropy ↑\uparrow
DFM (PSD)94.084.06
FMLM119.344.16
R-WGF-Sinkhorn (Ours)112.974.20

R-WGF-Sinkhorn achieves the highest entropy among the methods shown and improves on the reported FMLM baseline in both metrics.

Area_Chair_RrS3OpenReview ↗

[Urgent] Please check the authors' rebuttal soon

Dear Reviewers,

The authors have posted their responses to the issues you raised. As the discussion period will end soon, please take a look and have a discussion with the authors at your earliest convenience. Thank you very much!

Best, AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.