acceptodds
Under review as a conference paper at ICLR 2027

Unbalanced W-Flow: Unbalanced Wasserstein Drifting under Finite-Batch Marginal Fluctuations

21 views

Abstract

Teacher-free one-step generative models such as W-Flow learn directly from distribu- tional objectives, with the Sinkhorn divergence providing the strongest empirical perfor- mance among the energies studied in W-Flow. However, Sinkhorn divergence is based on balanced optimal transport, whose exact marginal constraints can induce undesirable transport under finite minibatch fluctuations. We introduce UW-Flow, which replaces the balanced Sinkhorn energy with a debiased unbalanced Sinkhorn divergence and derives the corresponding Wasserstein velocity. The resulting dynamics retain W-Flow’s cross- minus-self structure, recover the balanced limit, and preserve target identifiability under suitable assumptions. We further characterize the finite-batch velocity error of balanced and unbalanced transport by comparing each minibatch field with its population counter- part. Our analysis identifies conditions under which marginal relaxation yields smaller minibatch-to-population velocity error, and shows that for separated matched modes, bal- anced sensitivity grows with inter-mode separation while the corresponding UOT (Un- balanced Optimal Transport) sensitivity remains controlled by local transport geometry. Controlled experiments on fixed-support and continuous two-mode distributions support these predictions. On MNIST, CIFAR-10, and ImageNet-100, UW-Flow improves gener- ation under matched training budgets, with the largest gains in unconditional generation and consistent improvements in the class-conditional setting.

Primary area:
generative models
Keywords:
generative models

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_AvPQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes UW-Flow, a teacher-free one-step generative method that extends W-Flow by introducing unbalanced optimal transport. The method aims to reduce the undesirable transport caused by marginal fluctuations in finite minibatches. The authors replace the balanced Sinkhorn divergence with a debiased unbalanced Sinkhorn divergence and derive the corresponding Wasserstein velocity field. They provide theoretical analyses of target identifiability and finite-batch velocity estimation errors. The analysis shows that unbalanced transport can reduce velocity estimation errors under certain conditions, particularly for well-separated modes. Experiments on synthetic distributions, MNIST, CIFAR-10, and ImageNet-100 demonstrate improvements in generation quality over W-Flow.

Critical strengths
  • Clear motivation and well-designed method. The paper identifies an important limitation of balanced optimal transport under finite-batch marginal fluctuations. The proposed unbalanced formulation directly addresses this issue while retaining the main structure of W-Flow. The method is simple and has a clear connection to the identified problem.

  • Meaningful theoretical analysis. The paper provides a theoretical characterization of finite-batch velocity estimation errors for balanced and unbalanced transport. The separated-mode analysis offers useful insights into why marginal relaxation can improve stability. These results strengthen the motivation for the proposed method.

  • Solid experimental validation. The experiments cover synthetic distributions and multiple image generation benchmarks. UW-Flow consistently improves FID over W-Flow under matched training budgets, with particularly notable gains in unconditional generation. The batch-size ablations also provide empirical support for the theoretical analysis.

Critical weaknesses
  • The paper shows that unbalanced optimal transport can reduce finite-batch velocity estimation errors under certain conditions. However, the balanced and unbalanced methods are compared against their own population velocity fields, which are generally different. Therefore, a smaller estimation error does not necessarily imply better generation dynamics or convergence to the target distribution. The paper lacks direct evidence connecting the improved velocity stability to the observed gains in generation quality. Additional experiments measuring velocity errors and generation performance during training would strengthen this connection.

  • The theoretical stability advantage of UW-Flow is established mainly under fixed-support assumptions and a specific separated-mode setting. However, real-world data distributions can have more complex structures, including overlapping modes and varying degrees of imbalance. It remains unclear whether the theoretical advantages extend to these settings. Moreover, the experimental results depend on the relaxation parameter, which is selected based on generation performance. A more systematic evaluation across different distribution structures, transport batch sizes, and relaxation strengths would help establish when UW-Flow provides reliable improvements over balanced transport.

Questions for authors

Comparison with recent UOT-based one-step generation methods. The recent paper One-Step Generative Modeling via Unbalanced Optimal Transport (arXiv:2609.32708) studies a closely related problem and proposes UOT-GF, which relaxes only the target-data marginal while keeping the generated-sample marginal fixed. Could the authors clarify the methodological and theoretical differences between UW-Flow and UOT-GF? In particular, what are the advantages of relaxing both marginals rather than only the target marginal? A direct comparison under matched settings would help establish the distinctive contributions and practical benefits of UW-Flow.

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

This paper proposes UW-Flow, an extension of W-Flow that replaces the balanced Sinkhorn divergence with a debiased unbalanced Sinkhorn divergence. The main idea is to relax the strict marginal constraints of balanced optimal transport, which can introduce undesirable transport when minibatch distributions do not match. The authors derive the corresponding Wasserstein velocity field and analyze its finite-batch estimation error. They show that under certain conditions, unbalanced transport can reduce the sensitivity of the velocity field to minibatch fluctuations. Experiments on synthetic distributions, MNIST, CIFAR-10, and ImageNet-100 show improvements over the original W-Flow, especially for unconditional generation.

Critical strengths
  • The paper addresses a meaningful limitation of minibatch optimal transport in W-Flow. The motivation for relaxing marginal constraints is clear and reasonable.

  • The authors derive the velocity field under unbalanced Sinkhorn divergence and provide an analysis of finite-batch estimation errors. The separated-mode experiments also help illustrate the theoretical findings.

  • The experimental results are encouraging. UW-Flow consistently improves FID over W-Flow across different transport batch sizes on CIFAR-10, with particularly large improvements at smaller batch sizes.

Critical weaknesses
  • Limited methodological novelty. The main modification is replacing balanced Sinkhorn divergence with its unbalanced counterpart within the existing W-Flow framework. Unbalanced optimal transport has already been studied for minibatch transport and generative modeling. Although the theoretical analysis adds value, the methodological contribution seems relatively incremental.

  • The experimental evaluation is somewhat limited. The main comparisons are against W-Flow, and the ImageNet experiments are conducted on ImageNet-100 rather than the full dataset. It is therefore difficult to assess how much the method improves over other recent one-step generation approaches, or whether the observed improvements extend to more challenging settings.

  • Sensitivity to the marginal relaxation parameter. The results depend on the choice of tau, and the best value varies across datasets and settings. The paper does not provide a practical strategy for selecting this parameter, which may limit the applicability of the method.

Questions for authors
  • Theorem 1 shows that the minibatch UOT velocity can have a smaller error relative to its own population field. However, balanced and unbalanced transport have different population velocity fields. Is there any theoretical or empirical evidence showing that reducing this estimation error directly leads to better generation? For example, could the authors examine whether the measured velocity errors correlate with FID during training?

  • The optimal tau appears to vary across different datasets and experimental settings. Is there any intuition or practical guideline for selecting tau? Can it be related to the minibatch size, distribution complexity, or degree of marginal mismatch?

  • The experiments mainly compare UW-Flow against W-Flow. It would be helpful to include comparisons with other recent teacher-free one-step generative models under similar settings, to better understand the practical significance of the improvements.

  • Does unbalanced Sinkhorn transport introduce additional computational cost compared with balanced Sinkhorn transport? It would be useful to report training time or computational cost, particularly since one of the motivations is improving training under limited minibatch support.

Reviewer_9fe1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes UW-Flow, a one-step generative model. It builds on the drifting framework but replaces the drifting velocity field by the Wasserstein gradient of a debiased entropic unbalanced optimal transport (UOT) divergence. Using entropic UOT instead of entropic OT relaxes the hard marginal constraints, which helps mitigate the errors induced by mini-batch OT. The authors theoretically compare the population velocity field with its mini-batch approximation for both entropic OT and entropic UOT. They also compare the two approaches in practice on 2D datasets and large-scale image datasets, where UW-Flow generally achieves better performance.

Critical strengths

The paper is relatively clear and has a clear objective: improve W-Flow [1]. The motivation is supported by previous work showing that marginal relaxation can mitigate problems caused by mismatched mini-batch proportions. The analysis of mini-batch velocity errors is an interesting contribution. The experiments provide useful comparisons with W-Flow under matched architectures and training budgets and achieve consistant improvements.

Critical weaknesses

The methodological contribution is a relatively direct extension of W-Flow using an existing unbalanced Sinkhorn divergence. There is also substantial overlap with the recent work [2]. The mini-batch analysis appears to be the main distinct contribution here. Given the recency of this work, I would welcome a discussion of the relationship rather than treat its omission as a major issue by itself.

The theoretical results are interesting, but I wonder how directly it applies in practice.

Questions for authors

-I am a bit skeptical about the choice to relax both marginals, since relaxing the generated-sample marginal could reduce the contribution of some generated samples to the transport field. Could the authors elaborate on this choice? This question is related to [2], which uses source-fixed UOT.

  • Relation to direct divergence minimization: As stated in Prop 8 of [3], when instantiated with a W2 gradient, drifting is a parametric divergence minimization. Could the authors clarify this relationship and identify which practical choices break the equivalence ? In particular, in what sense does their method amount to minimizing the debiased entropic UOT divergence with respect to the generator parameters? I think this distinction is discussed in the appendix of [1], but it could be explained in more detail here for the divergence considered. It could permit to differentiate with earlier works like [4] and [5] in the balanced case.

  • Out of curiosity, the reported FID scores here do not appear to be state of the art (contrary to [1]). Could the authors clarify whether this gap is mainly due to computational constraints, architectural choices, or differences in evaluation protocols? This is not at all a major limitation, as the comparison with W-Flow is sufficient to support the main empirical claim.

  • The main theoretical comparison is stated for sufficiently large batches, but the largest experimental gains on cifar, for example, are obtained for B=4B=4 or B=8B=8, where the asymptotic approximation is least clearly justified. Could the authors elaborate on this? I also have the intuition that smaller values of τ\tau make the condition Aτ<L∞2A_\tau < L_\infty^2 in Theorem 1 easier to satisfy. Is there any monotonicity result supporting this intuition?

  • Table 7 indicates that you normalize the velocity, so the practical updates differ from the Wasserstein velocity described in the main text ? Does your velocity error comparison remain valid after this modification? If this is the cas, it should be mentioned in the main text.

Minor comments and typos:

  • Figure 1 appears to be missing. Figure 2 is not cited in the text and its caption could explicitly state that each panel uses its own population velocity.

-AτA_\tau is not explicitly defined in the main text

  • The convergence of the velocity field as τ→∞\tau \to \infty, asserted in Section 4.2, is not obvious to me (convergence of the energy alone does not establish convergence of their gradients).

References: [1] Han, Li & al. One-Step Generative Modeling via Wasserstein Gradient Flows. arXiv:2605.11755, 2026.

[2] Shen, Xia & al. One-step Generative modeling via unbalanced optimal transport. arXiv:2609.32708, 2026.

[3] Gretton, Wenliang & al. On the Wasserstein Gradient Flow Interpretation of Drifting Models. arXiv:2605.05118, 2026.

[4] Fatras, Sejourne & al. Unbalanced Minibatch OT. ICML, 2021.

[5] Genevay, Peyre & al. Learning Generative Models with Sinkhorn Divergences. AISTATS, 2018.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.