acceptodds
Under review as a conference paper at ICLR 2027

One-Step Generative Modeling via Unbalanced Optimal Transport

124 views

Abstract

Drifting models enable one-step generation by amortizing distribution transport into training, but this efficiency places greater demands on the transport field estimated at each update. In large-scale training, the field is computed from finite mini-batches of generated and real samples, which provide only imperfect approximations to the underlying distributions. Balanced optimal transport enforces exact mass matching within every mini-batch, making the estimated field sensitive to the particular composition of the real-data batch. We find that generated and real samples should be treated asymmetrically: letting the mass assigned to real samples adapt while keeping every generated sample fully transported improves generation across six feature-space metrics in controlled ablations, and is more robust to the relaxation strength than relaxing both marginals simultaneously, which falls below balanced transport under stronger relaxation. Motivated by this observation, we propose Unbalanced Optimal Transport Gradient Flow (UOT-GF), which keeps the generated-sample marginal fixed and relaxes only the real-data marginal. Under identical settings at DiT-B/2 on ImageNet-256, UOT-GF improves Fréchet Inception Distance (FID) from 1.53 to 1.46 over the balanced W-Flow baseline; scaling the same recipe yields 1.34 and 1.22 FID at L/2 and XL/2, the best FID among the one-step models we compare. We further derive the induced UOT transport force, establish a kinetic Vlasov–Fokker–Planck formulation whose overdamped zero-temperature limit recovers the drifting dynamics, and characterize non-target stationary states together with sufficient conditions for convergence.

Primary area:
generative models
Keywords:
Generative modeling, one-step generation, unbalanced optimal transport, gradient flow

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_Lnr1OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

The paper aims to improve upon a competitive drifting-model baseline [W-Flow] by resolving a balanced-OT constraint that is too stringent for data minibatches, where a small sample is not guaranteed to represent the real/generated distributions, and makes the training of W-Flow sensitive to minibatch composition. The authors address this by proposing UOT-GF, which modifies W-Flow using source-fixed unbalanced OT. The method fixes the source marginal while relaxing the target marginal, and makes the transport contributions from both directions symmetric, controlling reweighing using a KL penalty. Comprehensive empirical comparisons to related works including W-Flow across many metrics are conducted, demonstrating advantage over W-Flow in many cases/settings. The authors also derive the transport force and provide theoretical analyses of the gradient-flow dynamics.

Critical strengths
  1. The authors develop a well-motivated method that aims to address the minibatch imbalance sensitivity of W-Flow.
  2. A detailed comparison of the performance of various (U)OT methods, as well as settings within each method (Tables 1, 2). The authors demonstrate the importance of making the source-fixed transport contributions symmetric, rather than using only the forward direction or fully unbalanced OT.
  3. The paper includes thorough empirical evaluation of their method, across model sizes and different types of metrics, and compares against relevant baselines.
Critical weaknesses

Major Weaknesses

  1. There is limited evidence for the proposed explanation of why UOT-GF solves the minibatch issue in W-Flow. The authors motivate UOT-GF as a way to reduce misleading transport directions caused by imbalanced minibatches. While toy examples/visualisations (see Fig 2, 4-7, 9) are helpful, it is unclear whether UOT actually produces more reliable transport directions during training. For example, the authors could use a frozen generator and compare transport directions estimated from many minibatches for W-Flow/UOT-GF, and use large batch estimates from each method as references.

    • Minibatch sampling error is an issue that the authors motivate UOT-GF, and the authors provide this as a motivation for how UOT-GF may address problems caused by imbalanced minibatches.
    • However, when the actual data distribution is imbalanced (eg 20% pens and 80% pencils in the "writing utensil" real images class), UOT-GF may still reweigh the transport mass assigned to real samples, potentially leading to more desirable transport directions, which could also explain improvements in performance metrics.
    • The authors seem to focus on the imbalanced minibatch explanation as to why UOT-GF works well. Maybe there is another mechanism which the authors could discuss/demonstrate empirically.
  2. Improvements across different evaluation metrics are mixed. Although UOT-GF improves FID, the improvements in FDr6 and Inception-Score are less consistent across model scales (Table 3). More discussion of these differences would help clarify different trade-offs, since these metrics measure different things. In particular, L/2 comparison with W-Flow shows a small FID improvement and worse FDr6.

    • The authors should consider reporting FID across several evaluation seeds for the W-Flow-L and UOT-GF-L (row "KL tau = 0.985 - L" in Table 3) checkpoints. The FID difference between these two rows is 0.02 (1.36 - 1.34). The improvement is only 0.02 FID, so this would help assess how much the FID varies due to sampling variation.

Minor Weaknesses

  1. Could UOT have difficulty with underrepresented modes? UOT can reduce the influence of oversampled or spurious targets, but it might also weaken the learning signal for legitimate modes that the generator has not learned (or has difficulty learning). Maybe authors can do a toy experiment involving underrepresented modes to clarify this trade-off.

  2. The theoretical analysis studies idealised population dynamics (Eq 10), while actual training uses minibatch estimates and approximate transport computations (Eq 12-13). The authors do note this limitation in Appendix I; however, a toy experiment would help empirically establish the relevance of the theoretical analysis.

  3. The authors use strong assumptions in the convergence analysis. The convergence guarantee (Proposition 5) relies on assumptions that may be difficult to satisfy when the generator does not adequately cover the data distribution (Assumption 4; Eq 27-28). While the authors acknowledge this limitation, further discussion of its implications for practical training would be helpful, particularly when modes are underrepresented. Ideally this would be supported by a toy experiment.

  4. The text is quite dense, especially where math is involved.

Some of my concerns are related to limitations that the authors have themselves admitted to. Despite this, I think a more detailed treatment would improve the paper.

Questions for authors

See weaknesses.

[W-Flow] One-Step Generative Modeling via Wasserstein Gradient Flows, https://arxiv.org/pdf/2605.11755

3 comments

We thank the reviewer for suggesting a direct evaluation of mini-batch transport directions. We appreciate the time and effort invested in evaluating our work. The revised manuscript adds a frozen-generator experiment that compares how closely W-Flow and UOT-GF reproduce their own large-batch directions (Section H.10). We also added a controlled rare-mode recovery experiment to examine whether UOT-GF can help when the target distribution itself is imbalanced (Section H.11).

W1

Following the reviewer’s suggestion, we froze a single B/2 generator at 200K steps and compared balanced W-Flow (τ=1\tau=1) with UOT-GF (τ=0.985\tau=0.985, 0.950.95) over 100 paired mini-batch draws (Section H.10). All settings shared the same samples, mini-batch indices, and solver iteration count. We compared each method’s full update, including debiasing and CFG, with its own large-batch reference. Both UOT settings reduced angular discrepancy in four feature spaces, with paired 95% confidence intervals below zero for the full 1,024-sample reference and all ten 512-sample subsets. This shows that, on the frozen generator, UOT-GF more consistently reproduces its own large-batch direction from small batches in those four spaces.

To test whether UOT also helps when the target distribution is genuinely imbalanced, we used a two-component Gaussian mixture with a rare mode of 6.25% (Section H.11, Table 7). We initialized the rare component with either 8 or 0 of 512 particles, compared with a target of 32, and evolved the particles for 2,000 updates over five paired runs. Across the two initializations and two sampling schemes, UOT recovered 4.02%–4.18% right-region mass versus 3.28%–3.48% for balanced W-Flow, reducing mean terminal mass error by 24%–26%. The advantage persisted when every 64-sample target batch contained exactly four rare-mode samples. This control preserves the true imbalance while removing fluctuations in rare-mode counts, showing that these fluctuations alone cannot explain the recovery gain. The specific contribution of real-sample reweighting to the image-generation gains remains open.

W2

In Table 3, UOT-GF improves most metrics at B/2 and XL/2, with FID decreasing from 1.52 to 1.46 and 1.33 to 1.22, respectively. At L/2, FID, our primary metric, improves from 1.36 to 1.34, while FDr6 worsens. Across 50 evaluation seeds for the same L/2 checkpoints, the observed FID variation was approximately 0.0018, well below the reported gap of 0.02. These metrics capture different aspects of generation quality. FID measures distance in Inception feature space, FDr6 averages distances across six feature spaces, and Inception Score reflects classifier confidence and class diversity. We will discuss these trade-offs across model scales in the final version.

W3

We conducted the suggested rare-mode recovery experiment and report its setup and quantitative results in our response to W1 and Reviewer 4y3J (Q1). Symmetrized source-fixed UOT recovered more of the missing rare-mode mass than balanced W-Flow when the mode was initially underrepresented or absent, although recovery remained partial for both methods.

W4

To connect the population analysis in Eq. (10) to the update used in training, we added a finite-mini-batch analysis in Section 3.4 and a controlled particle experiment in Appendix H.11. With exact transport solves, the theory shows that symmetrized source-fixed UOT allows real-sample reweighting relative to balanced OT and, under the stated assumptions, yields lower expected one-step error than two-sided UOT under the full debiased update. The toy experiment tests the training implementation directly by evolving particles with the same debiased mini-batch transport field and 10 Sinkhorn iterations. Under these finite-iteration updates, UOT recovers more of the initially missing rare-mode mass than balanced W-Flow.

W5

When the generator underrepresents a target mode, the coverage assumptions in our convergence analysis can be difficult to verify. The transport update remains well defined, but energy descent alone cannot establish that the generator is recovering the missing mode or converging to the target. To examine the behavior of the energy empirically during training, we added an experiment evaluating the full debiased energy at 11 checkpoints for four classes and four sampling seeds per checkpoint–class pair. All 176 estimates were strictly positive, providing evidence of positivity for the evaluated distributions. Convergence to the target remains conditional on the stated assumptions. We will clarify this distinction in the final version. Further details are in our response to Reviewer 4y3J under “Weaknesses.”

W6

In the final version, we will add plain-language explanations alongside the main mathematical results so that readers less familiar with the theory can follow the intuition and understand what the results mean in practice.

AuthorsOpenReview ↗

Checking In on Our Response and Revised Manuscript

Dear Reviewer Lnr1,

With the discussion phase wrapping up soon, we wanted to briefly check in and hope that our rebuttal and new results have satisfactorily addressed all of your questions. If you feel that the previous concerns are now resolved without any major remaining issues, we would deeply appreciate it if you might consider updating your evaluation accordingly.

We truly appreciate the time and effort you have dedicated to reviewing our paper, and we remain at your disposal should you have any further questions.

Warm regards, The Authors

Thank you for the rebuttal.

The new results in section H.10, H.11, and additional W-Flow baselines provide useful additional evidence. I expected two additional sets of seeds but the evaluation over 50 inference seed sets also addresses my concern at size L/2

I still think the explanation for why UOT-GF improves performance is not fully established, though there is indeed a performance improvement. Specifically better agreement with each method's own large-batch transport directions does not necessarily imply better generation performance. However, the new experiments provide a bit more support for the motivation that the authors claim.

Additionally, the text in both the authors' response and the paper is difficult for me to read. I hope the authors will go through the paper to brush up on the writing and make sure things are clear and not too terse.

Overall, I think the authors have addressed most of my concerns sufficiently, and I will maintain my marginal-accept recommendation since the paper is decent.

I expect the authors to release code and artefacts (eg checkpoints) upon acceptance in support of open source research, another reviewer also mentioned open-sourcing code.

Reviewer_4y3JOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies how to generate images in one step. It points out a problem with methods such as W-Flow: balanced optimal transport requires strict mass matching in every mini-batch. This makes training sensitive to which samples happen to be in the batch. The authors propose UOT-GF, which keeps the generated-sample weights fixed but allows the real-data weights to change through a KL penalty. They support the method with detailed theory and experiments on ImageNet-256. Using DiT-XL/2, the method achieves an FID of 1.22 in one forward pass.

Critical strengths

Strong theoretical support. The authors carefully derive the method and explain how it connects to a broader mathematical model. They also study how the energy decreases and when the dynamics may stop before reaching the target distribution.

A clear and useful main idea. The method treats generated and real samples differently. By allowing only real-data weights to change, it reduces the effect of noisy samples and modes that appear too often in a batch. The toy examples and ImageNet analyses help explain why this works.

Strong results. The paper includes many controlled experiments across six feature-space metrics. The method works well at larger model sizes and reaches an FID of 1.22 on ImageNet-256. This is a strong result for a model trained from scratch that generates images in one step.

Critical weaknesses

One theoretical question remains open. The authors have not proved that the proposed energy is positive definite. It would help to explain what this missing proof means for the method and to add tests of training stability.

Rare modes need more study. The method can give less weight to some real samples. This may help with noise, but it could also reduce the influence of rare or difficult examples. Overall FID and recall may not show these effects. Results for individual classes and rare variations within each class would make the evidence stronger.

Questions for authors
  • Figure 4 shows that some real samples receive less weight, with mass ratios around 92%. Have you checked recall for individual ImageNet classes or rare variations within a class? Could the method give less weight to valid but difficult examples and therefore generate them less often?

  • The method solves two transport problems to compute the update. Compared with balanced W-Flow, how much extra GPU memory and time per training step does this require under the same settings?

3 comments

We thank the reviewer for raising questions about the energy’s positive definiteness, numerical training stability, and coverage of underrepresented modes. In response, we report empirical energy and training diagnostics, a per-class recall analysis, and a controlled rare-mode recovery experiment. We will clarify the scope of our convergence guarantees and carefully address the presentation comments in the final version.

Weaknesses

We thank the reviewer for highlighting this limitation. Our derivation gives a well-defined transport update from the proposed energy. The missing proof limits what we can conclude from energy descent. Even if the energy decreases, our current theory does not establish that the generated distribution approaches the target, whose status as the unique global minimizer remains open. Our convergence guarantees therefore remain conditional on the assumptions in the analysis. We will make this distinction explicit in the final version.

We complemented the theory with empirical diagnostics. We evaluated the full debiased energy at 11 training checkpoints for four classes, repeating each checkpoint–class evaluation with four sampling seeds. This design lets us check whether the sign of the energy estimate is consistent across training stages, class-conditioned distributions, and sample draws. All 176 estimates were strictly positive, providing empirical evidence of positivity for the evaluated distributions. However, the global condition required for positive definiteness, Ep(q)>0\mathcal E_p(q)>0 for every q≠pq\ne p, remains open. In practice, the first variation still defines the transport update, but a low or even zero energy value alone cannot certify that the generator has matched the target. We therefore assess numerical training stability directly using the diagnostics below.

We also added diagnostic curves from an existing 200K-step UOT-GF B/2 run (Fig. 10). The loss, gradient norms, and all 86 raw-velocity mean-square metrics remained finite throughout the recorded run. From 20K steps onward, all recorded pre-clipping gradient norms were below 0.247, and the transport-velocity RMS stayed within a stable range. At a fixed CFG scale of 1.2, FID improved overall from 12.344 at 5K to 1.507 at 200K, with intermediate fluctuations. These diagnostics support numerical stability for the monitored run, while a general convergence guarantee remains an open theoretical question.

Q1

We agree that improvements in aggregate metrics do not guarantee better coverage for every single class. To investigate this concern, we compared the 200K-step B/2 checkpoints at the same CFG scale of 1.19, using 50 held-out real images per class and 1,000 generated images per class for each of three paired generation seeds. Preliminary results across all 1,000 classes show that UOT-GF improves mean recall over balanced W-Flow by 0.503 percentage points in Inception feature space and 0.807 percentage points in DINOv2 feature space. However, mean recall across the three seeds was lower in 306 and 266 classes, respectively. Recall was lower for all three seeds in 64 and 46 classes, respectively. These results indicate that average gains coexist with class-specific reductions in measured recall, and we therefore cannot rule out adverse effects on individual classes. This comparison does not establish whether transport reweighting causes these reductions.

We further tested whether relaxing the target marginal weakens recovery of a genuine rare mode in a two-component Gaussian mixture (see Section H.11). The rare component accounts for 6.25% of the target distribution, so a correctly covered mode would contain 32 of the 512 equally weighted particles. We initialized it with either 8 or 0 particles and evolved the particles for 2,000 updates using the full debiased transport field of balanced W-Flow or symmetrized source-fixed UOT. The two methods shared initial particle positions and sampling indices in each of five paired runs. Alongside right-region mass, we measured mass near the rare-mode center to check whether particles reached the target component.

Across the four recovery conditions in Table 7, the final right-region mass was 3.28%–3.48% for balanced W-Flow and 4.02%–4.18% for UOT, corresponding to a 24%–26% reduction in mean terminal mass error. The advantage persisted when every 64-sample target mini-batch contained exactly four rare-mode samples. This control preserves the true imbalance while removing fluctuations in rare-mode counts, showing that the recovery benefit extends beyond reducing sensitivity to those fluctuations. Recovery remained partial for both methods.

Q2

On NVIDIA H20 GPUs, UOT-GF showed no increase in GPU memory usage under the same training settings. Compared with balanced W-Flow, the median time per training step increased from 1.96 s to 2.45 s at B/2 and from 3.88 s to 4.60 s at XL/2.

Thank you for the helpful rebuttal. The clarification on positive definiteness, the added training stability diagnostics, and the per-class recall and rare-mode experiments address my main concerns. The GPU memory and timing measurements also answer my question about training cost.

I find the approach interesting and think this is a strong piece of work. I’m raising my score from 3 to 4 and look forward to seeing the code made publicly available.

Dear Reviewer 4y3J,

Thank you so much for your kind words, your prompt engagement during the discussion phase, and for raising your score! We truly appreciate your insightful suggestions, they have helped us make the paper much stronger and more rigorous!

As promised, we will publicly release the full codebase, pretrained models, and weights immediately upon acceptance so the community can easily build on our work. Thank you once again for your time, support, and constructive review!

Warmest regards, The Authors

Reviewer_SoR7OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes UOT-GF, a one-step generation method that replaces the balanced Sinkhorn solve in W-Flow with an unbalanced one, which keeps the generated-sample marginal exact while softly relaxing the real-sample marginal through a KL penalty. The transport problem is solved in both the generated-to-real and real-to-generated directions, and the two resulting velocity estimates are averaged to form the training target. The paper also provides a theoretical analysis that connects the induced transport force to the drifting dynamics and gives conditions for convergence. Experiments on class-conditional ImageNet-256 show that the method improves FID over W-Flow from 1.53 to 1.46 at B/2 under matched settings and reaches 1.22 FID at XL/2 with a single network evaluation. The paper concludes that relaxing only the real-data marginal is at least as effective as relaxing both marginals.

Critical strengths
  1. Clear research question. The paper focuses on which marginal should be relaxed in drifting training, and put balanced OT, fully unbalanced OT and one-sided relaxation into one framework to compare, W-Flow is a special case of it.
  2. Extensive experiments. The paper does many ablations on marginal divergence, relaxation strength, Sinkhorn iterations and entropic regularization, and evaluates with not only FID but also many other metrics, which discuss the effectiveness of the method relatively sufficiently.
  3. The comparison with baseline is solid. The paper retains the balanced baseline by itself and it can match the released W-Flow checkpoint, and also reports the variance of different training seeds. This make the B/2 comparison more convincing.
Critical weaknesses
  1. The improvement might depend on carefully tuned settings. The gain over W-Flow is only 1.53 to 1.46 FID at B/2, and seems to depend on carefully searched hyperparameters. For example, only moving CFG from 1.19 to 1.15 changes FID to 1.51, almost the same as W-Flow. This may make it hard to extend the method to other models.
  2. The story of the paper does not match the final method. The paper says the method "relaxes only the real-data marginal", but the final method averages forward and reverse solves, so "Generated and real samples therefore each appear once on the fixed side and once on the relaxed side" (Section 1). Also, the forward-only version, which really follows this idea, gets 7.11 FID, even worse than balanced OT 7.07. So it is not clear relaxing "which marginal" is the real reason for improvement.
Questions for authors
  1. Can the theory explain why relaxing only the real-data marginal is better than relaxing both marginals or using balanced OT? It seems most of the theoretical results also hold for balanced W-Flow.
  2. Can you show balanced OT with the same solver settings at L/2 and XL/2, so the 1.34 and 1.22 FID have a fair baseline? And how robust is τ = 0.985 at large scale. For example, a small τ sweep at 200K steps?
  3. Is the CFG scale of the W-Flow baseline searched in the same careful way as your method?
3 comments

We thank the reviewer for the thoughtful and constructive feedback. The revised manuscript adds a mini-batch analysis of marginal relaxation, balanced OT baselines at L/2 and XL/2 with 10 Sinkhorn iterations, and an XL/2 sweep of τ\tau at 200K training steps. We will carefully address the presentation comments in the final version.

Weaknesses

The choice of marginal constraint matters in each directed solve. Our method fixes the source marginal and relaxes the target marginal before averaging the two plans. As shown in Section 3.4, this guarantees that each group retains at least half of its reference mass. Concurrent UW-Flow [1] instead studies two-sided KL-UOT, which relaxes both marginals within each solve. In our ablation, fully balanced OT achieves 7.07 FID, two-sided KL-UOT achieves 7.00 at its best tested setting, forward-only UOT achieves 7.11, and symmetrized source-fixed UOT achieves 6.86. These comparisons support both symmetrization and retaining a hard marginal constraint in each solve. In the final version, we will clarify that “relaxing only the real-data marginal” describes the forward solve and use “symmetrized source-fixed UOT” for the full method.

Q1

Our new analysis in Section 3.4 of the revised manuscript explains how relaxing the real-data marginal affects mini-batch updates. Compared with balanced OT, symmetrized source-fixed UOT can adapt the transport mass assigned to real samples within a mini-batch. It also provides a mass-retention guarantee that two-sided UOT [1] can violate. For every group II of generated samples, the effective coupling retains at least half of the group’s reference mass: \begin{equation} m_I(\bar{\mathbf P})\ge \tfrac12\omega^q(I). \end{equation} As the matching costs for a nonempty proper group II increase by a common offset Δcost\Delta_{\mathrm{cost}}, its share of the two-sided UOT plan decays as Θ(e−Δcost/(ε+ρ))\Theta(e^{-\Delta_{\mathrm{cost}}/(\varepsilon+\rho)}), even after normalizing the plan to unit mass.

The new theorem shows that this difference persists in the Euler update. In the large-offset regime, the expected fraction of squared normal error remaining after one step converges to \begin{equation} \begin{cases} 1-p_{\mathrm{hit}}(\eta-\eta^2/4), & \text{symmetrized source-fixed UOT},\\ 1, & \text{two-sided UOT}, \end{cases} \end{equation} where η\eta is the step size and phitp_{\mathrm{hit}} is the probability that the independent self batch contains a sample from the displaced group. Thus, when phit>0p_{\mathrm{hit}}>0, symmetrized source-fixed UOT achieves strictly lower expected one-step squared normal error for sufficiently large offsets. The precise statements and assumptions appear in the subsection “Mass Retention for Hard-to-Match Samples in Mini-Batches” of our revised submission.

Q2

We added balanced W-Flow baselines at L/2 and XL/2, each trained for 200K steps with 10 Sinkhorn iterations, matching UOT-GF’s iteration count. Under ADM, UOT-GF achieves 1.3413 vs. 1.3748 FID at L/2 and 1.2196 vs. 1.3613 at XL/2. The same ordering holds under JiT, with 1.3481 vs. 1.3720 at L/2 and 1.2252 vs. 1.3325 at XL/2.

We also ran a τ\tau sweep at XL/2 for 200K steps. For τ∈0.95,0.97,0.975,0.985\tau\in\\{0.95,0.97,0.975,0.985\\}, the ADM FIDs are 1.2767, 1.2388, 1.2215, and 1.2196, respectively. All four settings outperform the balanced XL/2 baseline with 10 Sinkhorn iterations under both evaluation protocols. Among the tested values, τ=0.985\tau=0.985 gives the lowest FID. Both comparisons are included in the revised manuscript.

FID under the JiT and ADM evaluation protocols.

Model / settingCFGFID (JiT) ↓\downarrowFID (ADM) ↓\downarrow
W-Flow-L (200k, i=10i=10)1.141.3720181.374806
W-Flow-XL (200k, i=10i=10)1.091.3325481.361274
Séjourné KL-UOT-B (τ=0.985\tau=0.985)1.191.5473851.535807
UOT-GF-L (τ=0.985\tau=0.985)1.151.3481071.341342
UOT-GF-XL 200k (τ=0.95\tau=0.95)1.141.2847211.276738
UOT-GF-XL 200k (τ=0.97\tau=0.97)1.141.2367421.238770
UOT-GF-XL 200k (τ=0.975\tau=0.975)1.141.2258001.221548
UOT-GF-XL 100k (τ=0.985\tau=0.985)1.141.2611741.261597
UOT-GF-XL 200k (τ=0.985\tau=0.985)1.141.2252181.219600

Q3

YES. We follow the same CFG procedure as W-Flow during training. At inference, we use the CFG scales carefully selected by the original W-Flow authors and tune UOT-GF separately. At XL/2, the selected scale is 1.09 for W-Flow and 1.14 for UOT-GF. This shift in the optimal guidance scale is consistent with UOT changing the training dynamics and, in turn, how the trained model responds to guidance.


[1] Jeon, Jongwook, et al. "Unbalanced W-Flow: Unbalanced Wasserstein Drifting under Finite-Batch Marginal Fluctuations." arXiv preprint arXiv:2610.18552 (2026).

AuthorsOpenReview ↗

Gentle Follow-up on Our Rebuttal

Dear Reviewer SoR7,

As the discussion period is drawing to a close, we wanted to gently follow up to see if you are satisfied with our responses and the revised manuscript. If our clarifications have resolved your concerns and no major weaknesses remain, we would be very grateful if you would consider reflecting this in a higher assessment of our work. Please let us know if there is anything else we can clarify!

Thank you again for your valuable time and constructive feedback, and please let us know if there is anything else we can clarify!

Best regards, The Authors

Thanks for the authors' detailed responses. The additional experiments have addressed most of my concerns, and I have no further questions. Hope these new experiments can be incorporated into the final version of the paper and I will maintain the current score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal Revisions and New Experiments

Dear Area Chair and Reviewers,

As the discussion period draws to a close, we would like to sincerely thank all three reviewers (SoR7, 4y3J, and Lnr1) for their constructive feedback. We are very encouraged that all reviewers initially rated our work positively (3, 3, 3) and that Reviewer 4y3J has already raised their score to 4 (Clear Acceptance) following our rebuttal.

To thoroughly address every question raised in the reviews, we conducted extensive new large-scale experiments, mechanism diagnostics, and theoretical analyses during the rebuttal period. Notably, each full ImageNet-256 benchmark run trained from scratch requires 64 GPUs, representing a substantial computational effort to ensure fair, rigorous, and conclusive comparisons:

  1. Large-Scale Baselines and τ\tau Sweep at L and XL (64 GPUs per run; for Reviewer SoR7):

    • Trained matched balanced W-Flow baselines from scratch at DiT-L/2 and DiT-XL/2 for 200K steps with 1010 Sinkhorn iterations (matching UOT-GF's solver budget). UOT-GF consistently outperforms balanced W-Flow under both ADM and JiT protocols (improving XL/2 ADM FID from 1.361 to 1.220).
    • Conducted a 4-point τ\tau sweep at DiT-XL/2 (200K steps) across τ∈0.95,0.97,0.975,0.985\tau \in \\{0.95, 0.97, 0.975, 0.985\\}. Every tested τ\tau outperforms the matched balanced XL/2 baseline, confirming strong robustness at scale.
  2. Finite Mini-Batch Theory (Section 3.4; for Reviewers SoR7 and Lnr1):

    • Proved a new finite-batch theorem showing that symmetrized source-fixed UOT guarantees at least 1/21/2 reference mass retention for any sample group (whereas two-sided UOT decays exponentially with cost offsets) and achieves strictly lower expected one-step squared normal error in the Euler update.
  3. Mini-Batch Direction Fidelity & Rare-Mode Coverage (Sections H.10 & H.11; for Reviewers 4y3J and Lnr1):

    • Frozen-generator mini-batch alignment (Section H.10): Across 100 paired draws and four feature spaces, UOT-GF stays closer in angle to its large-batch reference transport direction than balanced W-Flow.
    • Rare-mode recovery (Section H.11) & Per-class recall: On a controlled 6.25% rare-mode task, UOT-GF reduces terminal mass error by 24%–26% over balanced W-Flow. We also added a full 1,000-class ImageNet recall breakdown across three seeds in both Inception and DINOv2 spaces.
  4. Energy Positivity, Stability Diagnostics, Evaluation Variance, and Profiling (for Reviewers 4y3J and Lnr1):

    • Evaluated the full debiased energy across 176 checkpoint/class/seed combinations (all strictly positive), added 200K-step training stability trajectories (Fig. 10), verified L/2 FID variance across 50 evaluation seeds), and reported exact H20 GPU memory and step-time measurements.

All new results and analyses have been incorporated into the revised manuscript. We deeply appreciate the reviewers' and AC's time and consideration, and we remain happy to address any final questions!

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.