acceptodds
Under review as a conference paper at ICLR 2027

Global Transport Couplings for Classifier-Free Guided Flows

Abstract

Optimal-transport couplings have been shown to reduce training variance in unconditional flow models, but their role in conditional generation remains unclear. A natural approach constructs separate couplings for each condition, but this is impractical for large or continuous conditioning spaces found in modern image foundation models. We introduce Global Transport (GT), a global class-agnostic optimal-transport coupling, computed without class labels. GT can associate different conditions with different regions of the source noise, and consequently worsens performance without guidance. However, when combined with classifier-free guidance (CFG), GT consistently improves generation across domains, model scales, and sampling budgets. This reversal suggests that couplings for conditional flows should be evaluated both empirically and theoretically under the guided flow used at inference, rather than on unguided generation. We evaluate GT over both discrete class and continuous text conditioned image generation across model scales, and investigate how coupling choice alters guided trajectories. These results identify coupling design in the guided flow setting as a simple training time axis to improve performance without modifying existing architectures, samplers, or guidance mechanisms.

Primary area:
generative models
Keywords:
Flow-based generative models, Optimal Transport, Classifier-free guidance

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_PhpiOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces global transport which pairs noise and data using optimal transport. The paper presents an interesting findings where class-agnostic GT couplings can give improved performances in conditional generation task, outperforming the widely used CFG technique. The theoretical results are clear and show that GT can reduce the gap between conditional and unconditional guidance, with empirical results supporting the claim. The experiment results in ImageNet look promising, as the mini-batch approximation is practically tractable, and it shows reasonable performance improvements. The ablation studies well support the claims made in the paper such as different coupling techniques (independent, GT, and class-conditional) and weight scheduler.

Critical strengths

S1. The paper is well structured. Clear motivation and formulations. It was easy to follow and grasp the main contributions of the paper. The problem the paper is trying to solve, and the proposed solution are also straightforward.

S2. Technical novelty. The proposed method is quite straightforward and simple to implement. It is quite surprising finding and can be quite useful in practice. The theoretical results, though assumes ideal conditions, are clear and intuitive enough to understand the behavior of the proposed method. To my knowledge, the fact that class-agnostic coupling can help guided sampling is clearly a novel finding which previously have not been explored.

S3. Practical impact. The proposed method is tractable with mini-batch approximation, which gives potentials to scale. Considering that CFG is a widely used practice in modern generative models, the work has great potential to the field. The experiment results are broad, covering different model sizes, sampling steps, and weights, and overall look convincing.

Critical weaknesses

W1. Limited scope of experiments. While broadly covering the claims made in the paper, the experiments are done in relatively small scale dataset ImageNet, and the model sizes are significantly small compared to modern generative models. While I stand more optimistic that the proposed method will scale to larger generative models, it is not shown in the paper. And this is the main concern I have in this paper. Can the authors show how it performs in state-of-the-art models such as FLUX 3 or Qwen-Image?

W2. Ideal assumptions. The theoretical results are intuitive, but they rely on strong assumptions of ideal velocity fields. While these can be difficult to show, this assumption hardly holds true in real world apploications, due to discretization step errors, model capacity, optimization. Instead, can the authors validate the proposed theoretical results in a simple 2D experiments? To show the tightness of the gap bound quantitatively.

Questions for authors

Q1. Can this method applied to transfer tasks (e.g., img-img) where the source distribution is also a discrete set of data points.

2 comments
AuthorsOpenReview ↗

Response to Reviewer Phpi

We are grateful to the reviewer for their thoughtful review and positive appraisal of our work. We appreciate their recognition that GT is “easy to implement”, that our theoretical results are clear and intuitive, and that the method offers practical scalability through mini-batch approximation. We are particularly encouraged by their assessment that GT “has great potential to the field”, given the widespread use of CFG in modern generative models. We also share the reviewer's enthusiasm for the surprising empirical results, which hold across multiple conditions, model scales, and sampling budgets. We hope these findings inspire further investigation into the training axis as a means of improving CFG. Below, we address the reviewer's questions and suggestions, grouped by topic.

Scope of Experiments

We are grateful that the reviewer is optimistic about GT scaling to larger generative models utilized in industry applications. We agree that this would be an extremely exciting experiment for GT and that showing results scaling to FLUX 3 / Qwen-Image would be highly impactful. Unfortunately due to compute constraints, we are unable to complete these experiments within the discussion period due to the scale of the models. We do believe this is a useful set of experiments and are in the process of sourcing compute to complete them. We believe our imagenet-scale experiments, while not as large as current state-of-the-art models, are still quite useful to the community.

Empirical validation of theoretical results

We agree with the reviewer that theoretical results assume idealistic conditions and that empirical investigations on a proposed gap in synthetically constructed experiment is a useful analysis.

As part of the rebuttal we acknowledge the importance of further validating this claim in real-world applications. We therefore measure the prediction gap in ImageNet-256 experiments across GT, class-conditional OT and independent transport plans using SiT-B/2 model with results in Table R3.

Table R3

Couplingn = 256n = 1024n = 2048n = 4096n = 8192n = 32768
Independent5.95–––––
Class-cond. OT5.955.985.976.016.016.02
GT5.625.545.505.49––

We observe that GT has the lowest prediction gap across trajectory (250 Euler steps) correlating with our experimental results further validating our theoretical results. We thank the reviewer for raising this point and we include this analysis in section C.2.1 in the revised manuscript.

We highlight that the paper already provides similar analysis on the 40-Gaussian Mixture experiment (4 classes) in section 4. Figure 3 shows GT obtains the lowest prediction gap in a finite batch setting with velocity fields constructed as MLP networks empirically validates our theoretical results.

Transferability to image-image setting

Can this method applied to transfer tasks (e.g., img-img) where the source distribution is also a discrete set of data points.

The GT algorithm can be applied to transfer tasks such as img-img. However, we are somewhat skeptical that this is the right application for GT as the theory behind classifier-free guidance is only applicable in the Gaussian → data setting where we can perform approximate posterior inference by adding scores. In the data → data setting adding vectors no longer corresponds to adding scores nor approximate posterior inference. Therefore while GT can be applied in this setting procedurally we do not view it as a particularly promising direction at this time.

Conclusion

Thank you again for your time, encouraging feedback, and helpful suggestions. Please let us know if you have any further questions. We look forward to continuing the discussion.

We are grateful to the reviewer for their positive appraisal of our work and for helping us empirically strengthen our theoretical results by validating them in a real-world setting. As the rebuttal discussion is coming to a close, we would like to kindly ask whether our responses have adequately addressed your concerns. We would be happy to provide any additional clarifications or address any further questions during the remaining time of the rebuttal discussion.

Reviewer_Z6BrOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes global transport (GT) methods to reduce mismatches in condition pairs when training flow-matching (FM) models. The motivation for this paper is that p_{1}(x) is not equal to p_{1}(x|c), thus leveraging the Hungarian algorithm to reassign the noise in a mini-batch. The experimental results demonstrate its validity.

Critical strengths
  1. The proposed method is interesting. While p_{1}(x|c) is truly different from p_{1}(x). This leads to the CFG-drop harming the early training process. This paper proposes a simple yet effective method to alleviate this problem.

  2. The experimental results clearly demonstrate its validity.

Critical weaknesses
  1. The overall writing can be further improved. The main paper should clearly demonstrate how to calculate xt^\hat{x_{t}}. The current version requires the audience to look in the appendix for more details.

  2. Unclear mechanism behind conditional generation improvements. GT constructs noise–image pairs without explicitly considering the conditioning information. Although this changes the learned conditional and unconditional velocity fields, the paper does not fully explain why reducing global transport cost improves CFG-based generation. In particular, the theoretical analysis suggests that the prediction gap can vanish in the population-optimal transport limit, potentially eliminating the guidance signal. What determines the optimal coupling strength for conditional generation?

  3. Missing OT minibatch-size ablation. The paper fixes the OT minibatch size at 256 without systematically investigating its influence on generation performance. Since batch size controls the quality of the transport approximation, experiments across different OT batch sizes would clarify the relationship between transport cost, the CFG prediction gap, and generation quality.

Questions for authors

Is it possible to construct the GT coupling using a larger pool of samples, such as 256 noise–image pairs, while performing model optimization using a smaller training batch, such as 8? This would help determine whether the proposed method can be applied efficiently in memory-constrained training settings without sacrificing coupling quality.

To sum up, I tend to be positive about this paper due to its interesting findings. The rebuttal should focus on the weaknesses.

3 comments
AuthorsOpenReview ↗

Response to Reviewer Z6Br

We thank the reviewer for their positive appraisal of our work. We are happy to hear that the reviewer found GT “reduces mismatches” in conditional noise and sample pairs when training flow matching models. We are also grateful that the reviewer found “experimental results to demonstrate its validity”. We now directly address weaknesses and questions raised in the review grouped by topic.

Clarity

The main paper should clearly demonstrate how to calculate x^t\hat{x}_t.

As we do not use the x^t\hat{x}_t notation in this work we assume the reviewer was referring to an OT constructed source batch x0^=πGT(x0)\hat{x_0}=\pi_{GT}(x_0) which is used in Algorithm 1. We have therefore amended section 3 (General objective) to highlight the difference between x0x_0 (noise batch) and x^0\hat{x}_0 (rearranged noise batch).

The current version requires the audience to look in the appendix for more details.

We recognize that a reader currently needs to refer to the appendix to gather insight into derivations and proofs behind stated theoretical observations in section 4. We added pointers to relevant appendix sections into the main text to ease reader’s need to search for these in the appendix. Thank you for helping to improve the clarity of our work.

Why Global Transport works

The reviewer astutely notes that the paper does not fully explain why reducing global transport improves CFG-based generation. At this time we do not have a conclusive theoretical explanation of the effectiveness of GT. We regard this as an open question, and as Reviewer LruU noted, have hedged our theory appropriately in that light.

As to what determines the optimal coupling strength for conditional generation, we believe it depends on data (particularly dimension and structure), model, and optimization dynamics in a non-trivial way. At this point we can only offer empirical results as this is difficult to analyze theoretically.

Batch Size Sweeps for GT and class-conditional OT baseline

We are grateful to the reviewer for raising this important point and for pointing out that the OT batch is fixed to 256 in all our experiments. We agree that assessing this is especially important to provide a reliable baseline for class-conditional OT since it computes L2L_2 cost between the same class samples.

We sweep both class-conditional OT and GT for OT batch size of n=[512,1024,2048,4096,8192]n=[512, 1024, 2048, 4096, 8192] and we additionally conduct sweeps for class-conditional OT at n=[16384,32768]n=[16384, 32768] since we don’t see any improvement at smaller batch sizes (see Table R1). Larger batch size makes sense for class-conditional OT since at n=[16384,32768]n=[16384, 32768] we have ~16 and ~32 samples of each class represented in the OT batch (ImageNet-256 has 1,000 classes).

Table R1

Method (FID ↓)n=256n=512n=1024n=2048n=4096n=8192n=16384n=32768
SiT-B/2+GT4.154.114.174.174.044.05——
SiT-B/2+Class-cond. OT5.085.175.215.155.145.075.024.92

We observe that increasing batch size is beneficial for class-conditional OT with FID decreasing from 5.08 at n=256 to 4.92 at n=32,768. However, GT outperforms C$^2$OT at all batch sizes tested. Furthermore, even the smallest batch size GT model (n=256) outperforms the 128x larger class-conditional OT baseline (n=32,768) by a substantial margin (4.15 vs. 4.92).

This shows that GT can be successful at relatively small batch sizes, which increases its practical utility as a large batch requirement is a major compute bottleneck for large-scale generative models and has hampered existing OT-based methods.

We thank the reviewer for raising this point and we include the batch size sweep in sections 5.1 (Impact of batch size) and C.2.2 in the revised manuscript.

Is it possible to construct the GT coupling using a larger pool of samples, such as 256 noise–image pairs, while performing model optimization using a smaller training batch, such as 8? This would help determine whether the proposed method can be applied efficiently in memory-constrained training settings without sacrificing coupling quality.

Yes, this is possible, and a useful technique especially with much larger OT batch sizes. We use this technique for our batch size sweep for all methods where we keep the optimizer batch size at 256 and only ablate the OT batch size. This shows that GT can be applied in memory constrained settings without sacrificing coupling quality.

Conclusion

We sincerely thank the reviewer again for their helpful suggestions and positive feedback. We hope these revisions and additional experiments address your questions, and we would be happy to discuss any remaining points during the discussion window.

Thanks to the authors for the rebuttal. I have checked all the contents. The authors have addressed my concerns regarding the batch size and the writing problems. However, as acknowledged by the authors, the current manuscript cannot explain why GT can connect the condition pair via the reasonable theory. At this stage, I think the pros outweigh the cons, and thus I keep my positive rating.

Thank you for your follow up, continued positive appraisal of our work and for confirming that our batch size ablations and writing revisions addressed your concerns.

Regarding the theory, we agree that a full characterization of why finite-batch global couplings improve guided sampling is a rich open problem. At the same time, our Propositions 4.1–4.2 connect global transport cost to the CFG prediction gap and trajectory regularization (supported empirically from 2D mixtures to ImageNet-256 in Table R3). Much like CFG itself, where strong empirical utility preceded a complete theoretical picture, we hope our findings and initial bounds provide a useful foundation for future theoretical work in the community.

With the experimental and clarity questions from your review addressed, we would be grateful if you might consider reflecting these improvements in an updated assessment. Thank you again for your helpful engagement with our work.

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

GT trains CFG conditional flow models with condition-agnostic minibatch OT coupling (batch 256). It hurts unguided conditional sampling but consistently helps guided sampling, motivating evaluating couplings under the guided flow. Prop. 4.1 bounds the integrated prediction gap by excess transport cost; Prop. 4.2 orders costs GT ≤ C²OT ≤ independent. On ImageNet-256 (SiT-B/L/XL, 800K iters) GT beats independent at optimal w (64 NFE: 4.15/2.69/2.59 vs 5.16/2.95/2.78), reaches FID 1.91 with guidance interval, and shows gains for DMF distillation, dopri5, text-conditioned ImageNet and single-cell data. ≈3% overhead.

Critical strengths
  1. Clean, counter-intuitive finding that reverses Cheng & Schwing (2025) under CFG, consistent across scales, NFEs, distillation and text conditioning.
  2. Zero inference cost; theory is appropriately hedged.
Critical weaknesses
  1. The class-conditional OT baseline is degenerate. With β_m = 10 forbidding cross-class pairs, batch 256 and 1000 classes (~0.25 samples/class), within-class OT is a no-op; C²OT differs from independent by ≤0.1 FID everywhere. The claim of outperforming class-conditional couplings is therefore unsupported; soft-β or class-grouped batches were not tried.

  2. The mechanism contradicts the best variant. Lower cost → smaller gap predicts semi-discrete GT should be best, yet it is the worst at its optimum (B/2 FD_DINOv2 199.6 vs 123.6). A smaller gap also means GT partly cancels CFG, so “tolerates larger w” is a guidance rescaling, not extra robustness; at optimal w the gains are 0.15–0.19 FID at XL/2, single seed.

Questions for authors
  1. Report C²OT with soft β or class-grouped batches.

  2. Is Table 1’s 2.06 your reproduction or the official number? Align training budget and sampler.

  3. Why is SD-OT lowest-cost yet worst on FD_DINOv2? Does measured prediction gap track performance across variants?

  4. SiT-L/2 independent FID 5.50/5.74/2.95 at 16/32/64 NFE is non-monotone — same w sweep?

3 comments
AuthorsOpenReview ↗

Response to Reviewer LruU (Part 1/2)

We are grateful to the reviewer for their time and effort in reviewing our work. We appreciate that the reviewer identified GT to be “clean” and “counter-intuitive finding” consistent across scales, conditioning type, sampling budget and as a distillation strategy. We are also heartened that the reviewer appreciated “zero inference cost” and theoretical guarantees as one of the key strengths. We now address weaknesses and questions raised in the review, grouped by a theme.

Class-conditional OT baseline

The reviewer raises a very fair point: at n=256 across 1,000 classes, most classes appear at most once per batch, leaving limited room for within-class pairing (producing modest gains of up to 0.13 FID in Table 2). To directly test whether C$^2$OT was simply bottlenecked by batch size, we ran two new sets of experiments:

  • A sweep over matched batch sizes for both GT and C$^2$OT. Where we find that GT outperforms class-conditional OT across all batch sizes measured. Additionally, GT at batch size 256 outperforms the largest class-conditional OT batch size tested (32,768).
  • A C$^2$OT baseline with soft-beta which still underperforms GT.

We believe this additional experiment substantially strengthens our claim and provide further details of the experiment below and in sections 5.1 (Impact of batch size) and appendix C.2.2 of the revised manuscript.

OT Batch Size Sweep

We select OT batches of n=512 (≈0.5\approx0.5 samples per class), n=1024 (≈1\approx1 sample per class), 4096 (≈4\approx4 samples per class), 8192 (≈8\approx8 samples per class) for both class-conditional OT and GT to test whether increasing number of same-class samples positively improve both transport plans. We additionally train class-conditional OT with batches of n=16,384 (≈16\approx16 samples per class) and n=32,768 samples (≈32\approx32 samples per class) and provide the results in the table below (Table R1) as we observe limited to no improvement on n=[512,1024,4096,8192]n=[512, 1024, 4096, 8192]. We note that these batches are substantially larger than the one tested in the original C$^2$OT paper which used n=6400 for their class-conditioned imagenet experiments. We train with batch size 256 and compute the transport plan every i = n/256 batches.

Table R1

Method (FID ↓)n=256n=512n=1024n=2048n=4096n=8192n=16384n=32768
SiT-B/2+GT4.154.114.174.174.044.05——
SiT-B/2+Class-cond. OT5.085.175.215.155.145.075.024.92

We observe that increasing batch size is beneficial for class-conditional OT with FID decreasing from 5.08 at n=256 to 4.92 at n=32,768. However, GT outperforms C$^2$OT at all batch sizes tested. Furthermore, even the smallest batch size GT model (n=256) outperforms the 128x larger class-conditional OT baseline (n=32,768) by a substantial margin (4.15 vs. 4.92).

This shows that GT can be successful at relatively small batch sizes, which increases its practical utility as a large batch requirement is a major compute bottleneck for large-scale generative models and has hampered existing OT-based methods.

Class-conditional OT baseline with soft βm\beta_m

We further run an experiment with soft βm\beta_m as suggested by the reviewer using OT batch size of 256. We highlight however that this isn’t the class-conditional OT baseline as constructed in Cheng & Schwing (2025) as high βm\beta_m is deliberately chosen to construct class-grouped OT transport. Specifically in Section 3.3.2 of Cheng & Schwing 2025, they choose βm\beta_m to be infinite, and in practice large enough, using βm=100\beta_m=100 in their supplied code. Therefore, soft βm\beta_m can be thought of as a combined GT+class-conditional OT transport allowing transport between class-grouped and cross-class samples. We select βm=0.02\beta_m=0.02 based on our sweep of βm\beta_m in Table 8, as at this level the percentage of cross-class samples is ~45-50%.

Table R2 shows the sweep for soft-beta class-conditional OT across CFG scales sampled with Euler (64 steps).

Table R2

FIDCFG=1.0CFG=1.5CFG=2.0CFG=2.5CFG=3.0CFG=3.5CFG=4.0
SiT-B/2+Independent26.446.685.167.9311.0413.6515.76
SiT-B/2+Class-cond. OT26.726.705.087.8110.9013.4715.59
SiT-B/2+GT31.168.134.155.888.7211.3213.41
SiT-B/2+Class-cond. OT(β=0.02\beta=0.02)29.117.424.516.589.5012.1314.21

We observe that soft-β\beta still performs worse than GT.

AuthorsOpenReview ↗

Response to Reviewer LruU (Part 2/2)

Should semi-discrete be best?

The mechanism contradicts the best variant. Lower cost → smaller gap predicts semi-discrete GT should be best, yet it is the worst at its optimum (B/2 FD_DINOv2 199.6 vs 123.6). A smaller gap also means GT partly cancels CFG, so “tolerates larger w” is a guidance rescaling, not extra robustness; at optimal w the gains are 0.15–0.19 FID at XL/2, single seed.

While promising, we found current semi-discrete transport methods numerically challenging. Computation of the semidiscrete transport potential [Mousavi-Hosseini et al. 2026; Kong et al. 2026] requires stochastic optimization and it is possible that in our runs that it did not converge, despite significant effort and close consultation with authors of those works. This suggests that minibatch-OT enjoys clear convergence criteria, may be a better choice in practice, especially for GT. We believe this difficulty explains the relative underperformance of semi-discrete OT in our setting.

Other Clarifications

Is Table 1’s 2.06 your reproduction or the official number? Align training budget and sampler.

We confirm that SiT-XL/2 baseline has been retrained and assessed in the exact same training and inference setting as SiT-XL/2+GT. In both experiments we used 250 Euler steps and guidance intervals of [0.1, 0.7]. We clarify this in the main text (section 5.1).

Does measured prediction gap track performance across variants?

In Table R3 we show a measured prediction gap across the trajectory using the SiT-B/2 model on ImageNet-256. We observe that GT has the lowest prediction gap among three transport plans. We thank the reviewer for raising this question and we include this set of new empirical results in section C.2.1 of the revised manuscript.

Table R3

Couplingn = 256n = 1024n = 2048n = 4096n = 8192n = 32768
Independent5.95–––––
Class-cond. OT5.955.985.976.016.016.02
GT5.625.545.505.49––

SiT-L/2 independent FID 5.50/5.74/2.95 at 16/32/64 NFE is non-monotone — same w sweep?

We confirm the same w sweep was performed for all methods.

Conclusion

We thank the reviewer again for their careful reading and constructive feedback, which helped us significantly strengthen our empirical analysis. Please let us know if we can clarify anything further during the discussion period.

We would like to thank the reviewer for their time and effort in reviewing our work. We are grateful for the reviewer’s valuable feedback that helped strengthen our work and empirical results by investigating the impact of batch size and validating theoretical results. As the rebuttal discussion is coming to end, we would like to confirm whether our responses have effectively addressed raised concerns. We are happy to provide clarifications to any remaining concerns and remain enthusiastic to continue dicussion.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Response

We sincerely thank all three reviewers for their constructive feedback and encouraging evaluations. We are gratified that all reviewers recognized the core strengths of Global Transport (GT): its simplicity, zero inference overhead, and practical scalability [LruU, Phpi], its novel and counter-intuitive finding that condition-agnostic OT improves CFG generation [LruU, Z6Br, Phpi], and its consistent empirical gains across model scales, samplers, and modalities [LruU, Z6Br, Phpi].

To address the shared questions regarding OT batch size scaling, the class conditional OT baseline, and real world validation of our theory, we conducted several new experiments and updated the manuscript. We summarize the changes here and highlight changes in blue in the updated manuscript.

  • OT Batch Sweep [LruU, Z6Br]: To test whether C$^2$OT is bottlenecked by OT batch size, we have added an additional experiment testing C$^2$OT up to a 32k batch size and GT up to an 8k batch size showing that while C$^2$OT slightly improves at large pools, GT outperforms C$^2$OT across every batch size. Notably GT at batch size 256 substantially outperforms C$^2$OT at a 128x larger batch size (4.15 vs. 4.92 FID).
  • Empirical Validation of the CFG Prediction Gap at Scale [LruU, Z6Br, Phpi]: Complementing our finite-batch validation in 2D (Section 4, Fig. 3), we have directly measured the prediction gap on ImageNet-256 across couplings and OT pool sizes. Consistent with Proposition 4.2, GT achieves the lowest prediction gap across all pool sizes consistent with FID improvements. We also clarify the numerical optimization challenges of stochastic semi-discrete OT solvers in our individual responses.

We refer reviewers to the individual responses for detailed discussion results.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.