- Rating
- 2
- Confidence
- 2
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 3
Read review
Paper summary
This paper integrates the existing FlashSinkhorn solver into the flow matching training pipeline, making larger-scale online OT pairing computationally feasible. It theoretically analyzes how the coupling and candidate-set size affect gradient variance and integration error. The authors compare standard OT-CFM with EOT-CFM using larger pairing candidate sets on synthetic data, single-cell data, and image-generation tasks.
Critical strengths
-
The paper provides an implementation that can be integrated into existing PyTorch training pipelines. By incorporating FlashSinkhorn, it makes larger-scale online OT pairing feasible with modest additional overhead and improves OT-CFM training and generation performance on the evaluated tasks.
-
The paper further characterizes the theoretical relationship between OT pairing, velocity-target uncertainty, and stochastic-gradient variance, helping clarify how different pairing strategies affect gradient variance under ideal fitting conditions.
Critical weaknesses
-
The methodological novelty is limited. The pairing objective, entropic regularization, and FlashSinkhorn solver used in this paper all come from prior work. The main methodological addition is their integration into an online training pipeline. The distinctive value of this integration relative to existing pairing methods needs to be established more clearly.
-
The additional theoretical insights and their practical guidance are limited. The paper further explains properties of existing OT pairing methods, but its main conclusions rely on ideal fitting conditions. It does not sufficiently explain how these results can guide finite training processes, approximate coupling computation, or the choice of candidate-set size and regularization strength. Their implications for practical training and subsequent method development therefore remain unclear.
-
The experimental coverage and comparisons are insufficient. The image-generation experiments cover only CIFAR and ImageNet-64, without validation at higher resolutions or with larger modern generative models, limiting the evidence for broader scalability. Although the paper reports some computational overhead, it lacks comparisons with other relevant pairing methods under matched conditions in terms of both quality and total training cost. Its relative advantages therefore remain insufficiently established.
Questions for authors
-
What new insights does the theoretical analysis provide beyond existing understanding? Do these results change our understanding of the applicability or relative merits of different pairing methods, or guide the choice of pairing method, candidate-set size, and regularization strength in practical training? In particular, which conclusions remain applicable before ideal fitting is achieved?
-
Could the authors compare the training computational costs and final generation quality of the proposed approach with other relevant pairing methods under matched conditions? Where applicable, the comparison should include the total costs of precomputation, pairing, and model training. Quality under the same total computational budget, or the computational cost required to reach the same quality, would help clarify the practical advantages of the approach.
-
Are there results on larger models, higher resolutions, or more complex datasets that support the paper’s scalability claims? If such results are unavailable, could the authors clarify the scope of the conclusions supported by the current experiments and the potential limitations when extending the approach to these settings?
6 comments
Author response (2/2)
W3 & Q3: experimental coverage is insufficient
We add text-conditioned flow training experiments to evaluate our method on a more complex task relevant to modern generative AI applications. We trained on CUB task, providing both quantitative FID results below and qualitative examples in Figure 4.
| NFE 25 | NFE 100 | Adaptive | |
|---|---|---|---|
| OT-CFM-128 | 11.04 | 10.61 | 10.27 |
| EOT-CFM-4096 | 10.88 | 10.44 | 10.12 |
| EOT-CFM-8192 | 10.49 | 10.20 | 9.94 |
Beyond generation quality, we find that larger minibatch context also improves text alignment. For example, in the third panel of Figure 4, the prompt is "bird with a long beak, blue eyes, and dark gray and black feathers". However, both baselines incorrectly assign blue to the feathers, while ours with a larger coupling minibatch follows the proper feather color, which shows that larger coupling minibatch may provide a more consistent conditional training signal that helps associate textual attributes with their intended visual components.
Q2: matched condition's training cost and quality
We apologize for missing this point. On CIFAR-100, OT-CFM takes 12.2 h, our largest EOT-CFM-8192 takes 14.3 h, and EOT-CFM-scheduled takes 12.5 h. Shown in Figure 5, we also train those methods for additional steps to extend their training time to 14 h, and our larger minibatch one still performs better under both same training time and same training steps. We note that the most of the additional overhead comes from coupling computation, so with larger models and more complex text-conditioned task, this overhead will account for a smaller fraction of total training time as flow optimization processes takes most of the training time. We have added these to the updated manuscript.
Again, many thanks for your careful reading and thoughtful questions. We believe your valuable feedback has helped us further strengthen the significance of our contributions, particularly by better connecting our theoretical results to their practical implications. We hope that our responses have addressed your concerns, and we look forward to an open-minded discussion if any such concerns remain.
Sincerely,
Authors of Paper 187
Hi reviewer dJhy,
As the rebuttal period comes to the end, thanks for your time to check our response to see if there's any further questions or concerns.
Author response (1/2)
We sincerely thank you for providing these detailed and insightful points. We address each of them below:
W1, W2 & Q1: theory's practical guidance to method
We believe the key significance lies in how the relationship between minibatch size, gradient variance, and generation error is mathematically established. Taking gradient variance as an example, as explained in line 280–283, our key contribution is to connect minibatch size to velocity uncertainty through the transport-cost residual, and then connect this uncertainty to gradient variance.
Zoom into Theorem 4.15, the gradient variance is bounded by R (velocity uncertainty) and essential supremum J-norm (the largest squared model-Jacobian norm). Specifically, Jacobian norm is very large at the beginning and decreases as training progresses (see Figure 6 in Appendix B). By Propositions 4.12 and 4.14, a larger coupling minibatch can better control velocity uncertainty. This also empirically aligns with Figure 2, where the variance reduction gain is most significant during the first quarter of training. Together, these findings motivate using a large minibatch initially to suppress the uncertainty amplified by a large Jacobian, then switching to smaller minibatches as the Jacobian norm decreases.
We address this through an adaptive scheduling strategy (EOT-CFM-scheduled). Taking an example from CIFAR, it starts with a coupling minibatch of 8192 to first and gradually reduces it to 2048. On CIFAR-100, it achieves 7.29 FID, comparable to 7.25 with EOT-CFM-8192, while requiring only 12.5 hours of training, close to the 12.2 hours of OT-CFM with 7.89 FID. On ImageNet, OT-CFM takes 55.2 h with 13.26 FID, EOT-CFM-scheduled takes 56 h with 12.96 FID, and EOT-CFM-10K takes 58.8 h with 12.94 FID. Further training-time comparisons are provided in our response to Q2.
We also add text-to-image generation experiments to demonstrate applicability to conditional generation (further details in our response to W3 and Q3). On text-conditioned CUB, EOT-CFM-scheduled achieves 9.98 FID in 18.5 hours, compared with 9.94 FID in 20.2 hours for EOT-CFM-8192 and 10.27 FID in 18.1 hours for OT-CFM.
We believe these results demonstrate a concrete benefit of scalable online coupling, which provides the flexibility to adapt the coupling context in ways that previous coupling methods could not achieve, addressing W1 and Q1 regarding your concern on "the distinctive value of this integration relative to existing pairing methods".
We first clarify that optimality is a common setup in FM analysis, and we have expanded the discussion of prior work in lines 203–208. Optimality removes the variance contribution from imperfect fitting of the conditional-mean velocity, providing a common basis for comparing gradient variance across different couplings.
Without ideal fitting, the corresponding bound in Theorem 4.15 only includes an additional fixed-round fitting error term, and this uncertainty-amplification mechanism still remains applicable. We provide the derivation and discussion in Appendix B.