- Rating
- 3(+1)
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
Summary This paper proposes Contrastive Beckmann Maps (CBM) to address the limited generation quality and training plateau of existing Beckmann Transport Models (BTM). By exploiting the linearity of probability currents, the authors introduce CBM-Transport and CBM-Endpoint to incorporate contrastive guidance during training without learning a velocity field.
Results The method preserves one-step generation (1 NFE) while improving FID from 20.8 to 6.7 on ImageNet 256×256 and from 7.51 to 5.32 on CIFAR-10.
Critical strengths
S1. Theoretically motivated contrastive guidance for BTM\
The paper leverages the linearity of probability currents to derive a contrastive training objective for endpoint-map learning without requiring a velocity field. It further provides theoretical analyses of conservation along the contrastive current and negative-region exclusion under idealized assumptions, supported by a toy experiment.
S2. Substantial empirical improvements with efficient fine-tuning
CBM-T substantially improves FID with only a short fine-tuning period while preserving one-step inference without additional network evaluations. The paper provides controlled comparisons starting from identical BTM checkpoints and a null control, demonstrating the importance of negative supervision. CBM-E also improves training efficiency when learning from scratch.
Critical weaknesses
Major Weaknesses
W1. Insufficient alignment between the problem motivation, proposed method, and empirical evidence
The paper motivates CBM by highlighting class confusion and the FID plateau of BTM, and proposes contrastive training to suppress transport toward incorrect classes. It also suggests high gradient variance as a possible explanation for the plateau. However, the experiments primarily report improvements in FID and classifier entropy without directly establishing whether class confusion is reduced or how it relates to the observed plateau. In particular, lower classifier entropy does not necessarily indicate improved classification into the correct target class.
-
Evaluate target-class accuracy, mislabeling rates, and confusion matrices on ImageNet and CIFAR-10.
-
Analyze the relationship between class confusion and FID improvements across training checkpoints.
-
Provide direct evidence for the proposed high-gradient-variance explanation.
W2. Insufficient understanding of CBM-T optimization stability and degenerate solutions
Appendix Figure 9 and Table 11 show that CBM-T substantially improves FID during a short fine-tuning period but consistently deteriorates with continued training. A similar pattern occurs when CBM-T is applied after CBM-E. More importantly, the toy experiments in Appendix A.2 report near-constant-map collapse when training CBM-T from scratch and divergence at larger guidance strengths. A constant map can trivially eliminate the transport residual, although it does not generally satisfy the positive boundary condition. Proposition 4 characterizes an ideal stationary condition but does not guarantee stable convergence toward it.
-
Track mislabeling rates, precision/recall, and transport or conservation residuals throughout training.
-
Determine whether prolonged training continues to improve the intended negative suppression at the expense of diversity, or instead reflects optimization instability.
-
Analyze sensitivity to guidance strength, CBM-T fine-tuning duration, and simultaneous CBM-E/CBM-T training.
W3. Limited clarity of the BTM-specific contribution relative to prior contrastive generation methods
Signed contrastive target distributions and negative-region exclusion have already been investigated in Signed Rectified Flow, while additive contrastive targets have also been considered in Drifting. The main novelty of CBM therefore appears to lie in adapting these principles to Beckmann probability currents and endpoint-map learning. While technically meaningful, the paper does not sufficiently establish the additional theoretical insights or distinct empirical advantages offered by this Beckmann-specific formulation.
-
Clarify the theoretical contributions beyond Signed Rectified Flow and Drifting.
-
Compare the proposed current-based formulation against simpler negative-pair repulsion or alternative contrastive objectives under matched training conditions.
W4. Limited diversity of experimental settings and robustness of design choices
The experiments are limited to DiT-based class-conditional image generation on CIFAR-10 and ImageNet, and the pretraining results rely on a single seed. Furthermore, several design choices are not sufficiently justified or evaluated for robustness.
-
CBM-E applies its contrastive loss only for (t 0.5), but no ablation is provided to justify this threshold.
-
The CBM-T configurations differ between CIFAR-10 ((), gradient balancing disabled, boundary weight 0.05) and ImageNet XL fine-tuning (, boundary balance 0.25). The rationale for these choices and their sensitivity are insufficiently discussed.
Minor Weaknesses
W1. Citation and grammatical errors
An unresolved citation appears as (?) around Line 89. Several grammatical issues are also present, such as "may due to" and "If switch the loss."
W2. Figure presentation
In Figure 6, the short CBM-T fine-tuning stage following CBM-E training is indicated only by a star marker. Explicitly illustrating the transition and subsequent fine-tuning trajectory would make the experimental setup easier to understand. The alignment and spacing of some figure captions could also be improved.
W3. Reference Error
There appears to be a reference formatting error involving an unmatched } on page 8 (page number) and page 9 (the “Under review as a conference paper” header).
Questions for authors
Q1. BTM-specific formulation: What are the key theoretical or practical advantages of formulating contrastive guidance through Beckmann probability currents, compared with Signed Rectified Flow or other existing contrastive generation methods?
Q2 Generalizability beyond class conditioning: Does the proposed contrastive objective naturally extend to other conditioning modalities, such as text-conditioned generation? Are there any theoretical or practical limitations in defining negative pairs for these settings?
5 comments
Follow-up on our Response
Dear Reviewer hXWN,
Thank you again for reading our paper and providing the feedbacks. We have evaluated target class accuracy, analysed the stability of CBM-T later in training, clarified contribution over drifting and signed reflow. We also trained ImageNet with CIFAR parameters and vice versa to show our result is not sensitive to parameters. Thank you also for suggesting Text-to-Image (T2I), which our methods adapt naturally to. Our T2I is early in training. Would be great if you can take a moment to see if they addresses your concerns, we are happy to clarify further if needed.
Best Authors
Response to Authors
I appreciate the authors’ responses and additional experiments.
My concerns about class confusion (W1), the contribution relative to prior work (W3, Q1), and the dataset-specific CBM-T settings (W4) are resolved. The explanation of the CBM-T degradation (W2) is also partially convincing.
However, I have not yet received an adequate explanation or experimental evidence on the following points:
- Evidence for the gradient-variance explanation of the BTM plateau
- Whether the CBM-T degradation reflects lost diversity or optimization instability, and the near-constant-map collapse
- t ≤ 0.5 threshold in CBM-E (W4)
I would appreciate any further comments on these points.
text-to-image, target-class accuracy, understanding CBM-T degradation, clarification of contribution
Thank you for the review. We address the main concerns below.
Does CBM reduce class confusion? (W1) Yes. Target-label accuracy (%) on the Fig. 4 samples (also in our reply to KfhP):
Top-1 accuracy rises with both classifiers, and confidently wrong samples roughly halve. Across guidance strengths, accuracy and FID peak at the same ω, so the FID gain comes with less class confusion.
CBM-T stability (W2). Why CBM-T drifts when left on, and why we use it as a short dose, is discussed in our reply to Cufo (W1, Q1). In short, the CBM-T gradient norm grows with ω, so it trains with larger gradients than BTM and drifts if left on too long. The recipe of CBM-E followed by a CBM-T dose pretrains stably from scratch at XL scale and reaches FID 11.2 at about 125 epochs.
Relation to Signed Rectified Flow and Drifting (W3, Q1). Both apply the contrast to something BTM does not have. Signed Rectified Flow contrasts a time-dependent velocity field and samples by integrating it, so it does not learn a one-step map. Drifting repels in the extra feature space of a pretrained encoder with batches of 4,096 to 8,192. BTM has neither velocity nor extra feature space , which is why guidance for it was left open in the BTM paper. Our observation is that the probability current is linear in the target, but velocity is not linear in the target for autonomous flows that BTM builds on. So the contrast should be applied to the current while still training a one-step map. Contrary to drifting, we used no extra feature encoder; our batch size can be as few as 4 per gradient (batch ablation in our reply to Q1of KfhP) contrary to drifting which needs at least 4096 batch size. On simpler negative-pair repulsion: CBM-E (Sec. 3.3) is exactly that, pushing T(x∣c) away from T(x∣c−), and Fig. 9 and Table 11 compare it with CBM-T from the same checkpoints.
Sensitivity of the CBM-T settings (W4). The CIFAR-10 and ImageNet settings differ mainly in the guidance strength ω, which we chose for each dataset by a sweep: on CIFAR-10 ω=0.5,1,2,4 give FID 6.59, 6.03, 5.66, 5.76. The gradient balancer and the boundary weight matter much less. To show this, we fine-tuned ImageNet XL/2 with the CIFAR-10 setting (Table A). We also finetuned CIFAR-10 model with ImageNet settings as in Table B. Swapping the parameters between ImageNet and CIFAR doesn't matter. Guidance strength ω does matter. But similar to the CFG literatures, it is normal that ω is tunned separately for each task.
Table A. DiT-XL/2, CBM-T fine-tune of the BTM checkpoint, FID after 6,000 updates (one seed).
Table B. CIFAR-10, CBM-T fine-tune of the BTM checkpoint (FID 7.51), FID after 2,500 updates.
Text conditioning (Q2). The construction carries over directly: the negative for a caption is an image from the batch paired with a different caption. The analysis does not require negatives to be disjoint from the positive. Mismatched captions approximate the unconditional distribution, which is what CFG subtracts with an empty prompt, and occasional near-matches only weaken the contrast. We have started a text-to-image run but it is a very large dataset and limited by compute and rebuttal time it is early in training. At epoch 4 it reaches GenEval 0.06, which is far too early to draw conclusions, but the setup trains end to end and we will report the finished run.
Thanks for flagging typos and the unresolved citations, they will be fixed in the revision.