acceptodds
Under review as a conference paper at ICLR 2027

Contrastive Beckmann Maps Improve One-step Generation

Abstract

Beckmann Transport Models (BTM) recently opened a promising avenue for learning one-step generative maps end-to-end, bypassing the need for distillation, intermediate targets, or external features. However, its quality has so far lagged behind state-of-the-art methods, which largely depend on condition strengthening techniques, such as classifier-free guidance for flows and diffusions: guidance combines velocity or score predictions, and a native map has neither. We address this limitation by introducing Contrastive Beckmann Maps (CBM). We observe that the object defining a BTM, its probability current, is linear in the target distribution, so guidance can be applied to the current itself. CBM subtracts the current toward mismatched image–label pairs, obtained by pairing each label with another image in the batch. Under the assumptions of our analysis, the ideal map avoids the negative region of the contrastive target, while sampling still requires one network call. Contrary to CFG, CBM is a theoretically grounded training technique tailored to the BTM framework, repelling correct samples from negatively drawn samples during the training phase. On ImageNet , fine-tuning a converged BTM model with CBM for around 15% of an epoch lowers one-step FID from 20.8 to 6.7 and 5.5 with postprocessing. Trained from scratch, CBM reaches the same FID of an optimized BTM with fewer epochs. On CIFAR-10, CBM improves one-step FID from 7.5 to 5.3.

Primary area:
generative models
Keywords:
One-step Generation, Distribution Reweighting, Time-independent Flow

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.67 over 3 reviews, above 99% of the others.
Reviewer_hXWNOpenReview ↗
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 ≤\leq 0.5), but no ablation is provided to justify this threshold.

  • The CBM-T configurations differ between CIFAR-10 ((ω=2\omega=2), gradient balancing disabled, boundary weight 0.05) and ImageNet XL fine-tuning ((ω=1)(\omega=1), 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
AuthorsOpenReview ↗

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):

ConvNeXt-L top-1ConvNeXt-L top-4CLIP top-1CLIP top-4confident and wrong
BTM57.676.737.659.016.2
CBM82.094.061.782.18.7

Top-1 accuracy rises with both classifiers, and confidently wrong samples roughly halve. Across guidance strengths, accuracy and FID peak at the same ω\omega, 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\omega, 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)T(x\mid c) away from T(x∣c−)T(x\mid 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\omega, which we chose for each dataset by a sweep: on CIFAR-10 ω=0.5,1,2,4\omega = 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\omega does matter. But similar to the CFG literatures, it is normal that ω\omega 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).

ω\omegaGradient balancer and boundary weightFID
2ImageNet setting (balancer on, boundary balance 0.25)9.10
2CIFAR-10 setting (balancer off, 5× lower boundary weight)9.08
1Balancer off, boundary weight fixed at the balancer's average7.07
1CIFAR-10 setting (balancer off, 5× lower boundary weight)7.34

Table B. CIFAR-10, CBM-T fine-tune of the BTM checkpoint (FID 7.51), FID after 2,500 updates.

ω\omegaGradient balancer and boundary weightFID
1Balancer on, as on ImageNet6.03
2Balancer on, as on ImageNet5.66
2Balancer off, boundary weight 0.255.47
2CIFAR-10 setting (balancer off, boundary weight 0.05)5.34

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.

AuthorsOpenReview ↗

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

Reviewer_hXWNOpenReview ↗

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:

  1. Evidence for the gradient-variance explanation of the BTM plateau
  2. Whether the CBM-T degradation reflects lost diversity or optimization instability, and the near-constant-map collapse
  3. t ≤ 0.5 threshold in CBM-E (W4)

I would appreciate any further comments on these points.

Reviewer_CufoOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In this work the authors propose Contrastive Beckmann Maps, a way to add guidance to Beckmann Transport Models. The need for guidance comes from BTMs tendency to mix up neighbouring classes and to plateau at a high FID since CFG is not possible in the BTM formulation. Since the probability current is linear in the data distribution, the authors guide it by subtracting the current toward mismatched data–label pairs, obtained by shuffling labels within the batch. Experiments show improved FID on Imagenet compared to the vanilla variant.

Critical strengths
  • Simple and well motivated

The idea is simple and well motivated with clear gains on every experiment. The observation that currents add where velocities dont is clean, and it addresses an open problem the BTM paper explicitly left unsolved.

  • Cheap and easy to adopt

Sampling stays at one NFE, with no teacher, reward or feature encoder. The method plugs into existing BTM checkpoints. Furthermore, it behaves like one would expect CFG to do.

Critical weaknesses
  • Theory gain gap

CBM-T, the variant the theory is about, only helps in a fine-tuning manner. Training longer with it makes results worse, and from scratch its unstable, which is odd if its fixed point is supposed to be the right map. The from-scratch speedup comes from CBM-E, which has no theory behind it at all. Id tone down the theoretically grounded claim.

  • No adaptable guidance scale

The guidance strength is fixed during training time. w is baked into the weights, and so changing the precision/recall trade-off means fine-tuning again when a different level is needed. With CFG, the scale can be picked at sampling time. iMF and Drifting keep that flexibility by conditioning the network on the guidance strength.

minor:

  • Overlapping novelty

The additive contrastive target is shared with Drifting models which have a similar mechanism.

Questions for authors
  • Why does CBM-T degrade when trained longer, if its fixed point is the desired contrastive map? Do you have an explanation, for example optimization instability, or the non-uniform time schedule changing the effective target?

-Is there a way to condition on guidance scale as iMF and Drifting, so the guidance strength can be chosen at sampling time? Would a single scale conditioned model match the scale fine-tunes in Table 1?

2 comments
AuthorsOpenReview ↗

Training guidance scale and explain degradation

Thank you for carefully reading our paper and for coming with a helpful review. The guidance-scale question led to our best new result, and we explain the degradation of CBM-T as due to an increase in gradient norm due to guidance ω\omega.

Choosing ω\omega at sampling time (W2, Q2) works , and a single conditioned model beats the separate fine-tunes. We condition the map on ω\omega as iMF does, fine-tune once from the BTM checkpoint using CBM-T algorithm with ω∼U[0,2]\omega \sim U[0,2], and pick ww at sampling time. Sampling is still one network call. After 6,000 updates (FID, precision / recall):

wwone conditioned modelseparate fine-tune at ω=w\omega=w
0.56.84, 70.9 / 47.58.96, 63.1 / 55.5
16.13, 74.6 / 43.06.69, 70.5 / 48.1
1.56.79, 75.9 / 39.37.36, 71.9 / 43.0
27.83, 74.7 / 36.38.39, 69.5 / 42.0

The conditioned model has lower FID at every ww. The same holds when pretraining DiT-B/2 from scratch with ω\omega-conditioning and CBM-E algorithm, against the paper's fixed-strength CBM run (FID):

epochpapertrained conditioning, eval at w=1w=1trained conditioning, eval at w=1.5w=1.5
577.972.871.8
1065.260.858.9
2052.448.347.3
4045.441.841.7

So the flexibility comes for free, and both experiments will go into the paper main text.

Theory vs. CBM-T in practice (W1, Q1). We will drop "theoretically grounded" and say what the analysis covers: Props. 4 and 5 characterize the ideal fixed point of CBM-T and what that map samples. CBM-E will be presented as the endpoint-space counterpart, backed by experiments.

Degradation of CBM-T: Proposition 4 characterises the fixed point of the stop-gradient update, but does not guarantee convergence to the fixed. The sign-flipped negative term keeps acting on a noisy estimate of the negative current from mismatched pairs, so training can overshoot the fixed point. This matches the data: the degradation starts earlier and grows faster with as shown in the table below:

2000 updates4000 updates6000 updates
17.106.846.74
48.0310.2312.63

The numbers are FID-50k of CBM-T fine-tunes started from the same XL checkpoint.

From the same XL checkpoint, FID still improves at =1 after 6000 updates but rises from 2000 updates on at =4. The overshoot of the fixedpoint is consisted with the magnitude of the gradient increases with guidance strength ω\omega as shown below:

earlylate
00.720.71
11.863.05
46.0215.24

From the CBM-T loss, one can see the gradient roughly scales with guidance strength ω\omega. At high ω\omega, not only is the FID bad but also the deterioration starts early due to a large gradient magnitude and overshooting. In the toy example, the exact flow reaches the target at every , while training does not settle at the fixed point when =1.5 from BTM and diverges at =3. The time schedule only reweights the conserved current (Remark 7) and leaves the fixed point unchanged. The time schedule does not explain the degradation that depends on . We therefore use CBM-T as a short dose. We found that a CBM-T dosage after CBM-E pretraining scales. DiT-XL/2 trained from scratch this way reaches FID 11.2 at about 120 epochs, against 17.6 for BTM after 1,280 (see our reply to KfhP).

Overlap with Drifting. The additive target is shared. Drifting contrasts in a pretrained 2.1B MAE encoder feature space and trains with batches of 4,096 to 8,192. CBM contrasts probability currents in data space with no encoder and works with 4 images instead of 4096 per update. A batch size of 4 for CBM-T gives FID 6.55 on the XL fine-tune and larger batches are worse, see our reply to KfhP Question 1. Contrasting the current is also what gives the exclusion proposition as in Prop. 5.

Reviewer_CufoOpenReview ↗

Concerns resolved

Thank you for addressing my concerns and for adding the additional experiment in the short time. I am glad the guidance conditioning worked. I will raise my initial rating.

Reviewer_KfhPOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Contrastive Beckmann Maps (CBM), which adds training-time guidance to one-step Beckmann Transport Models using mismatched image–label pairs as negatives. The authors introduce two losses: a contrastive transport loss for fine-tuning and an endpoint contrastive loss for training from scratch. They also show, under idealized assumptions, that the transport map avoids regions where the contrastive target measure is negative. Experiments on ImageNet-256 and CIFAR-10 show improved FID and faster convergence. On ImageNet-256, fine-tuning reduces BTM’s FID from 20.8 to 6.7, or 5.5 with additional postprocessing, though recall decreases as guidance becomes stronger.

Critical strengths
  1. CBM substantially improves an existing BTM with only a short fine-tuning stage, reducing ImageNet-256 FID from 20.8 to 6.7 while retaining one-step sampling. The continued-training and matched-pair controls help establish that the improvement comes from the contrastive signal.

  2. The method introduces guidance through probability currents without requiring a teacher or an additional feature encoder. The analysis connects the proposed update to a contrastive current and, under stated assumptions, shows that the resulting map avoids negative regions of the target measure.

Critical weaknesses
  1. The contrastive guidance idea is closely related to Drifting Models. Drifting uses large batches of 4,096 or 8,192, which can limit its practicality. CBM uses batch 256 for training from scratch and a much smaller batch for XL fine-tuning, but its sensitivity to batch size is unclear. An ablation with batch sizes 256, 1,024, and 4,096, under comparable training budgets, would help establish whether CBM works reliably with small batches. This could be a meaningful practical advantage over Drifting.

  2. The from-scratch experiments are limited to DiT-B/2 and fewer than 200 epochs in Fig. 6. The XL results instead rely on fine-tuning a pretrained BTM, leaving the performance of CBM trained from scratch at larger scale unclear. I understand that compute may be limited, but an XL experiment with a longer schedule, such as 720 epochs, would strengthen the evaluation. Reporting total training compute would also make the comparison with iMF and Drifting in Table 3 more informative.

  3. Negative samples are selected without considering class similarity. Since reducing class confusion is a central motivation, it would be useful to compare random negatives with negatives from similar classes—for example, using a cat rather than a car when the target is a dog. This would help clarify whether the current sampling strategy makes effective use of the contrastive signal.

Questions for authors

In addition to the ablations requested above, could the authors report target-label top-1 and top-4 classification accuracy for the samples in Fig. 4? Since the model is class-conditional, lower classifier entropy alone does not establish better condition adherence: the classifier could be more confident in the wrong class. These metrics would directly test whether CBM reduces class confusion.

3 comments
AuthorsOpenReview ↗

batch size ablation, XL from scratch, and nearby classes as negatives

Thanks for the constructive suggestions, all of them are very helpful and we ran all of them.

Batch size (W1). CBM works better with small batches instead of large batches. Large batches are not needed for CBM as they do for drifting. XL fine-tune from Table 1 (ω=1\omega=1):

images / updateupdatesimages seenFID
416,00064k6.55
32 (paper)3,00096k6.67
256750192k6.28
1,024192197k7.45
4,09648197k10.34

Four images per update already match the paper's FID with fewer images seen. At the same image budget, larger batch size gives worse FID. A CBM negative is just another image in the batch paired with the wrong label, and the loss is per-sample, so batch size only affects gradient noise. Drifting estimates its field over many samples at once and needs batches of 4,096 to 8,192. This is a practical advantage of CBM and we will highlight it in the paper.

XL from scratch (W2). DiT-XL/2 trained from scratch with CBM reaches FID 24.1 at about 125 epochs same as size B/2, and 11.2 after a short CBM-Transport dose. The XL pretraining FID curve can be found in this link https://anonymous.4open.science/ is constantly below B/2 and BTM reports 17.6 after 1,280 epochs, so CBM beats BTM with roughly 10x fewer epochs. Limited by computing speed, we will keep the run going toward 720 epochs for the final version and add epochs and training compute to Table 3.

Similar-class negatives (W3). Using negatives from only similar classes hurts. We fine-tuned from the BTM checkpoint with ω=1\omega=1 and CBM-T algorithm and drawing each label's negative only from its nearest classes:

negativesFIDprecisionrecall
10 nearest classes9.8160.558.1
100 nearest classes7.2766.652.9
random (paper)6.6970.548.1

Pretraining B/2 from scratch with the CBM-E algorithm with negative samples only from nearby classes is consistently worse than using randomly labelling the images.

epochrandom(paper)10 nearest classes as negatives
1065.279.7
2052.461.6
4045.452.1
5044.650.1

Nearby classes look alike at high noise, where the class is decided, so the current from nearby classes almost cancels the positive one and the guidance gets weaker as we observe precision decreases and recall going up. The precision and recall trend for training with nearest negatives is similar to lower CFG guidance strength. Random negatives approximate the unconditional distribution, which is exactly what CFG subtracts since class confusion may still confuse far apart classes.

Target-label accuracy (%) on the samples of Fig. 4. Confident and wrong is defined as the classifier assigns probability 0.5 to a wrong class.

ConvNeXt-L top-1ConvNeXt-L top-4CLIP top-1CLIP top-4confident and wrong
BTM57.676.737.659.016.2
CBM82.094.061.782.18.7

CBM is sharper because it puts mass on the correct class, not on a wrong class.

AuthorsOpenReview ↗

Follow up on our response

Dear Reviewer KfhP,

Thank you again for the constructive suggestions. All three suggestions were helpful and we ran all of them. On batch size, we ablated 4, 32, 256, 1,024, and 4,096 images per update at comparable budgets: CBM works better with small batches, four images per update already matches the paper’s FID with fewer images seen, and larger batches are worse at the same image budget. This is a practical advantage over Drifting, which needs 4k-8k. On scale, DiT-XL/2 trained from scratch with CBM reaches FID 24.1 at about 125 epochs and 11.2 after a short CBM-Transport dose, with its pretraining curve consistently below B/2 — roughly 10× fewer epochs than BTM’s 17.6 at 1,280 epochs, and we will keep the run going toward 720 epochs and add epochs and training compute to Table 3 for the final version. On negative sampling, drawing negatives only from nearby classes hurts in both the fine-tuning and from-scratch settings. We attribute this to the contrastive current from nearby classes nearly cancelling the positive one at high noise, where the class is decided, which shows up as lower precision and higher recall, much like weaker CFG.

We also added the target-label accuracy you asked for on the Fig. 4 samples: CBM raises ConvNeXt-L top-1 from 58 to 82 and top-4 from 77 to 94, and reduces confident-and-wrong predictions from 16% to 9%. So the lower classifier entropy does reflect mass on the correct class rather than confidence in a wrong one.

It would be great if you can take a minute to see if these address your concerns. We are happy to clarify further or run additional ablations if needed.

Best, Authors

The author addresses my concerns about batch size and negative class. It is surprising to see that the optimal batch size is 256, which is very small even for Sit-XL training (batch size 1024). I do agree that it is a big advantage that CBM only requires small batch size 256 rather than big batch size 8,192 in drifiting. But the FID is still significantly higher than drifiting (1.54).

I think the authors did a good job in the rebuttal, therefore i increase my score to clear acceptance.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.