acceptodds
Under review as a conference paper at ICLR 2027

Mixed-Field Matching: Time-Conditioned Transport with Energy-Based Refinement

Abstract

We introduce Mixed-Field Matching (MFM), which trains one network whose time input follows the noise level up to a threshold and stays fixed above it. Diffusion and flow-matching models give the network the noise level at every step of generation, while recent equilibrium models remove it entirely. We argue that the need for the noise level changes during generation. Far from the data, one noisy sample is consistent with many noise levels and many clean images, and the noise level reduces this uncertainty about where the sample should move. Close to the data, the manifold hypothesis suggests that the sample itself carries this information. MFM sampling first transports noise toward the data with the time-conditioned field and then refines each sample using a single time-independent field. A penalty that sets the field to zero on clean training examples keeps finished samples in place, so extra refinement steps do not degrade them. On a two-dimensional example where equilibrium matching (EqM) moves every sample onto the data but gives the modes the wrong proportions, MFM assigns samples to modes about as accurately as flow matching (FM). On ImageNet-256 with DiT-XL/2, at the same training and sampling cost as the baselines, MFM lowers FID by 15% relative to EqM and 28% relative to FM without guidance, and by 14% and 33% at a shared classifier-free guidance scale.

Primary area:
generative models
Keywords:
flow matching, diffusion models, generative models, equilibrium matching, energy-based models, noise conditioning, sampling, image generation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_2c3LOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Mixed-Field Matching (MFM), a generative method that combines time-conditioned transport with time-independent refinement in a single network. For interpolation times below a threshold (a), the network receives the true time and predicts a flow-matching velocity. Above (a), the time input is fixed, while the target magnitude decreases linearly to zero at the clean endpoint. Sampling first integrates the time-conditioned field and then repeatedly applies the fixed refinement field. The authors additionally introduce an endpoint penalty that encourages zero predictions on clean samples and an in-batch estimator that applies this penalty without additional network evaluations. The paper analyzes the regression cost of fixing the time input and gives a local convergence result for refinement. Experiments cover ImageNet-256, CIFAR-10, two-dimensional distributions, inpainting, and field composition, with comparisons primarily against flow matching and equilibrium matching.

Critical strengths
  1. The method is simple, intuitive, and uses a single network.

  2. The endpoint penalty provides useful stability under additional refinement steps.

  3. The paper is clearly written and includes extensive implementation details and several metrics.

Critical weaknesses
  1. The central mechanism is not isolated. MFM differs from FM in its time input, target decay, target multiplier, endpoint penalty, and sampler. It also differs from EqM in both time conditioning and sampling. A particularly important confound is that MFM uses λ=4\lambda=4, whereas FM uses an unscaled target. Appendix C shows that changing MFM from λ=4\lambda=4 to λ=1\lambda=1 worsens ImageNet FID by more than five points, suggesting that target scaling alone may explain a substantial part of the improvement.

  2. A key baseline is missing on ImageNet. [1] showed that removing noise conditioning can improve flow matching on CIFAR-10, reducing FID from 3.01 to 2.61 with Euler sampling. Consistently, MFM’s noise-unconditional FM baseline achieves 2.760 FID on CIFAR-10, close to MFM’s 2.744. Therefore, a matched ImageNet comparison is necessary to establish whether retaining time conditioning during transport is actually beneficial. EqM is not a sufficient substitute because it uses a different target and sampler.

  3. The empirical comparison is not fully controlled. FM uses an SDE sampler, EqM uses gradient descent, and MFM uses Euler transport followed by fixed-point refinement. The main DiT-XL/2 results use one training run per method. At the best tested guidance scales, MFM improves over EqM only from 3.736 to 3.532 FID, while the B/2 results indicate that guided differences can be comparable to run-to-run variation.

[1] Q. Sun, Z. Jiang, H. Zhao, and K. He. “Is Noise Conditioning Necessary for Denoising Generative Models?” Proceedings of the 42nd International Conference on Machine Learning (ICML), PMLR 267:57469–57502, 2025.

Questions for authors
  1. What is the performance of FM trained with (\lambda=4), with the output rescaled during sampling as done for MFM?

  2. Can you report fully noise-unconditional FM on ImageNet using the same setting?

  3. Since the learned field is non-conservative, in what precise sense is the refinement “energy-based,” and what justifies adding such fields for composition?

4 comments
AuthorsOpenReview ↗

Response to 2c3L

Thank you for your detailed review and for the specific controls you proposed. We ran all of them, and we believe the paper is stronger as a result. Our response has two parts. This part covers the new controls (W1, Q1, Q2). The second part covers the samplers, the repeated runs and the energy interpretation (W3, Q3).

W1. "The central mechanism is not isolated ... target scaling alone may explain a substantial part of the improvement."

Thank you for raising this important point. Following your comment, we trained models that isolate each component and added them to the revised paper (new Table 14). On DiT-B/2, FM, EqM and MFM are trained with the same multiplier λ = 4, the same target, the same training cost and no endpoint penalty, and are sampled with the same Euler update and 250 function evaluations. FM has an FID of 32.40, EqM 32.62 and MFM 29.87. FM is trained with the official EqM code, which reproduces our FM with λ = 1 (37.23 compared with 37.45).

As you anticipated, target scaling lowers FM's FID, from 37.45 to 33.99 with FM's own target scaled by 4, and to 32.40 with MFM's target. At the same λ, MFM's FID is a further 2.5 lower, and MFM with the endpoint penalty (31.32) is also lower than both. Together, these results give the matched comparison you asked for, in which MFM has a lower FID than both FM and EqM.

Q1. "What is the performance of FM trained with λ = 4, with the output rescaled during sampling as done for MFM?"

Thank you for suggesting this experiment. We trained FM with λ = 4 in two ways (new Table 14). With FM's own target scaled by 4 and the output divided by 4 during sampling, FM has an FID of 33.99 without guidance and 14.59 at guidance scale 1.5. With MFM's target it has 32.40 without guidance. MFM has 29.87 and 12.37 at the same training cost and with the same Euler update. Wang & Du report 41.89 for the second model (their Table 4). We obtain 42.92 when the output is divided by λ alone, and 32.40 when it is divided by the full target scale, which recovers the velocity at every time. We therefore report the lower FM number (caption of Table 14).

Q2 / W2. "Can you report fully noise-unconditional FM on ImageNet using the same setting?"

Yes, and thank you for the reference to Sun et al. Following your request, we added noise-unconditional FM to the revised paper as a row of the new Table 14. It has an FID of 35.73 with λ = 4, trained with the official EqM code, and MFM has 29.87. Sun et al. report the same comparison on ImageNet-256 with SiT-B/2 (their Figure 8). Without guidance, removing the time input raises FID from about 37 to about 41 there, in agreement with Wang & Du (36.68 to 40.81, their Table 4). Table 14 also gives a matched test of the time input during transport, since EqM and MFM differ only in the time input below the threshold. EqM has 32.62 and MFM 29.87, so retaining the time input during transport lowers FID on ImageNet-256.

This is the second part of our response. It covers the samplers, the repeated runs and the energy interpretation.

W3. "FM uses an SDE sampler, EqM uses gradient descent, and MFM uses Euler transport followed by fixed-point refinement. The main DiT-XL/2 results use one training run per method. At the best tested guidance scales, MFM improves over EqM only from 3.736 to 3.532 FID."

Thank you for these three points. We address them in turn.

  • Sampler. In the new Table 14, every method uses the same Euler update and 250 function evaluations. EqM's gradient-descent step is an Euler step for a time-independent field (Wang & Du, Appendix D.1), the same update as MFM's refinement. In the DiT-XL/2 table, FM keeps its SDE sampler, which gives it a lower FID than Euler steps (35.90 compared with 37.45 on DiT-B/2).
  • Number of runs. As you suggested, we trained a second run of every method (Table 6). Each MFM run has a lower FID than each EqM and FM run, at both model sizes, with and without guidance. Averaged over the two DiT-XL/2 runs, MFM's FID is 11% lower than EqM's without guidance and 10% lower at guidance scale 1.5 (3.78 compared with 4.21).
  • Best guidance scales. The two numbers you cite are each method's best tested scale after 80 epochs, from one run each, and MFM's FID is the lower one there as well. With longer training, MFM-XL/2 achieves an FID of 1.94 after 700 epochs. This is lower than DiT-XL/2 and SiT-XL/2 after 1400 epochs (2.27 and 2.06) and within 0.01 of EqM-XL/2 after 1400 epochs with the same Euler update. MFM therefore reaches this FID with half the training (Table 1, Figure 4).

Q3. "Since the learned field is non-conservative, in what precise sense is the refinement energy-based, and what justifies adding such fields for composition?"

Thank you for this question, which helped us state the claim precisely. Following it, we changed the title from "Energy-Based Refinement" to "Implicit Energy-Based Refinement", and a new paragraph in Appendix A.3 states the sense we mean, which is that of EqM's implicit energy-based models.

Above the threshold, MFM's field is time-independent and is trained on EqM's target. For a noisy sample z and its clean example x1, this target is a positive constant times (x1 − z), which is minus the gradient of a quadratic energy centred at x1. Near the data, the ideal field points to the nearest point of the data manifold, and it is minus the gradient of the squared distance to the manifold, scaled by a constant. As in EqM, the energy is implicit, and the network is trained on these targets without an architectural constraint. As you note, the trained field is therefore not exactly conservative, and Appendix C.4 measures this for both MFM and EqM. For this reason, our convergence result (Proposition 2) does not assume a conservative field. It treats refinement as a fixed-point iteration and holds for any field.

Composition has the same justification as in EqM (their Section 5.4), where adding the fields of two models corresponds to adding their implicit energies. On CIFAR-10, the summed MFM fields have the lowest FID on the shared classes (Table 3).

Thank you again for your suggestions, which have improved the paper. We hope that the new controls and clarifications address your concerns. If they do, we would be grateful if you would consider updating your score, and we are happy to answer any further questions.

Thank you for providing the additional experiments, which have strengthened the empirical evidence. However, my concerns regarding the proposed mechanism have only been partially addressed.

While the theory explains the conditions under which ignoring temporal information results in only a minor increase in regression error, it does not specify how to determine the precise switching point. Instead, a fixed threshold is chosen empirically based on generation quality and applied uniformly within each setting. Although this is a reasonable empirical choice, the associated costs would become prohibitively high in the context of large-scale generative training.

I would rate this submission at 2.5, it does not quite meet the threshold for a score of 3.

Thank you for reading our response, and for noting that the additional experiments strengthen the empirical evidence. Your further question on the threshold is a practical one, and we are glad to answer it.

"... it does not specify how to determine the precise switching point. ... the associated costs would become prohibitively high in the context of large-scale generative training."

As you note, the analysis gives the conditions under which a fixed time input costs little, and the value itself is selected by sample quality. In our experiments, this selection was inexpensive.

  • One selection, on the small model. We compared three thresholds (0.7, 0.8 and 0.9) on DiT-B/2 after 80 epochs, which required two additional DiT-B/2 runs. In this comparison, FID is not very sensitive to the exact threshold. Without guidance it differs by 0.5 between 0.7 and 0.8, and at the best tested guidance scale all three thresholds are within 0.23 (6.92 to 7.15, Table 9).

  • No selection for the large models. The selected value, a = 0.8, was used unchanged for DiT-L/2, for DiT-XL/2 with five times as many parameters, and for 700 epochs. The main results of the paper (Table 1, Figures 4 and 5) therefore use a threshold that was not tuned at those scales, and with it MFM has the lowest FID at both model sizes (Table 15).

  • The same scalar as in EqM. The threshold is the cutoff of the training target, which EqM's target also has, and MFM uses the same scalar for the time input. On ImageNet-256, the selected a = 0.8 coincides with the cutoff of the EqM baseline (Appendix C.1).

  • What the analysis provides. Section 3.4 gives the criterion, which is that the threshold should lie where a noisy sample already identifies its clean example or its noise level. Appendix A.2 checks it with exact optimal fields for two moons embedded in higher dimension, where lower thresholds become admissible as the dimension grows. The analysis therefore indicates the range of thresholds to compare, and the comparison selects the value.

These results suggest that the threshold can be selected with a small model on the same data, at the cost of a few small-model runs, and need not be searched at the scale of the final model. This is the procedure that we followed, and following your comment we now describe it in the revised paper (Appendix B.2). We hope that this addresses your remaining concern.

If it does, we would be grateful if you would consider updating your score, and we are happy to answer any further questions.

Reviewer_E8U9OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MFM, which leverages the observation that the necessity for time conditioning varies across the stages of the generative process. MFM employs a time-conditioned transport approach based on FM during the initial phase, followed by a time-independent refinement process using EqM in the later phase. Furthermore, the paper also adds an endpoint penalty to leave clean data unchanged during refinement.

Critical strengths
  • This paper provides an analysis of the mode proportion issue in EqM. It demonstrates that even when samples reach the data manifold, they may fail to accurately reproduce the target distribution. This finding supports the design rationale behind MFM, which maintains initial time conditioning.

  • The proposed method demonstrates practical improvements in both generation performance and refinement stability.

Critical weaknesses

W1: The connection between the hybrid design (FM + EqM) rationale and its experimental validation is not sufficiently clear. While it is understandable that time conditioning can be omitted in the later stages, the additional benefits expected from combining early FM with late EqM are less clear. This makes it difficult to evaluate the necessity of the paper's core design choices.

W2: There is an insufficient explanation regarding the independent contribution of the endpoint penalty. Although its effect on refinement stability is presented, its contribution to overall performance is not clearly distinguished from that of the hybrid structure. Therefore, it is difficult to definitively assess the validity of the penalty itself and its specific contribution to the final proposed method.

Questions for authors

Q1: Could you explain the intended advantages of combining an initial FM phase with a subsequent EqM phase, compared to using either a full FM or a full EqM approach? Furthermore, could you clarify how the current experiments validate these design expectations, and provide further comparisons to substantiate this?

Q2: Based on the penalty ablation for the MFM presented in the paper, could the author provide a clearer explanation of the independent contribution of the endpoint penalty? Furthermore, in image generation, could the author compare the results obtained by applying the same penalty to the EqM versus not applying it, in order to disentangle the benefits of the endpoint penalty from those of the hybrid architecture?

2 comments
AuthorsOpenReview ↗

Response to E8U9

Thank you for your review, and for your positive comments on the mode-proportion analysis and on the improvements in generation and refinement stability. Your two questions helped us explain the design more clearly, and we added the comparison you suggested to the revised paper (new Table 14). We answer each question below.

Q1 / W1. "Could you explain the intended advantages of combining an initial FM phase with a subsequent EqM phase, compared to using either a full FM or a full EqM approach? ... could you clarify how the current experiments validate these design expectations, and provide further comparisons?"

Thank you for this question. It concerns the central design choice of the paper, so we are glad to clarify it. The combination is intended to keep one advantage of each method, and we test both.

  • From FM, it keeps the correct mode proportions. Far from the data, one noisy sample is consistent with many clean images at many noise levels, and a field without a time input must average over them. On a 2D target with unequal mode probabilities, the total-variation distance between the sampled and the target mode proportions is 0.288 for EqM and 0.011 for MFM (Section 4.5).
  • From EqM, it keeps a single time-independent field near the data. With the endpoint penalty, samples remain stable under additional refinement steps (Figure 11a), and the same field gives MFM the lowest FID for composition and for inpainting with half or more of the image hidden (Table 3).

As the further comparison you asked for, we trained the three methods on DiT-B/2 with the same target multiplier λ = 4, the same target and training cost, no endpoint penalty, and the same Euler update with 250 function evaluations (FM with the official EqM code). Full FM has an FID of 32.40, full EqM 32.62 and MFM 29.87. The models differ only in the time input, so the combination has a lower FID than either full method.

Q2 / W2. "Could the author provide a clearer explanation of the independent contribution of the endpoint penalty? ... compare the results obtained by applying the same penalty to the EqM versus not applying it?"

Thank you for suggesting this comparison, which separates the two effects directly. We ran it on DiT-B/2 and added it to the revised paper (new Table 14, FID without guidance).

MethodWithout the penaltyWith the same penalty
EqM32.6235.43
MFM (ours)29.8731.32

MFM has the lower FID in both cases, and MFM with the penalty also has a lower FID than EqM and FM without it. The penalty increases FID for both methods (by 1.5 for MFM and 2.8 for EqM), so it is not the source of MFM's lower FID, which comes from the time input.

The penalty's own contribution is to refinement stability and to sFID.

  • Stability. From 50 to 200 refinement steps, FID changes by less than 0.2 with the penalty and rises by 9.7 without it (Figure 11a, FID on 5,000 samples, separate-batch penalty). The in-batch estimator is also stable on CIFAR-10, where FID changes by at most 0.13 between 250 and 700 function evaluations and rises from 3.70 to 13.75 without the penalty (Table 7).
  • sFID. The penalty lowers sFID for both methods, from 6.83 to 5.86 for MFM (Table 14).

Thank you again for your questions, which helped us make the design and the role of the penalty clearer. We hope that these additions address your concerns. If they do, we would be grateful if you would consider updating your score, and we are happy to answer any further questions.

I appreciate the author's rebuttal and the additional experiments.

The comparison in Table 14, which uses matched training targets and computational budgets, supports the performance advantages of the hybrid design. The results obtained by applying the same endpoint penalty to EqM help clarify the respective contributions of the hybrid design and the penalty.

The additional evidence substantially addresses my concerns, so I will raise my score.

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

The paper proposes Mixed-Field Matching (MFM), which combines time-conditioned transport with autonomous refinement in a single network. The network receives the true interpolation time until a threshold (a), after which its time input is fixed. The refinement target is proportional to the displacement toward the clean training example. An endpoint penalty encourages the field to vanish on clean samples, and an in-batch estimator incorporates this penalty without increasing the number of network evaluations per training update.

The paper provides a conditional-expectation characterization of the regression cost of removing time information and a local refinement bound under a uniform approximation assumption. Experiments on ImageNet-256 and CIFAR-10 show improvements over the evaluated FM and EqM baselines. Additional experiments examine refinement stability, inpainting, composition, and mode-proportion recovery on synthetic distributions.

Critical strengths
  1. Meaningful improvements with a simple implementation and matched computational budgets. On ImageNet-256 with DiT-XL/2, the in-batch MFM variant achieves an unguided FID of 12.64, compared with 14.92 for EqM and 17.44 for FM, at the same reported training cost and 250 sampling evaluations. At guidance scale 1.5, the corresponding values are 3.532, 4.111, and 5.301. The advantage over EqM remains, although it becomes smaller, when each method uses its best tested guidance scale. These results support the practical usefulness of the proposed training and sampling scheme.

  2. The synthetic experiments distinguish convergence to the data support from recovery of the target probabilities. Section 4.5 and Appendix D.3 evaluate both distance to the modes and the proportions assigned to them, using population fields and learned networks. The learned-field comparison also controls the endpoint penalty between EqM and MFM. These experiments provide useful evidence that successful convergence to data points can coexist with substantial distributional error, and that time-conditioned transport can mitigate this problem.

Critical weaknesses
  1. The main contribution is not properly isolated. On ImageNet, FM uses λ=1\lambda=1, while MFM uses λ=4\lambda=4 and an endpoint penalty. The paper does not compare against noise-unconditional FM or EqM with the same endpoint penalty. Therefore, the reported gains may come from target scaling or regularization rather than the mixed-field schedule itself.

  2. The stability claim is stronger than the theory supports. Proposition 2 assumes a uniform approximation bound over a whole neighborhood, but the endpoint penalty only constrains the field at clean training samples. A zero field at the endpoint does not imply local contraction or convergence of nearby samples.

  3. The evidence for robust superiority is limited. The main DiT-XL/2 results use one training run, and the gains at the best guidance scales are comparable to run-to-run variation. On CIFAR-10, noise-unconditional FM nearly matches MFM, which weakens the claim that retaining time conditioning is generally necessary.

Questions for authors

See the weakness.

1 comment
AuthorsOpenReview ↗

Response to hGMB

Thank you for your constructive review, and for recognizing the improvements at matched cost and the value of the mode-proportion experiments. Your suggestions led to three new controls, a second run of every method and a revised Proposition 2. We believe the paper is stronger as a result, and we address each point below.

W1. "The reported gains may come from target scaling or regularization rather than the mixed-field schedule itself."

Thank you for raising this important point. We agree that these controls are needed to isolate the contribution of the time input. Following your suggestion, we trained all three controls you list on DiT-B/2 and added them to the revised paper (new Table 14). Every model uses the same Euler update and 250 function evaluations.

  • Target scaling. With the same λ = 4, the same target and no endpoint penalty, FM has an FID of 32.40, EqM 32.62 and MFM 29.87. These three models differ only in the time input. FM is trained with the official EqM code, which reproduces our FM with λ = 1 (37.23 compared with 37.45).
  • Regularization. With the same endpoint penalty, EqM has 35.43 and MFM 31.32. MFM with the penalty also has a lower FID than FM and EqM without it.
  • Noise-unconditional FM. With λ = 4 it has an FID of 35.73, which is 5.9 higher than MFM's.

Together, these controls give the matched comparison you asked for. MFM has the lowest FID with and without the penalty, so we attribute the improvement to the time input.

W2. "A zero field at the endpoint does not imply local contraction or convergence of nearby samples."

Thank you for this observation, which helped us state the result more precisely. Following your comment, we revised Proposition 2 so that it states the two requirements separately (Section 3.3, Appendix A.3). The first is that the field is small on the data manifold, which is what the endpoint penalty trains at the clean examples. The second is that the field's deviation from the direction to the nearest data point changes slowly near the data (a Lipschitz condition with a small enough constant). This second condition gives the contraction of nearby samples. Under both, each refinement step reduces the distance to the data manifold by a constant factor, up to a small remainder.

We also measured this directly on the trained ImageNet-256 model (Appendix A.3). The residual of the field, which is its norm rescaled so that it equals the distance to the data manifold for the ideal field, falls to about 0.03 within 200 refinement steps and stays between 0.03 and 0.04 up to 1,000 steps, for latents of norm 51.5. The residual of EqM's field at the data is 6.8 times larger. Accordingly, from 50 to 200 refinement steps FID changes by less than 0.2 with the penalty and rises by 9.7 without it (Figure 11a, FID on 5,000 samples, separate-batch penalty).

W3. "The main DiT-XL/2 results use one training run, and the gains at the best guidance scales are comparable to run-to-run variation. On CIFAR-10, noise-unconditional FM nearly matches MFM."

Thank you for this suggestion. Following it, we trained a second run of every method (Table 6). On DiT-XL/2, each MFM run has a lower FID than each EqM and FM run, with and without guidance. Averaged over the two runs, MFM's FID is 11% lower than EqM's without guidance and 10% lower at guidance scale 1.5. Without guidance, the difference between the two means is 1.66, compared with at most 0.97 between two runs of one method. The same ordering holds for the further models of Tables 14 and 15 (DiT-B/2 and DiT-L/2).

Your remark on CIFAR-10 is also helpful. On ImageNet-256, noise-unconditional FM has an FID 5.9 higher than MFM's (Table 14), so the small difference on CIFAR-10 agrees with the trend in our other results, where MFM's improvement is larger for the larger models. MFM has the lowest FID from DiT-B/2 to DiT-XL/2 (new Figure 5), and from DiT-S/2 to DiT-XL/2 its FID decreases 4.5 times, compared with 4.0 times for EqM (Tables 6 and 15). The time input also has a larger effect at this scale. Without guidance, removing it from FM raises FID by 4.1 (Wang & Du, Table 4), and Sun et al. show a similar increase for SiT-B/2 (their Figure 8).

Thank you again for your suggestions, which have improved the paper. We hope that the new experiments and the revised proposition address your concerns. If they do, we would be grateful if you would consider updating your score, and we are happy to answer any further questions.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Response to All

We sincerely thank the reviewers for their constructive feedback and suggestions, which have made the paper stronger. We are glad that the reviewers find the method simple and intuitive, with a single network (2c3L), that it gives meaningful improvements at matched training and sampling cost (hGMB, E8U9), that the endpoint penalty gives useful stability under additional refinement steps (2c3L, E8U9), that the mode-proportion analysis supports the design (hGMB, E8U9), and that the paper is clearly written (2c3L).

Summary. The reviewers' main request was to separate the effect of MFM's time input from the target multiplier λ and the endpoint penalty. We ran every control they requested. MFM has the lowest FID in each matched comparison, with and without the endpoint penalty. The result holds in repeated training runs, and MFM reaches the FID of the published EqM-XL/2 (within 0.01) with half the training. All changes are marked in red in the revised PDF.

1. Is the improvement due to MFM's time input, or to λ and the endpoint penalty? (hGMB, E8U9, 2c3L)

All three reviewers asked for this comparison, and we agree that it is important. Following their suggestions, we trained the requested controls on DiT-B/2 (new Table 14). The three models below have the same λ = 4, the same target and training cost, no endpoint penalty, and the same Euler update with 250 function evaluations, so they differ only in the time input.

MethodTime inputFID, no guidance
Flow matching (FM)at every time32.40
Equilibrium matching (EqM)none32.62
MFM (ours)below the threshold, held fixed above29.87

FM is trained with the official EqM code, which reproduces our FM with λ = 1 (37.23 compared with 37.45). With the endpoint penalty, as suggested by Reviewers hGMB and E8U9, MFM has 31.32 and EqM 35.43. Table 14 also includes FM with its own target scaled by 4 (33.99) and noise-unconditional FM (35.73), as requested by Reviewers 2c3L and hGMB. With or without the penalty, MFM has a lower FID than every FM and EqM model in the table.

2. Is the improvement robust across runs, model sizes and training lengths? (hGMB, 2c3L)

Motivated by the reviewers' comments on single training runs, we trained a second run of every method.

  • Second runs. Each MFM run has a lower FID than each EqM and FM run, on DiT-B/2 and DiT-XL/2, with and without guidance. On DiT-XL/2, at equal training cost and averaged over the two runs, MFM's FID is 11% lower than EqM's without guidance and 10% lower at guidance scale 1.5 (Table 6).
  • Model size. MFM has the lowest FID from DiT-B/2 to DiT-XL/2, and the difference is larger for the larger models. From DiT-S/2 to DiT-XL/2 its FID decreases 4.5 times, compared with 4.0 times for EqM (Tables 6 and 15).
  • Longer training. MFM has the lowest FID from 40 epochs onward. After 700 epochs, MFM-XL/2 achieves an FID of 1.94 at guidance scale 1.5. This is lower than DiT-XL/2 and SiT-XL/2 after 1400 epochs and within 0.01 of EqM-XL/2 after 1400 epochs with the same Euler update, so MFM reaches this FID with half the training (Table 1, Figure 4).

3. Do the theoretical statements match the trained model? (hGMB, 2c3L)

  • Proposition 2. Following Reviewer hGMB's observation, it now states the two requirements for contraction separately, and Appendix A.3 reports the corresponding measurement on the trained model.
  • Energy interpretation. Following Reviewer 2c3L's question, the title is now "Implicit Energy-Based Refinement", and Appendix A.3 states the sense precisely, following EqM.

We believe the paper is clearer and stronger thanks to the reviewers' suggestions. We thank them again for their time and effort, and we are happy to answer any further questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.