- 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
-
The method is simple, intuitive, and uses a single network.
-
The endpoint penalty provides useful stability under additional refinement steps.
-
The paper is clearly written and includes extensive implementation details and several metrics.
Critical weaknesses
-
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 , whereas FM uses an unscaled target. Appendix C shows that changing MFM from to worsens ImageNet FID by more than five points, suggesting that target scaling alone may explain a substantial part of the improvement.
-
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.
-
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
-
What is the performance of FM trained with (\lambda=4), with the output rescaled during sampling as done for MFM?
-
Can you report fully noise-unconditional FM on ImageNet using the same setting?
-
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
Part 2
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.
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.