acceptodds
Under review as a conference paper at ICLR 2027

Improved Training of Strong Stochastic Flow Maps

19 views

Abstract

Stochastic flow maps promise few-step sampling of stochastic differential equations (SDEs) by learning to directly jump between times along the solution path, yet training these models to learn this solution map remains difficult. The primary obstacle is the Brownian path itself: it is an infinite dimensional object which needs to be passed to the neural network and is nowhere differentiable which makes the implementation of derivative-based losses which power the state-of-the-art deterministic flow maps just out of reach. We identify two core design principles which address these issues: a) represent the Brownian path with a basis which carries higher-order local information, and b) train with an objective that forces the network to use the full representation of the path not just the first order. To address this we leverage the shifted Legendre polynomial expansion with a Wong–Zakai styled objective unlocking derivative-based losses. Finally, we propose a recipe for fine-tuning large pre-trained flow models into strong stochastic flow maps. We ablate through this design space on several toy problems and use our findings to transform a large flow model into a strong stochastic flow map on ImageNet to achieve a two-fold increase in performance at four NFE.

Primary area:
generative models
Keywords:
stochastic flow maps, SDEs, diffusion models, flow maps, few-step models, stochastic analysis

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_5V2DOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes iSSFM, an improved training framework for strong stochastic flow maps. The key idea is to use a shifted Legendre polynomial representation of Brownian paths together with a Wong–Zakai approximation, which transforms the underlying SDE into a smooth random ODE and enables derivative-based training objectives such as Improved MeanFlow. The authors further develop a fine-tuning strategy that converts pretrained flow models into strong stochastic flow maps.The method is evaluated on several toy problems and ImageNet-256, demonstrating substantial improvements over existing semigroup-based training.

Critical strengths
  1. Interesting methodological contribution. The combination of shifted Legendre polynomial representations and Wong–Zakai approximation provides an effective way to extend derivative-based training objectives to strong stochastic flow maps. In particular, the proposed formulation enables the use of higher-order Brownian path information during training.

  2. Strong empirical results. The proposed method substantially improves few-step sampling performance over the semigroup-based baseline. On ImageNet-256, it reduces FID from 12.97 to 5.36 at 4 NFE. The experiments also include useful comparisons of Brownian representations and training objectives.

Critical weaknesses
  1. Limited evaluation of strong approximation accuracy. While the paper demonstrates strong improvements in sampling quality, the ImageNet experiments evaluate pathwise consistency against the model's own 32-step solution rather than a high-accuracy numerical solution of the pretrained SDE. A direct comparison against the underlying SDE would provide stronger evidence of the learned strong stochastic flow map's accuracy.

  2. Limited downstream applications. The Feynman–Kac corrector experiment provides an interesting application, but it is restricted to a low-dimensional Gaussian mixture. The practical advantages of strong stochastic flow maps over deterministic or weak stochastic alternatives remain insufficiently demonstrated, particularly for high-dimensional generative modeling tasks.

Questions for authors
  1. Applicability to pretrained diffusion models. Can the proposed Wong–Zakai-based training method be directly applied to pretrained score-based diffusion models? If so, what is the motivation or practical advantage of constructing an SDE from a pretrained flow matching model rather than directly learning the stochastic flow map of a pretrained diffusion SDE?

  2. Necessity of strong stochastic flow maps. Could the authors elaborate on the practical importance of preserving strong pathwise correspondence? Beyond the Feynman–Kac corrector example, are there important downstream applications that specifically require strong stochastic flow maps rather than weak stochastic flow maps or deterministic alternatives? In particular, what advantages does the strong property provide that cannot be achieved by matching the appropriate transition distributions?

6 comments
AuthorsOpenReview ↗

Response to 5V2D part 1

We thank reviewer 5V2D for their thoughtful review of our paper and interest in the downstream applications. We are happy the reviewer found the work to be an "[i]nteresting methodological contribution" and to have "[s]trong empirical results". Below we answer the questions and concerns raised by the reviewer.

Limited evaluation of strong approximation accuracy. While the paper demonstrates strong improvements in sampling quality, the ImageNet experiments evaluate pathwise consistency against the model's own 32-step solution rather than a high-accuracy numerical solution of the pretrained SDE. A direct comparison against the underlying SDE would provide stronger evidence of the learned strong stochastic flow map's accuracy.

At the reviewer's suggestion we compared these methods against an underlying 512-step Euler–Maruyama solve (revised Figure 4b; Appendix H.3, Table 7). We compare each one of our core changes: the exponential lift, closed-form martingale, drift matching loss, and Wong–Zakai iMF loss to show the impact on strong error. N.B., as our parameterizations use an exponential lift we compare those to the exponentially lifted SDE Yt=αtXt\boldsymbol Y_t = \alpha_t \boldsymbol X_t.

Table R.1. Strong mean squared error w.r.t. the teacher SDE (Table 7 of the revised PDF).

ModelRef.NFE 1NFE 2NFE 4NFE 8NFE 16NFE 32
SSFMXt\boldsymbol X_t0.3540.3510.3290.3130.3190.328
+ exp. liftYt\boldsymbol Y_t0.3590.3570.3330.3220.3350.350
+ closed-form martingaleYt\boldsymbol Y_t0.3220.3040.2690.2580.2530.238
+ drift matchingYt\boldsymbol Y_t0.3270.3030.2530.2220.1910.147
+ Wong–Zakai iMF lossYt\boldsymbol Y_t0.7270.5280.4480.3640.2630.170
Teacher EM (plain)Xt\boldsymbol X_t9988.55.8750.9340.8280.6660.389
Teacher EM (lifted)Yt\boldsymbol Y_t0.5790.6750.5410.4840.3830.292

Our improved parameterization (exponential lift, closed-form martingale, drift matching) yields the lowest strong error, and it is below the teacher's own lifted Euler–Maruyama solver at matched NFE. iSSFM, which adds the Wong–Zakai iMF loss, gives the best fidelity among the stochastic flow maps and converges strongly.

Applicability to pretrained diffusion models. Can the proposed Wong–Zakai-based training method be directly applied to pretrained score-based diffusion models? If so, what is the motivation or practical advantage of constructing an SDE from a pretrained flow matching model rather than directly learning the stochastic flow map of a pretrained diffusion SDE?

Yes we can do both! We chose to write the paper using the flow matching formalism but everything holds mutatis mutandis for score matching models using utθ(x)=α˙tαtx+νt22∇log⁡ptθ(x)⏟=stθ(x).\boldsymbol u_t^\theta(\boldsymbol x) = \frac{\dot\alpha_t}{\alpha_t} \boldsymbol x + \frac{\nu_t^2}{2} \underbrace{\nabla \log p_t^\theta(\boldsymbol x)}_{= \boldsymbol s_t^\theta(\boldsymbol x)}. It is merely a choice of how the base model was trained, did it learn the velocity, score, noise prediction, or data prediction? We can then fine-tune from any of these formalisms. We clarified this in footnote 2 of the revised PDF.

AuthorsOpenReview ↗

Response to 5V2D part 2

Weakness. Limited downstream applications. The Feynman–Kac corrector experiment provides an interesting application, but it is restricted to a low-dimensional Gaussian mixture. The practical advantages of strong stochastic flow maps over deterministic or weak stochastic alternatives remain insufficiently demonstrated, particularly for high-dimensional generative modeling tasks.

Question. Necessity of strong stochastic flow maps. Could the authors elaborate on the practical importance of preserving strong pathwise correspondence? Beyond the Feynman–Kac corrector example, are there important downstream applications that specifically require strong stochastic flow maps rather than weak stochastic flow maps or deterministic alternatives? In particular, what advantages does the strong property provide that cannot be achieved by matching the appropriate transition distributions?

Yes, we find that for image editing with SDEs this pathwise property is crucial.

We ran some additional experiments (new Section 4.3, Figure 5, Table 4; Appendix H.4) doing editing of 240 real ImageNet-256 images consisting of 30 source to target class pairs each with 8 images. Every method received the same NFE budget BB, counting both inversion from the real image and generation. We map a source image X1\boldsymbol X_1 to a noisy state Xt0\boldsymbol X_{t_0}, change the class condition, and generate from Xt0\boldsymbol X_{t_0}. DMF inverts its ODE, whereas the SDE methods noise the source forward, record the Brownian increments that carry it back to X1\boldsymbol X_1 under the source class [1], and replay those increments under the new class; this requires a map that follows a given Brownian path. Methods are compared at equal preservation to the source image measured in LPIPS [3] from the original image. We compare DMF, DMF (denoise–renoise), the teacher SDE following [1] (Cycle-SDE), SSFM, and iSSFM, as well as SSFM (weak) and iSSFM (weak).

Table R.2. Image editing with different models (Table 8 of the revised PDF).

MethodNFECLIP @ LPIPS 0.12CLIP @ 0.20PickScore @ 0.12PickScore @ 0.20ImageReward @ 0.12ImageReward @ 0.20
DMF1620.6221.7619.4519.60−1.25−1.08
Teacher SDE (Cycle-SDE)1621.01—19.40—−1.22—
SSFM (weak)16—19.65—19.26—−1.40
iSSFM (weak)16—19.73—19.35—−1.36
SSFM1620.5521.6719.4019.53−1.24−1.11
iSSFM (ours)1621.0122.1519.4519.59−1.18−1.04
DMF3220.7221.8919.4119.55−1.23−1.07
Teacher SDE (Cycle-SDE)3220.9721.8819.4319.44−1.21−1.16
SSFM (weak)32—19.67—19.22—−1.40
iSSFM (weak)32—19.82—19.32—−1.36
SSFM3220.7222.0019.4219.58−1.22−1.07
iSSFM (ours)3221.0522.3419.4719.64−1.17−0.99
Teacher SDE (Cycle-SDE), many-step reference38–10421.0421.9419.4419.48−1.21−1.19

N.B., entries with — denote that the method never preserves the source this well; even its mildest edit is further away. Note, even the mildest edit of DMF (denoise–renoise) has LPIPS 0.25 at NFE 16 and 0.30 at NFE 32 (row omitted from table since it failed). The teacher SDE uses the same recorded Brownian increments as iSSFM with B/2B/2 Euler–Maruyama steps of the frozen teacher. The many-step reference runs on the full 64-step rollout, so its NFE ranges from 38 to 104 depending on t0t_0. (weak) denotes the same network run from the noised source on an independent Brownian path instead of the recorded one, i.e., using only its transition distributions, thus a weak map.

Figure 5 of the revised PDF shows qualitative edits at 32 NFE: iSSFM changes the object while keeping the pose, layout, and background of the source, DMF and DMF (denoise–renoise) redraw the scene, and iSSFM (weak) keeps the class but not the layout.

We summarize the main takeaways:

  • iSSFM edits best at matched NFE, leading DMF and SSFM on every metric at NFE 32 (Appendix H.4, Table 9).
  • DMF (denoise–renoise) cannot edit conservatively: it largely creates an entirely new image.
  • SSFM roughly ties DMF; our improvements are what make pathwise editing beat DMF.
  • The strong property is necessary: the weak maps lose 2–2.5 CLIP at equal preservation and cannot edit conservatively.
  • The teacher SDE needs many more steps and still trails iSSFM.

Additionally, concurrent work to ours [2] has explored SSFMs in physical time to accelerate molecular dynamics by learning to take larger jumps in solving the Langevin SDE.

AuthorsOpenReview ↗

Response to 5V2D part 3

Concluding remarks

We thank the reviewer for their insight and helpful suggestions. We believe that the new experiments exploring both the strong property and its applications for high-dimensional generative modeling tasks have greatly improved the quality of this paper. We kindly ask that in light of these new experiments and clarifications the reviewer would consider them in their review. We are happy to engage in further discussion.

References

[1] Nie, S., Guo, H. A., Lu, C., Zhou, Y., Zheng, C., & Li, C. (2024). The Blessing of Randomness: SDE Beats ODE in General Diffusion-based Image Editing. ICLR 2024.

[2] McCallum, S., Rindtorff, N., Tong, A., & Foster, J. (2026). Langevin Flow Maps: Efficient Molecular Dynamics and Transition Path Sampling. arXiv preprint arXiv:2610.05998.

[3] Zhang, R., Isola, P., Efros, A. A., Shechtman, E., & Wang, O. (2018). The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. CVPR 2018.

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

This paper proposes an improved training recipe for strong stochastic flow maps, which approximate SDE solutions conditioned on a Brownian realization. The method uses a shifted-Legendre approximation of the Brownian path and a Wong–Zakai formulation to enable derivative-based flow-map objectives. It also introduces a parameterization for converting pretrained deterministic flow models into stochastic maps. Experiments cover toy SDEs, ImageNet generation, and Feynman–Kac correction.

Critical strengths

The paper addresses a meaningful technical obstacle in applying derivative-based objectives when the conditioning Brownian path is nowhere differentiable. The suggested solution is principled which is done by combining smooth path approximations with interval-composition rules , and the comparisons across representations and objectives help motivate the final design. The ImageNet improvement over the matched semigroup baseline looks good, e.g., reducing 4-NFE FID from 12.97 to 5.36. The Feynman–Kac experiment also provides a relevant application where tracking stochastic trajectories has value beyond image generation.

Critical weaknesses

I have some minor issues on the experiments and the objective of the paper.

  1. I believe the ImageNet experiments do not clearly establish why stochastic flow-map learning is beneficial for image quality. The main comparison shows that the proposed recipe improves over another stochastic map, but omits the pretrained deterministic DMF and a restart (i.e., denoise–renoise) sampling baseline at matched NFE. Since these alternatives can already generate diverse samples, the practical benefit of learning a Brownian-path-conditioned map remains unclear. Preserving a particular SDE solution is a meaningful goal, but it should be distinguished from improving image generation. As such, the claim should be justified whether the stochastic flow maps are relevant for the high-quality image generation, or in learning the distribution of samples made by SDE.

  2. The role of ImageReward, PickScore, and HPS in the ImageNet evaluation is insufficiently motivated. These metrics evaluate learned preferences or compatibility with class-derived text prompts, rather than fidelity to the ImageNet distribution. Higher reward can reflect preferred visual characteristics without indicating better distribution matching. FID and FD-DINO provide more direct evidence for the distributional claim; the paper should clarify what additional property the reward metrics establish and treat them as supplementary evidence.

  3. Finally, the large-scale pathwise evaluation measures consistency with each model’s own 32-step rollout rather than accuracy against the teacher SDE under the same Brownian realization. This does not distinguish a faithful approximation from an internally consistent but biased map. The theoretical convergence of the smooth-path ODE as the representation order increases does not resolve this issue for the learned, finite-order model.

Questions for authors
  1. How does iSSFM compare with the pretrained deterministic DMF and a denoise–renoise sampler at matched NFE? Are existing results available that clarify the image-quality benefit of stochastic map learning?

  2. What additional property are ImageReward, PickScore, and HPS intended to establish in this class-conditional setting? Please distinguish perceptual preference or class compatibility from distributional fidelity.

  3. Is pathwise evaluation against a sufficiently resolved teacher-SDE trajectory available? How should the semigroup baseline’s lower self-consistency error in Figure 4 be interpreted alongside iSSFM’s better FID?

5 comments
AuthorsOpenReview ↗

Response to tMgy part 1

We thank reviewer tMgy for their detailed review of our work and the interest in our experimental results. We are glad that the reviewer found our paper to address "a meaningful technical obstacle in applying derivative-based objectives". Below we answer their questions.

How does iSSFM compare with the pretrained deterministic DMF and a denoise–renoise sampler at matched NFE? Are existing results available that clarify the image-quality benefit of stochastic map learning?

This is a great question.

Part 1. For sampling from noise, the pre-trained DMF with denoise–renoise steps is a better few-step sampler, with an FID of 1.82 at 4 NFE for DMF vs. 5.96 for iSSFM (and 2.33 vs. 2.37 at 8 NFE). We do however, observe better FID at 16 NFE, viz. 3.72 (DMF) vs. 2.07 (iSSFM).

Part 2. DMF with denoise–renoise cannot sample the transition kernel pt∣s(⋅∣Xs)p_{t|s}(\cdot|\boldsymbol X_s), let alone follow a given Brownian path, making it a different object from SSFM/iSSFM. To more properly study what this difference means we ran some additional experiments where we compare SSFM to our iSSFM and to DMF on image editing (new Section 4.3, Figure 5, Table 4; Appendix H.4). Viz., editing of 240 real ImageNet-256 images consisting of 30 source to target class pairs each with 8 images. Every method received the same NFE budget BB, counting both inversion from the real image and generation. We map a source image X1\boldsymbol X_1 to a noisy state Xt0\boldsymbol X_{t_0}, change the class condition, and generate from Xt0\boldsymbol X_{t_0}. DMF inverts its ODE, whereas the SDE methods noise the source forward, record the Brownian increments that carry it back to X1\boldsymbol X_1 under the source class [2], and replay those increments under the new class; this requires a map that follows a given Brownian path. Methods are compared at equal preservation to the source image measured in LPIPS [1] from the original image. We compare DMF, DMF (denoise–renoise), the teacher SDE following [2] (Cycle-SDE), SSFM, and iSSFM, as well as SSFM (weak) and iSSFM (weak).

Table R.1. Image editing with different models (Table 8 of the revised PDF).

MethodNFECLIP @ LPIPS 0.12CLIP @ 0.20PickScore @ 0.12PickScore @ 0.20ImageReward @ 0.12ImageReward @ 0.20
DMF1620.6221.7619.4519.60−1.25−1.08
Teacher SDE (Cycle-SDE)1621.01—19.40—−1.22—
SSFM (weak)16—19.65—19.26—−1.40
iSSFM (weak)16—19.73—19.35—−1.36
SSFM1620.5521.6719.4019.53−1.24−1.11
iSSFM (ours)1621.0122.1519.4519.59−1.18−1.04
DMF3220.7221.8919.4119.55−1.23−1.07
Teacher SDE (Cycle-SDE)3220.9721.8819.4319.44−1.21−1.16
SSFM (weak)32—19.67—19.22—−1.40
iSSFM (weak)32—19.82—19.32—−1.36
SSFM3220.7222.0019.4219.58−1.22−1.07
iSSFM (ours)3221.0522.3419.4719.64−1.17−0.99
Teacher SDE (Cycle-SDE), many-step reference38–10421.0421.9419.4419.48−1.21−1.19

N.B., entries with — denote that the method never preserves the source this well; even its mildest edit is further away. Note, even the mildest edit of DMF (denoise–renoise) has LPIPS 0.25 at NFE 16 and 0.30 at NFE 32 (row omitted from table since it failed). The teacher SDE uses the same recorded Brownian increments as iSSFM with B/2B/2 Euler–Maruyama steps of the frozen teacher. The many-step reference runs on the full 64-step rollout, so its NFE ranges from 38 to 104 depending on t0t_0. (weak) denotes the same network run from the noised source on an independent Brownian path instead of the recorded one, i.e., using only its transition distributions, thus a weak map.

Figure 5 of the revised PDF shows qualitative edits at 32 NFE: iSSFM changes the object while keeping the pose, layout, and background of the source, DMF and DMF (denoise–renoise) redraw the scene, and iSSFM (weak) keeps the class but not the layout.

We summarize the main takeaways:

  • iSSFM edits best at matched NFE, leading DMF and SSFM on every metric at NFE 32 (Appendix H.4, Table 9).
  • DMF (denoise–renoise) cannot edit conservatively: it largely creates an entirely new image.
  • SSFM roughly ties DMF; our improvements are what make pathwise editing beat DMF.
  • The strong property is necessary: the weak maps lose 2–2.5 CLIP at equal preservation and cannot edit conservatively.
  • The teacher SDE needs many more steps and still trails iSSFM.
AuthorsOpenReview ↗

Response to tMgy part 2

What additional property are ImageReward, PickScore, and HPS intended to establish in this class-conditional setting? Please distinguish perceptual preference or class compatibility from distributional fidelity.

We included the ImageReward, PickScore, and HPS metrics as an additional supplement to the main distributional metrics of FID and FD-DINOv2, as mentioned by the reviewer these metrics measure preferred visual characteristics. We have made this more clear in Section 4.2 of the revised pdf.

Is pathwise evaluation against a sufficiently resolved teacher-SDE trajectory available? How should the semigroup baseline’s lower self-consistency error in Figure 4 be interpreted alongside iSSFM’s better FID?

We observe that our improvements increase our strong convergence property, i.e., our error shrinks more quickly as the number of steps grows larger and we also retain better fidelity.

Below we compare each of our core changes, the exponential lift, closed-form martingale, drift matching loss, and Wong–Zakai iMF loss, along with an ODE flow map (DMF) and DMF with denoise–renoise sampling to show the impact on strong error. We measure the strong error w.r.t. an underlying ground truth 512-step Euler–Maruyama solve; this replaces the self-consistency plot in the revised Figure 4b, with full results in Appendix H.3, Table 7. N.B., as our parameterizations use an exponential lift we compare those to the exponentially lifted SDE Yt=αtXt\boldsymbol Y_t = \alpha_t \boldsymbol X_t.

Table R.2. Strong mean squared error w.r.t. the teacher SDE (Table 7 of the revised PDF).

ModelRef.NFE 1NFE 2NFE 4NFE 8NFE 16NFE 32
SSFMXt\boldsymbol X_t0.3540.3510.3290.3130.3190.328
+ exp. liftYt\boldsymbol Y_t0.3590.3570.3330.3220.3350.350
+ closed-form martingaleYt\boldsymbol Y_t0.3220.3040.2690.2580.2530.238
+ drift matchingYt\boldsymbol Y_t0.3270.3030.2530.2220.1910.147
+ Wong–Zakai iMF lossYt\boldsymbol Y_t0.7270.5280.4480.3640.2630.170
DMF-XL/2+Yt\boldsymbol Y_t1.2021.2001.2161.2451.2531.244
DMF-XL/2+ (denoise–renoise)Yt\boldsymbol Y_t1.2021.1951.2981.2001.2041.224
Teacher EM (plain)Xt\boldsymbol X_t9988.55.8750.9340.8280.6660.389
Teacher EM (lifted)Yt\boldsymbol Y_t0.5790.6750.5410.4840.3830.292

Concluding remarks

We thank the reviewer again for their time and careful review of our work. We believe that our new experiments directly comparing to DMF have created a much stronger and more interesting paper and we thank the reviewer for encouraging us to do so. We hope the reviewer will reconsider their review now with the insight from these new experiments. We are happy to answer any additional questions.

References

[1] Zhang, Richard, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In 2018 IEEE/CVF conference on computer vision and pattern recognition, pp. 586-595. IEEE, 2018.

[2] Nie, Shen, Hanzhong Guo, Cheng Lu, Yuhao Zhou, Chenyu Zheng, and Chongxuan Li. The blessing of randomness: SDE beats ODE in general diffusion-based image editing. In International Conference on Learning Representations, vol. 2024, pp. 41763-41799. 2024.

Dear reviewer tMgy,

We are thankful for your time and insightful review, particularly your emphasis on comparison to the teacher DMF model and the benefits of stochastic map learning. As the end of the rebuttal period is quickly approaching we wanted to highlight that we have now compared to the deterministic DMF and DMF with denoise-renoise sampling, see Table R.1. We also have reported the strong error w.r.t. a sufficiently resolved teacher-SDE trajectory for the various models (Table R.2).

We are happy to answer any remaining or additional questions. If our responses and additional experiments address your concerns, we would be grateful if you would consider updating your assessment.

Reviewer_bwe6OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper improves the training of neural maps that approximate additive-noise SDE solutions conditioned on a Brownian path. It combines local Legendre representations, Wong–Zakai smoothing, derivative-based objectives, and an exponential-integrator parameterization. Toy experiments examine pathwise errors, while ImageNet experiments demonstrate substantial few-step quality improvements over a matched semigroup-training baseline.

Critical strengths
  1. The paper develops a coherent training approach to a genuine difficulty in learning stochastic dynamics. The combination of controlled experiments and substantial generation-quality gains supports a meaningful methodological contribution.
  2. A coherent solution to a genuine training obstacle. Smoothing the Brownian driver makes derivative-based flow-map objectives usable while retaining a pathwise approximation interpretation (§3.2; Appendix C.3). Figure 3b and Table 3 evaluate strong errors against an independently integrated SDE reference, providing evidence that the proposed objectives use higher-order path information effectively.
Critical weaknesses
  1. The evidence is stronger for improved generation quality than for faithful reproduction of the intended stochastic dynamics at scale. This limits how broadly the reported improvements can support the paper's strongest pathwise-accuracy claims.
  2. ImageNet self-consistency does not establish fidelity to the intended SDE. Figure 4b measures agreement with each model’s own 32-step rollout. A consistently biased model could perform well under this test, and iSSFM shows higher self-consistency error than SSFM despite better FID. Proposition C.3 concerns the exact smoothed dynamics as representation order increases, not the accuracy of the trained low-order map. Consequently, the large-scale evidence supports image quality more strongly than strong pathwise fidelity.
Questions for authors
  1. On the same Brownian realizations, how do iSSFM and the semigroup baseline compare against a converged numerical solution of the frozen teacher SDE? Does the FID gain preserve pathwise accuracy, or trade it away?
  2. Do the representation rankings persist when each coefficient budget is trained separately and the closest KL-based method is evaluated with its original training objective? A controlled toy experiment would directly address the claimed representation advantage.
5 comments
AuthorsOpenReview ↗

Response to bew6 part 1

We thank reviewer bwe6 for their detailed review and insightful questions about the pathwise fidelity of the learned stochastic flow maps. We are happy that the reviewer finds iSSFM to be a "coherent solution to a genuine training obstacle." Below we address the questions raised during the review.

On the same Brownian realizations, how do iSSFM and the semigroup baseline compare against a converged numerical solution of the frozen teacher SDE? Does the FID gain preserve pathwise accuracy, or trade it away?

We thank the reviewer for the suggestion and evaluate the pathwise fidelity of the underlying SDE. We find that the FID/FD-DINOv2 gain does not trade away the strong convergence property.

We compare each one of our core changes, the exponential lift, closed-form martingale, drift matching loss, and Wong–Zakai iMF loss to show the impact on strong error. In the table below we measure the strong error w.r.t. an underlying ground truth 512-step Euler–Maruyama solve (revised Figure 4b; Appendix H.3, Table 7). N.B., as our parameterizations use an exponential lift we compare those to the exponentially lifted SDE Yt=αtXt\boldsymbol Y_t = \alpha_t \boldsymbol X_t. The semigroup baseline of our submission corresponds to the "+ drift matching" row below, and iSSFM to the "+ Wong–Zakai iMF loss" row.

Table R.1. Strong mean squared error w.r.t. the teacher SDE (Table 7 of the revised PDF).

ModelRef.NFE 1NFE 2NFE 4NFE 8NFE 16NFE 32
SSFMXt\boldsymbol X_t0.3540.3510.3290.3130.3190.328
+ exp. liftYt\boldsymbol Y_t0.3590.3570.3330.3220.3350.350
+ closed-form martingaleYt\boldsymbol Y_t0.3220.3040.2690.2580.2530.238
+ drift matchingYt\boldsymbol Y_t0.3270.3030.2530.2220.1910.147
+ Wong–Zakai iMF lossYt\boldsymbol Y_t0.7270.5280.4480.3640.2630.170
Teacher EM (plain)Xt\boldsymbol X_t9988.55.8750.9340.8280.6660.389
Teacher EM (lifted)Yt\boldsymbol Y_t0.5790.6750.5410.4840.3830.292

Our improved parameterization (exponential lift, closed-form martingale, drift matching) yields the lowest strong error. iSSFM, which adds the Wong–Zakai iMF loss, gives the best fidelity among the stochastic flow maps (Tables R.2a and R.2b) and converges strongly. We also ablate all of our improvements on the FID and FD-DINOv2 metrics below (revised Table 1; full results in Table 6).

N.B., in our original tables SSFM denoted SSFM + exp. lift + closed-form martingale + drift matching, i.e., SSFM + improved parameterization (FID 12.97 vs. 5.36 for iSSFM at 4 NFE, each trained for up to 200k iterations with the best EMA checkpoint). We have since replicated the original SSFM recipe in this setting and ablated our improvements one at a time, training every model for the same 100k iterations. The submitted baseline thus corresponds to the "+ drift matching" row (FID 14.67 at 4 NFE), and iSSFM reaches 5.96. The conclusion is the same, i.e., at 4 NFE iSSFM reduces FID by about 2.5× over the semigroup baseline and about 5× over the original SSFM recipe (31.00).

Table R.2a. FID (Tables 1 and 3 of the revised PDF).

ModelNFE 1NFE 2NFE 4NFE 8NFE 16
SSFM197.1183.5631.0011.094.82
+ exp. lift208.2590.1628.158.073.16
+ closed-form martingale163.5660.2118.334.952.98
+ drift matching184.7457.9914.674.152.03
iSSFM (+ WZ-iMF)111.6153.115.962.372.07

Table R.2b. FD-DINOv2 (Tables 1 and 3 of the revised PDF).

ModelNFE 1NFE 2NFE 4NFE 8NFE 16
SSFM2116.91061.3475.1229.1135.6
+ exp. lift2112.21113.1447.4190.8104.0
+ closed-form martingale1561.2723.0301.6125.772.6
+ drift matching1817.6772.3288.1128.270.8
iSSFM (+ WZ-iMF)994.2583.5143.075.255.1
AuthorsOpenReview ↗

Response to bwe6 part 2

Do the representation rankings persist when each coefficient budget is trained separately and the closest KL-based method is evaluated with its original training objective? A controlled toy experiment would directly address the claimed representation advantage.

We agree this is a useful addition to our analysis and summarize the results below, now training each method and coefficient NN budget separately (revised Figure 3a). We observe that the shifted Legendre basis still results in a lower strong error than the other expansions.

We also add in the Lagrangian loss with the sinusoidal Karhunen–Loève representation in addition to the standard semigroup objective; both are used in [1]. As in our other experiments, the Lagrangian objective seems to be less stable compared to the semigroup objective, suffering a much higher strong error. We report the average strong error w.r.t. a 512-step Euler–Maruyama solve of each model's own learned SDE on the same Brownian paths.

Table R.3. Checkerboard strong RMSE (Figure 3a of the revised PDF).

MethodNFEN=1N=2N=3N=4N=5N=6N=7N=8
Shifted-Legendre11.237$_{\pm 0.002}$0.468$_{\pm 0.002}$0.386±0.004_{\pm 0.004}0.320±0.005_{\pm 0.005}0.288±0.004_{\pm 0.004}0.285±0.008_{\pm 0.008}0.285±0.007_{\pm 0.007}0.283_{\pm 0.005}\
Global KL10.865$_{\pm 0.002}$0.805_{\pm 0.002}\0.667$_{\pm 0.002}$0.515$_{\pm 0.003}$0.460$_{\pm 0.003}$0.428$_{\pm 0.005}$0.385$_{\pm 0.007}$0.373$_{\pm 0.003}$
Global KL + local increment11.238$_{\pm 0.002}$0.881$_{\pm 0.002}$0.722$_{\pm 0.002}$0.502$_{\pm 0.003}$0.474$_{\pm 0.002}$0.412$_{\pm 0.006}$0.393$_{\pm 0.003}$0.367$_{\pm 0.004}$
Local dyadic increments11.237$_{\pm 0.002}$0.739$_{\pm 0.003}$0.741$_{\pm 0.001}$0.603$_{\pm 0.003}$0.602$_{\pm 0.002}$0.574$_{\pm 0.004}$0.574$_{\pm 0.003}$0.573$_{\pm 0.003}$
Global KL + DS–Lagrangian11.587$_{\pm 0.040}$1.481$_{\pm 0.056}$1.368$_{\pm 0.029}$1.275$_{\pm 0.038}$1.255$_{\pm 0.068}$1.177$_{\pm 0.050}$1.189$_{\pm 0.038}$1.167$_{\pm 0.052}$
Shifted-Legendre20.738$_{\pm 0.002}$0.342$_{\pm 0.005}$0.279±0.006_{\pm 0.006}0.261±0.004_{\pm 0.004}0.260±0.005_{\pm 0.005}0.259±0.008_{\pm 0.008}0.261±0.009_{\pm 0.009}0.261_{\pm 0.006}\
Global KL20.660$_{\pm 0.002}$0.627_{\pm 0.004}\0.618$_{\pm 0.002}$0.596$_{\pm 0.001}$0.508$_{\pm 0.004}$0.449$_{\pm 0.004}$0.427$_{\pm 0.005}$0.423$_{\pm 0.002}$
Global KL + local increment20.738$_{\pm 0.002}$0.703$_{\pm 0.002}$0.589$_{\pm 0.003}$0.575$_{\pm 0.003}$0.570$_{\pm 0.003}$0.485$_{\pm 0.003}$0.449$_{\pm 0.003}$0.443$_{\pm 0.004}$
Local dyadic increments20.738$_{\pm 0.002}$0.572$_{\pm 0.004}$0.572$_{\pm 0.001}$0.550$_{\pm 0.002}$0.551$_{\pm 0.003}$0.406$_{\pm 0.007}$0.407$_{\pm 0.003}$0.388$_{\pm 0.004}$
Global KL + DS–Lagrangian21.402$_{\pm 0.035}$1.285$_{\pm 0.053}$1.230$_{\pm 0.025}$1.180$_{\pm 0.032}$1.152$_{\pm 0.058}$1.077$_{\pm 0.047}$1.097$_{\pm 0.029}$1.074$_{\pm 0.048}$
Shifted-Legendre40.571$_{\pm 0.002}$0.254±0.010_{\pm 0.010}0.219±0.008_{\pm 0.008}0.225±0.009_{\pm 0.009}0.224±0.006_{\pm 0.006}0.225±0.008_{\pm 0.008}0.227±0.007_{\pm 0.007}0.222_{\pm 0.008}\
Global KL40.629$_{\pm 0.003}$0.632$_{\pm 0.003}$0.570$_{\pm 0.002}$0.528$_{\pm 0.004}$0.521$_{\pm 0.003}$0.517$_{\pm 0.002}$0.481$_{\pm 0.004}$0.462$_{\pm 0.003}$
Global KL + local increment40.571$_{\pm 0.001}$0.568_{\pm 0.002}\0.534$_{\pm 0.003}$0.531$_{\pm 0.003}$0.497$_{\pm 0.005}$0.486$_{\pm 0.002}$0.482$_{\pm 0.002}$0.465$_{\pm 0.004}$
Local dyadic increments40.571$_{\pm 0.002}$0.392_{\pm 0.004}\0.389$_{\pm 0.004}$0.336$_{\pm 0.005}$0.337$_{\pm 0.003}$0.279$_{\pm 0.008}$0.279$_{\pm 0.004}$0.269$_{\pm 0.006}$
Global KL + DS–Lagrangian41.099$_{\pm 0.015}$1.046$_{\pm 0.026}$0.924$_{\pm 0.011}$0.868$_{\pm 0.035}$0.876$_{\pm 0.047}$0.847$_{\pm 0.035}$0.834$_{\pm 0.016}$0.802$_{\pm 0.021}$

Concluding remarks

We thank the reviewer for their time and careful review of our work. We believe that the new more detailed examination of the strong error on the ImageNet experiments has resulted in a much better paper, along with the more comprehensive study on representing the Brownian path. We kindly ask the reviewer to reconsider their review in light of these new experiments and clarifications. We are happy to continue the discussion and provide any further clarifications.

References

[1] Pan, Z., Potaptchik, P., Yao, W., Albergo, M. S., and Pidstrigach, J. (2026). Itô maps for any-step SDEs. In: arXiv preprint arXiv:2606.11156

Dear reviewer bwe6,

We are very appreciative of your time and constructive feedback on our submission. We hope that our rebuttal responses, in particular the strong error w.r.t. the underlying SDE rollout (Table R.1) and updated comparison of the Brownian path representations (Table R.3), have addressed your questions.

As the end of the rebuttal period is quickly approaching, we would be happy to answer any remaining or additional questions. If our responses and additional experiments address your concerns, we would be grateful if you would consider updating your assessment.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Global response

We thank all reviewers for their careful reviews. All three reviewers asked how faithfully iSSFM follows the underlying SDE at scale, and two asked what the strong property offers beyond matching transition distributions. We have addressed all the concerns and questions in this rebuttal. We summarize the new experiments and the corresponding changes to the revised PDF below, with all the new and changed text in the revised PDF marked in blue.

Pathwise accuracy against the teacher SDE (bwe6, tMgy, 5V2D). We compare every model against a converged 512-step Euler–Maruyama solve of the frozen teacher SDE on the same Brownian path (Figure 4b; full results in Appendix H.3, Table 7). Our parameterization does indeed converge strongly with the error decreasing with the number of steps, with the non-strong baselines unable to do so.

Ablation from SSFM to iSSFM (bwe6, tMgy). We replicate the original SSFM recipe and add our changes one at a time, training every model for the same 100k iterations (Table 1). The submitted Table 1 (FID 12.97 vs. 5.36 at 4 NFE) compared iSSFM with a semigroup baseline that already used our parameterization (exponential lift, closed-form martingale, drift matching), each trained for up to 200k iterations with the best EMA checkpoint. That baseline corresponds to the "+ drift matching" row of the new ablation (FID 14.67 at 4 NFE), and iSSFM reaches 5.96. The conclusion is the same, i.e., at 4 NFE iSSFM reduces FID by about 2.5× over the semigroup baseline and about 5× over the original SSFM recipe (31.00).

Few-step SDE editing of real images (tMgy, 5V2D). A new Section 4.3 (Figure 5, Table 4; full results and paired comparisons in Appendix H.4, Tables 8–9) edits 240 real ImageNet-256 images at matched NFE. Replaying the source image's recorded Brownian path through iSSFM beats deterministic DMF editing, DMF (denoise–renoise), the teacher SDE at the same NFE, and SSFM at equal preservation of the source. The same network on an independent Brownian path, i.e., using only its transition distributions and cannot make conservative edits, so the improvement comes from the strong property. Figure 5 of the revised PDF shows qualitative edits at 32 NFE.

Brownian representations (bwe6). On the 2D checkerboard we now train one model per representation and coefficient budget with five seeds, and add the global Karhunen–Loève representation with its original Lagrangian objective (Figure 3a). The shifted Legendre representation has the lowest strong error for every budget beyond a single coefficient.

Other changes. We clarify in Section 4.2 that the preference metrics are supplementary to FID and FD-DINOv2 (tMgy), and to make room for the editing experiments we moved the Feynman–Kac corrector experiment to Appendix H.5.

Dear reviewers,

Thank you for your reviews. The author-reviewer discussion closes on November 18, so please respond to the authors' rebuttal before then. Please say whether it addressed your concerns, and update your score if your view has changed.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.