acceptodds
Under review as a conference paper at ICLR 2027

FlowPINN: Physics-Informed Neural Flow Maps for One-Step Generation

29 views

Abstract

Conditional flow matching (CFM) learns a velocity field and then generates every sample with a numerical ODE solve. We replace that pair with a single object: a neural *flow map* whose velocity is recovered by automatic differentiation in and which is trained by minimizing the residual of the flow ODE once, in the manner of a physics-informed neural network (PINN). Generation becomes a single network evaluation , . The integration work moves from sampling time, where it recurs for every sample, to training time, where it is paid once. The construction turns on one distinction that is easy to get wrong — flow matching supplies an unbiased velocity target at the network's *input*, whereas the flow ODE constrains the velocity at its *output*. Two naive substitutions of a flow map into the flow-matching loss that have exactly computable and degenerate minimizers turn to be wrong. We resolve the issue by incorporating the flow ODE residual at network output into the loss, penalized by the terminal conditional at distribution level, and hence propose a physics-informed flow matching based on PINN style solving (FlowPINN). Using benchmark examples of checkerboard and eight-Gaussian, and heavy-tailed funnel distributions, we demonstrate that one-step generation by FlowPINN reaches the accuracy (measured by energy distance) that is worth multiple steps of Euler solver or Runge-Kutta method in the classical CFM. In high-dimensional image generation problems (MNIST and CIFAR-10), FlowPINN matches a hundred-step solve to within the hardware noise band while sampling two orders of magnitude faster, and outperforms the state-of-the-art (SOTA) models including shortcut and meanflow.

Primary area:
generative models
Keywords:
One-step flow matching, flow ODE, physics-informed flow map

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_K8JeOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FlowPINN, a physics-informed flow matching framework designed for one-step (1-NFE) generative modeling. Instead of learning a velocity field and numerically integrating the flow ODE at sampling time, FlowPINN directly parameterizes the solution map of the flow ODE and trains it once by minimizing a balanced physics-informed residual alongside a distributional terminal penalty. Consequently, sample generation requires only a single forward pass without any numerical solver. Empirical evaluations on low-dimensional synthetic distributions, medium-dimensional Neal's funnel, and image benchmarks (MNIST and CIFAR-10) demonstrate that FlowPINN achieves competitive one-step generation quality compared to multi-step ODE solvers and few-step baselines such as MeanFlow and Shortcut Models.

Critical strengths
  1. Solid Mathematical Formulation: The mathematical construction and theoretical analysis of the proposed method are solid, systematically formulating four distinct algorithmic variants (FLOWPINN-P, FLOWPINN-L, FLOWPINN-C, and FLOWPINN-2T).

  2. Simple Core Idea and Clear Presentation: The core idea of learning the flow map by minimizing the ODE residual once during training is straightforward, and the paper is clearly written and easy to follow.

Critical weaknesses
  1. Overly Simple and Low-Resolution Datasets: The image generation experiments are restricted to MNIST (28×2828\times 28) and CIFAR-10 (32×3232\times 32), with a maximum resolution of only 32×3232\times 32. The authors should evaluate the method on higher-resolution benchmarks such as ImageNet-256, CelebA-HQ, and LSUN-Church.

  2. Lack of Latent-Space Generation: The experiments focus exclusively on pixel-space generation, lacking any evaluation or discussion on the widely used latent-space generation setting.

  3. Insufficient Baselines: The compared few-step baselines are overly simple, including only MeanFlow and Shortcut Models. Stronger and more recent one-step models—such as Improved MeanFlow [1], Align Your Flow (AYF) [2], and π\pi-Flow [3]—are missing.

  4. Theoretical Overlap Between FLOWPINN-2T and MeanFlow: The theoretical formulation and core equations of FLOWPINN-2T (Eqs. 18–19) are essentially identical to existing MeanFlow, differing only in the time-axis convention.

  5. Degradation on Mainstream DiT Backbones: As acknowledged in the Limitations section, the method degrades severely (losing 41%–68% in FID) when applied to mainstream DiT architectures.

  6. Contradictory and Unfair Teacher-Guided Comparison on CIFAR-10: While the main text claims FLOWPINN-L trains from scratch without a teacher and achieves 8.69 FID on CIFAR-10 in 29 A100-hours, Appendix D.4 reveals that this result actually relies on a pretrained, frozen CFM teacher (distillation, costing an extra 11 hours), whereas true joint training from scratch reaches only 33.76 FID. This contradicts the main text and makes the comparison against train-from-scratch baselines (MeanFlow at 9.88 and Shortcut at 23.65) unfair.

[1] Geng, Zhengyang, et al. "Improved mean flows: On the challenges of fastforward generative models." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2026.

[2] Sabour, Amirmojtaba, Sanja Fidler, and Karsten Kreis. "Align your flow: Scaling continuous-time flow map distillation." Advances in Neural Information Processing Systems 38 (2026): 146459-146512.

[3] Chen, Hansheng, et al. "pi-flow: Policy-based few-step generation via imitation distillation." International Conference on Learning Representations. Vol. 2026. 2026.

Questions for authors

Please refer to Weaknesses section.

3 comments
AuthorsOpenReview ↗

Comparison to iMeanFlow added

We thank the reviewer for acknowledging the solid math foundation and clear presentation of our paper! Below we try our best to address your concerns.

Weaknesses:

  1. MNIST and CIFAR-10 are standard image benchmarks in generative models. We have used them to demonstrate the numerical advantages of proposed methods over other models in fair comparisons. With available resources, we wish to complete tests on these extra high-dimensional datasets within the given discussion period. However, a paper's quality and contribution to the community should not depend solely on the problem size afforded. For example, Neural ODEs by Chen et al (2018) only reports MNIST, minibatch OT-CFM by Tong et al (2024) studies CIFAR-10 in the pixel space, and Boffi et al. (2025, LSD), Kim et al. (2024, CTM) also use the same datasets or images at this level.

  2. This is a great suggestion and we have acknowledged in the revision. We understand latent-space is a powerful technique to scale up applications that can be applied to benefit our work. However, the main purpose for this paper is to establish single-time flow map for one-time generation by adopting PINN-style solving. Given the existing contents and the capacity limit, we think it would be a good future direction.

  3. Thanks for pointing to the references. While we congratulate these authors on publishing their works on top conferences, we notice that they appear in this year. For example, "improved mean flows (iMeanFlow)" got accepted only three months ago! We have managed to add iMeanFlow in the revision. Below we quote iMeanFlow rows from Tables 1, 4, and 5 in the revision.

Table 1: MNIST on the UNet backbone

MethodNFEFID ↓clf-FID ↓IS ↑MEMO ≈ 1ms/1k ↓Params
CFM + midpoint10024.23 ± 0.120.79 ± 0.142.041 ± 0.0150.995 ± 0.004126927.28M
FlowPINN-L124.29 ± 0.751.60 ± 0.212.034 ± 0.0090.997 ± 0.00410914.49M
iMeanFlow130.62 ± 2.089.88 ± 0.982.086 ± 0.0081.001 ± 0.0051177.29M
iMeanFlow424.92 ± 0.4610.13 ± 1.052.060 ± 0.0071.000 ± 0.0054637.29M

Table 4: The two-dimensional benchmarks.

MethodNFECheckerboard ED ↓Checkerboard MEMO ≈ 1Checkerboard off-supp. ↓8-Gaussians ED ↓8-Gaussians MEMO ≈ 18-Gaussians in-mode ↑
CFM + RK41000.0021 ± 0.00091.050.0850.0036 ± 0.00180.990.983
FlowPINN-L10.0024 ± 0.00081.120.1620.0039 ± 0.00211.010.934
iMeanFlow10.0490 ± 0.00721.030.0850.0485 ± 0.02730.880.964
iMeanFlow20.0439 ± 0.00681.040.0780.0594 ± 0.02490.920.982
iMeanFlow40.0362 ± 0.01061.020.0620.0848 ± 0.02760.960.985

Table 5: Neal's funnel distribution.

MethodNFEd=10 ED ↓d=10 |Δ log p| ↓d=10 neck KS ↓d=100 ED ↓d=100 |Δ log p| ↓d=100 neck KS ↓
CFM + RK41000.0019 ± 0.00074.70.0190.2672 ± 0.18993750.031
FlowPINN-L10.0039 ± 0.00276.60.0270.2774 ± 0.038417.20.180
iMeanFlow10.1006 ± 0.01121.50.2310.2375 ± 0.019626190.389
iMeanFlow20.1024 ± 0.01493.40.2320.1580 ± 0.01768390.331
iMeanFlow40.1105 ± 0.01593.30.2370.1633 ± 0.020410930.323
  1. The connection and distinction between FlowPINN-2T and MeanFlow was in Appendix B of the submission and has been moved to Section 3.4. Indeed, we did NOT accredit the MeanFlow identity (18) to ourselves. On the contrary, we only regard FlowPINN-2T as a two-time relaxation of our single-time flow map that nicely connects to MeanFlow. A paragraph in Appendix B of revision now describes the implementation differences between FlowPINN-2T and MeanFlow.

  2. As commented in the response to the first reviewer, the models' performance depends on the capacity of the backbone. First, In Table 6 (old Table 5), FlowPINN methods still beat MeanFlow and Shortcut at NFE=1 on DiT-6 (6 blocks) backbone even there is some gap to CFM-100NFE. Second, in new Table 7, increasing the size by 5050\\%, a deeper DiT-9 (9 blocks) improves all results of DiT-6 and closes the gap by up to 4646\\%. Finally, among models being compared in the paper, UNet is shipped in both MeanFlow (torch version) and LSD; while Shortcut and iMeanFlow use DiT -- we do not judge which is mainstream but UNet is not something used only by us.

  3. We acknowledge this issue and have re-run the CIFAR-10 experiments with no teacher, at the same budget and with 5 random seeds. Please refer to Table 2 (old Table 6) for the updated results and responses to other two reviews.

We hope our answers address your concerns and we will be happy to take any follow-up questions.

Thank you very much for your responses and the additional experiments in the rebuttal. I appreciate the PINN-based perspective for formulating single-time flow maps, as well as your efforts in re-running CIFAR-10 without a teacher, so I will raise my score to 2.

That being said, while the high-level core idea is intuitive, I still consider the current manuscript rather rough in execution and in need of further polishing. For example, there are clear formatting issues with the figures: 1:1 square CIFAR-10 images in Figure 5 and plots in Figure 3 are squashed out of proportion, leaving both the images and the text inside visibly flattened.

Most importantly, my primary concern remains the lack of evaluation at mainstream higher resolutions such as ImageNet-256. To convincingly demonstrate the superiority of your method, you either need to comprehensively reproduce recent strong SOTA baselines under your own experimental setup, or scale up to standard higher-resolution benchmarks where you can directly compare against official numbers reported in the literature. Right now, on your highest-dimensional benchmark CIFAR-10 (Table 2), only MeanFlow and Shortcut are compared—even the newly added iMeanFlow and LSD are absent there, not to mention recent strong baselines like alpha-Flow, AYF and π\pi-Flow—and FlowPINN-L (7.33 FID) still slightly trails MeanFlow (7.06 FID) at 1-NFE. As for citing earlier general ODE/flow works like Neural ODE (2018) or minibatch OT-CFM, those predated today's standardized evaluation consensus for one-step flow-map generation. In the current one-step/few-step generation literature (including Shortcut, MeanFlow, iMeanFlow, AYF, and π\pi-Flow), ImageNet benchmarks are well-established and mature, so I strongly encourage you to include them in future revisions.

AuthorsOpenReview ↗

iMeanFlow and LSD for CIFAR-10 added

Thank you for raising your score! We also appreciate your acknowledgment of our effort for improvement.

We agree with you on the figures so we have fixed them in the revision.

We understand your concern so we have added the missing puzzles -- more recent iMeanFlow and LSD results to Table 2 (old Table 6), with relevant rows quoted below:

Table 2: CIFAR-10 on the UNet backbone

MethodNFEFID ↓clf-FID ↓IS ↑memo ≈ 1ms/1k ↓Params
CFM + midpoint1003.01 ± 0.030.08 ± 0.019.635 ± 0.0251.000 ± 0.0035175655.68M
FlowPINN-L17.33 ± 0.030.48 ± 0.038.481 ± 0.0330.999 ± 0.00252355.57M
iMeanFlow113.53 ± 0.060.36 ± 0.018.785 ± 0.0280.998 ± 0.00251655.74M
iMeanFlow47.96 ± 0.040.26 ± 0.019.405 ± 0.0621.000 ± 0.003205955.74M
LSD114.69 ± 0.050.82 ± 0.017.841 ± 0.0320.999 ± 0.00251655.74M
LSD44.90 ± 0.030.22 ± 0.018.893 ± 0.0240.999 ± 0.002205955.74M

We can see that both iMeanFlow and LSD still trail FlowPINN-L at NFE=1. iMeanFlow at NFE=4 becomes comparable with FlowPINN-L; LSD at NFE=4 beats FlowPINN-L and is comparable with FlowPINN-2T at NFE=4. All the comparisons are done at the same model size described in Table 2 and may be different from their published numbers. We have new Table 8 for their training recipe in the revision to make it transparent.

We have made our best efforts to complete the comparison with 4 SOTA models on all examples including image benchmarks. Please note that iMeanFlow was published only 4~5 months ago and it took 50~70 hours to train each of these models on CIFAR-10 at the budget of 440K steps. They are models all in the pixel space so training takes time. We have ImageNet running with good trend on the recorded logs; however, it is less likely to finish within the discussion period.

We agree with the reviewer that results on larger datasets may further strengthen the numerical evidences. However, we have demonstrated the numerical advantages of the proposed FlowPINN methods with sufficient proof. We hope the reviewer can appreciate our new results and agree with us on this. Thanks again for your nice suggestions for improving this work!

Reviewer_kUmzOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes FlowPINN, which learns a flow map by minimizing an ODE residual, enabling one-step generation. It introduces learned-denoiser and shared-network variants, a separable time parametrization, and evaluations on synthetic distributions, MNIST, and CIFAR-10.

Critical strengths

S1. The analysis of two naive regression formulations provides useful insight into why they fail to learn the intended flow map.

S2. The separable time parametrization reduces the cost of computing time derivatives.

Critical weaknesses

W1. Unclear methodological novelty.

The core residual closely resembles Lagrangian map distillation [1] with a fixed starting time. The Introduction’s contrast between prior composition-based methods and the proposed ODE-residual formulation therefore overlooks an existing residual-based approach. The paper should clarify and isolate the contributions of denoiser learning, weighting, terminal loss, and separable parametrization. Appendix B also identifies FlowPINN-2T’s objective with MeanFlow [3], but the implementation differences explaining their separate results remain unclear.

W2. Insufficient evidence for comparative superiority.

The main CIFAR-10 FlowPINN-L result uses a pretrained CFM field. Although Appendix D.4 discloses its additional 11-hour training cost, the main-text comparison reports only 29 hours against MeanFlow’s 33. MeanFlow’s reported FID of 9.88 also differs substantially from the original 2.92 [3]; this is not necessarily an error, since training budgets differ (200k versus 800k iterations), but the recipe differences warrant clarification. Direct comparisons with closely related LMD/LSD methods [1, 2] are missing, and image experiments are limited to MNIST and single-seed CIFAR-10.

W3. Inconsistent reporting and overstated conclusions.

Table 6 reports FlowPINN-L/CFM FIDs of 8.69/3.01, whereas Appendix D.4 gives 9.38/3.25. FlowPINN-L’s clf-FID likewise differs between the table and text (0.38 versus 0.52), without an explanation of different settings. The abstract’s broad claim of matching a 100-step solver on image benchmarks also conflicts with the substantial CIFAR-10 gap, which §4.4 explicitly acknowledges.

W4. Minor presentation issues.

Figure 6 appears vertically compressed, reducing readability. Training settings, variant differences, and cost accounting are scattered across sections, making it difficult to connect each result to its precise configuration.

References

[1] Boffi, Albergo, and Vanden-Eijnden. Flow Map Matching with Stochastic Interpolants: A Mathematical Framework for Consistency Models. 2024; revised 2025. https://arxiv.org/abs/2406.07507

[2] Boffi, Albergo, and Vanden-Eijnden. How to Build a Consistency Model: Learning Flow Maps via Self-Distillation. NeurIPS 2025. https://arxiv.org/abs/2505.18825

[3] Geng et al. Mean Flows for One-Step Generative Modeling. NeurIPS 2025. https://arxiv.org/abs/2505.13447

Questions for authors

Please refer to the weaknesses above.

3 comments
AuthorsOpenReview ↗

Comparison to LSD added

Thanks to the reviewer for highlighting the insight we put in the single-time formulation of flow-map by analyzing two failure modes and acknowledging the benefit of separable time parametrization in reducing the expensive JVP computation. Below is the point-to-point response to your concerns.

Weaknesses

W1: The biggest the difference from two-time flow-maps including Lagrangian map distillation (LMD) [1] and Lagrangian self-distillation (LSD) [2] is that our flow-map is a single-time formulation, which we regard as a more primitive object. As commented in the response to the first reviewer, different from the two-time flow-maps, the theoretic challenge of single-time formulation lies in where and how the velocity field is anchored in the flow-map network (refer to Figure 1 and the paragraph below Proposition 2). Additionally, even sharing similar ODE-residuals, both LMD and LSD have pointwise boundary conditions which is doomed to fail in the single-time setting and the terminal condition has to be imposed in the distribution level (refer to answer of Q3 in the first review) -- this has been further highlighted in the 'Relation to few-step and flow-map models' paragraph of the revision. Lastly, we very much appreciate the smart works of [1][2] and they had been cited in the introduction (line 44) and literature review (line 69-70) together with other two-time flow-maps so existing approaches in the literature were NOT overlooked. To emphasize that, we have added comparison with LSD (LMD requires a known teacher btb_t so LSD is more relevant) to the revision (see W2 below). FlowPINN-2T is regarded as a two-time relaxation of our single-time formulation also nicely connected to MeanFlow, which we acknowledged in the Appendix B in the submission and has been moved to Section 3.4 in the revision. A paragraph in Appendix B of revision now describes the implementation differences between FlowPINN-2T and MeanFlow.

W2: Thanks for pointing out the issue of CIFAR-10, as did by other reviewers. We acknowledge that and have re-run the CIFAR-10 experiments with no teacher, at the same budget and with 5 random seeds. Inconsistent numbers have also been fixed in the revision. We have added new results of LSD to Tables 1, 4, and 5 in the revision also quoted below.

Table 1: MNIST on the UNet backbone

MethodNFEFID ↓clf-FID ↓IS ↑MEMO ≈ 1ms/1k ↓Params
CFM + midpoint10024.23 ± 0.120.79 ± 0.142.041 ± 0.0150.995 ± 0.004126927.28M
FlowPINN-L124.29 ± 0.751.60 ± 0.212.034 ± 0.0090.997 ± 0.00410914.49M
LSD129.27 ± 1.533.59 ± 0.782.075 ± 0.0140.998 ± 0.0031187.29M
LSD424.32 ± 0.422.46 ± 0.552.055 ± 0.0140.996 ± 0.0024677.29M

Table 4: The two-dimensional benchmarks.

MethodNFECheckerboard ED ↓Checkerboard MEMO ≈ 1Checkerboard off-supp. ↓8-Gaussians ED ↓8-Gaussians MEMO ≈ 18-Gaussians in-mode ↑
CFM + RK41000.0021 ± 0.00091.050.0850.0036 ± 0.00180.990.983
FlowPINN-L10.0024 ± 0.00081.120.1620.0039 ± 0.00211.010.934
LSD10.0024 ± 0.00081.080.1180.0042 ± 0.00241.050.931
LSD20.0023 ± 0.00081.070.1150.0041 ± 0.00241.040.942
LSD40.0022 ± 0.00091.050.0920.0039 ± 0.00231.040.965

Table 5: Neal's funnel distribution.

MethodNFEd=10 ED ↓d=10 |Δ log p| ↓d=10 neck KS ↓d=100 ED ↓d=100 |Δ log p| ↓d=100 neck KS ↓
CFM + RK41000.0019 ± 0.00074.70.0190.2672 ± 0.18993750.031
FlowPINN-L10.0039 ± 0.00276.60.0270.2774 ± 0.038417.20.180
LSD10.0067 ± 0.00256.40.0410.8706 ± 0.27198690.066
LSD20.0034 ± 0.00144.90.0240.4992 ± 0.34724800.076
LSD40.0023 ± 0.00094.60.0230.3988 ± 0.33454580.056

W3: Thanks for your careful reading. These inconsistent records have been corrected in the revision. The overstated the claim has also been fixed in the abstract.

W4: We have increased the height of Figure 6 in the revision.

We hope our new results convince you that our FlowPINN has theoretic novelty that is distant from existing two-time flow maps and solid numerical advantages. We will be happy to engage in the discussion and take any follow-up questions.

Reviewer_kUmzOpenReview ↗

Response to rebuttal

Thank you for the response and the additional experiments, especially the teacher-free CIFAR-10 re-run and the new LSD comparisons.

However, the following concerns remain unresolved:

  • Novelty relative to LMD/LSD: FlowPINN-L/C remain close to LMD/LSD restricted to s=0. The claimed distinction does not hold, since LSD imposes no pointwise terminal condition. The distributional terminal loss helps only on the funnel (Table 8) and could equally be added to existing flow-map methods.
  • CIFAR-10 evidence: LSD/iMeanFlow are missing from Table 2, FlowPINN-L does not outperform MeanFlow at NFE=1, and the training recipe and the gap to the original MeanFlow result are not explained.
  • Claims: The abstract still claims that FlowPINN matches “an up to hundred-step solve to within the hardware noise band” on MNIST and CIFAR-10. This holds only for MNIST; on CIFAR-10 the gap to the 100-step solve remains large (7.33 vs. 3.01).

I will therefore maintain my score.

In addition, Figures 2 and 3 are still poorly made in the revision. Figure 2 merely restates Section 3.3 and adds no information as a figure. The overall quality of the figures needs improvement.

I hope this feedback helps improve the paper in its next version.

AuthorsOpenReview ↗

Connection/Distinction to LSD added in Appendix E; iMeanFlow and LSD results added for CIFAR-10

Thanks for acknowledging our new results. We appreciate your constructive comments and understand your concerns for which we try to address with new results.

  • Novelty relative to LMD/LSD: We agree with you that the residual in (13) is the Lagrangian condition (Boffi et al 2024; 2025) restricted to s=0s=0. However, one cannot simply regard FlowPINN as LSD at slice s=0s=0. To clarify the connection and distinction between these two, we have (all changes are marked in red in the revision):

    1. improved the claim in literature review to be more accurate: with particular Lagrangian self-distillation (LSD) (Boffi et al 2025) having loss as a double-time integral, in contrast to ours a single-time integral plus a distributional terminal condition;
    2. acknowledged the connection under Equation (14): the residual agrees with Lagrangian condition (Boffi et al 2024; 2025) on the slice s=0s=0; with the denoiser it is (online) Lagrangian map distillation restricted to one-step generation, rather than self-distillation across the whole triangle s≤ts\leq t;
    3. added a whole Appendix E with added ablation study on LSD variants on 2×22\times 2 grid: slice/triangle domain x self/separate teacher in Table 9.

First of all, LSD at s=0s=0 does not exist as a method. By its self-distillation mechanism, LSD has a teacher as the diagonal vt,tv_{t, t} of the same two-time network trained by flow matching on s=ts=t. Keeping only the rows on the slice s=0s=0 removes the self-teacher and leaves only Lagrangian map distillation (LMD) (Boffi et al 2024) on one slice with a teacher trained online. Second, we numerically investigate the behavior of LSD variants on the 2×22\times 2 grid where (triangle, self-teacher) is the original LSD and (slice, separate teacher) resembles our FlowPINN-L (regardless of terminal condition). Table 9 in the revision, together with evidences by other tables, highlight three major differences:

  1. Moving the Lagrangian rows from the triangle (LSD's defining domain) to the slice (FlowPINN's defining domain) consistently lowers the one-evaluation distance.
  2. FlowPINN-L, with separable basis network and balanced residual form, is more effective and stable than LDS restricted to s=0s=0.
  3. FlowPINN-L has better one-step generation performance than LSD.

In conclusion, the connection and distinction with LSD does not diminish the novelty and creativity, but strengthen the relationship of FlowPINN as a single-time flow map and existing two-time flow map literature.

  • CIFAR-10 evidence: We have added iMeanFlow and LSD rows for CIFAR-10 results to Table 2 (old Table 6) -- the results based on our setting (backbone, batch size, etc) show that they trail FlowPINN-L trained at the same budget. Table 8 of the training recipe has been added in the revision for transparency.

  • Claims: We have swapped the describing terms for vector benchmarks and image applications, which is more accurate now: "Using benchmark examples of checkerboard ... that is worth a hundred steps ... In high-dimensional image generation ... FlowPINN-L matches multiple-step solve".

We agree with you that Figure 2 adds no more information so we have moved it back to Appendix. All other figures that were squashed have been fixed for better visualization.

Thanks again for all your comments and we believe that they have been very useful in improving this work. We hope our clarification on the connection/distinction to LSD and added experimental results have addressed your concerns.

Reviewer_1mDCOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper puts forward FlowPINN, a one-step framework that learns neural flow maps without ODE integration at inference. Inspired by physics-informed neural networks, it enforces the initial condition and minimizes an ODE residual with terminal distribution matching, shifting computation to training. To estimate marginal velocity at the map output, the authors introduce a denoiser and efficient variants using separable time representations and shared architectures. They also explain why two naive flow-matching adaptations fail and discuss the connection to MeanFlow through a two-time formulation. Experiments on synthetic distributions, MNIST, and CIFAR-10 show competitive one-step generation and sampling efficiency compared with ODE solvers and few-step baselines. FlowPINN approaches multi-step integration quality on low-dimensional benchmarks and MNIST. However, its strongest CIFAR-10 results depend on a pretrained denoiser, and performance varies across architectures.

Critical strengths

The main strength of this paper is its interesting perspective on one-step generation, directly learning neural flow maps through ODE residual minimization instead of numerical integration during sampling. This provides an alternative to conventional flow matching by shifting computational costs from inference to training. Another positive aspect is the theoretical analysis of two straightforward flow-map formulations, which explains why naive adaptations fail and motivates the proposed approach. The authors also introduce practical designs, including separable time representations and shared-network architectures, to improve computational efficiency and reduce memory consumption. Experiments on synthetic distributions, MNIST, and CIFAR-10 demonstrate competitive one-step generation performance against existing methods such as MeanFlow and Shortcut. The results are particularly promising on low-dimensional benchmarks and MNIST, where FlowPINN approaches multi-step ODE solver quality with a single network evaluation.

Critical weaknesses

My main concern is that the experimental results do not fully support FlowPINN's claimed advantages. First, its best CIFAR-10 result (FID 8.69) relies on a pretrained denoiser, whose additional training cost is not fully reflected in the comparison. The jointly trained variant achieves a much worse FID of 33.76, raising concerns about its effectiveness when trained from scratch. Comparisons under matched total training budgets would be more convincing. Second, FlowPINN performs well with UNet but degrades considerably with DiT. Although the authors attribute this to the separable time representation, further evidence is needed to explain this architectural sensitivity. Third, the theoretical contribution is somewhat limited, as the formulation largely adapts existing PINN techniques to flow matching. While the analysis of two unsuccessful formulations is useful, it does not establish clear theoretical advantages over existing one-step methods. Finally, the evaluation focuses on relatively small-scale datasets, and CIFAR-10 results use only a single random seed, making robustness and scalability difficult to assess.

Questions for authors
  1. The best CIFAR-10 result (FID 8.69) relies on a pretrained denoiser, while the jointly trained variant achieves a much worse FID of 33.76. It is necessary to explain this substantial performance gap and provide comparisons under matched total training budgets, or otherwise the contribution is limited.

  2. FlowPINN performs well with UNet but shows considerable performance degradation with DiT. Could you investigate the reasons for this architectural sensitivity and whether alternative time parameterizations could improve its generalizability?

  3. The proposed objective combines an ODE residual with terminal distribution matching. Why not provide ablation studies to clarify the individual contributions of these two components and demonstrate the necessity of the PINN-based formulation?

  4. The CIFAR-10 experiments are conducted with only one random seed, and the evaluation is limited to relatively small-scale datasets. For a more fair comparison, it is better to report results across multiple seeds and evaluate FlowPINN on more challenging generation benchmarks to demonstrate its robustness and scalability.

1 comment
AuthorsOpenReview ↗

CIFAR-10 experiments re-run

We thank the reviewer for appreciating the new perspective of our paper on single-time formulation of flow-map for one-step generation, elucidation on the theoretic challenges with analysis of two failure modes, and effort of practical algorithm designs. We understand your concerns and have re-run the experiments to be clarified below.

Weaknesses

We agree with the reviewer that in CIFAR-10 training FlowPINN-L with a frozen desnoiser goes against the claim and makes unfair comparison to alternatives regarding the time budget. Therefore, in the revision, we have re-run the CIFAR-10 experiments with no teacher, at the same budget and with 5 random seeds. The FID 33.76 was an obsolete number in the submission and has been fixed, together with a few inconsistent records (See below answer to Q1). Regarding the performance under UNet vs DiT, the order of different algorithms was consistent in different backbones and new experimental results indicate the gap to CFM-100NFE in DiT was attributed to the its capacity -- increasing DiT by 5050\\% in size helps closing the gap by up to 4646\\%. (See below answer to Q2). We agree with that it appears straightforward to combine FM with PINN; however, different from the two-time formulation, the theoretic challenge lies in where and how the velocity field is anchored in the flow-map network -- that is what we struggled through and we thank for the reviewer's appreciation on that. Finally, we hope the evaluation results under random seeds in CIFAR-10 strengthen the evidences and ease your concerns.

Questions:

Q1: We have re-run the experiments of CIFAR-10 with all algorithms at the same budget and with random seeds. Table 2 (old Table 6) shows that at NFE=1, FlowPINN-L, reaches FID 7.33±0.037.33\pm 0.03 against MeanFlow's 7.06±0.037.06\pm 0.03 and Shortcut's 23.06±0.1223.06\pm 0.12; for NFE=4, FlowPINN-2T has the lowest FID 4.40±0.044.40\pm 0.04. Thanks to the separable basis representation, FlowPINN-L trains at shorter time compared with MeanFlow and FlowPINN-2T that require expensive JVP computation.

Q2: First of all, Table 6 (old Table 5) presents results using DiT backbone with 6 blocks only for reference. Even in this table, FlowPINN methods still beat MeanFlow and Shortcut at NFE=1. The gap to CFM-100NFE is explained by the capacity of DiT-6. In the following table of FID scores in MNIST (also Table 7 in the revision), the last column shows a deeper DiT with 9 blocks improves all results of DiT-6 and helps closing such gap by up to 4646\\%.

Table 7: Contrasting UNet and DiT backbones on MNIST

MethodNFEUNet (7.3M) ↓DiT (6 blocks, 7.3M) ↓DiT (9 blocks, 10.9M) ↓Change
CFM459.50 ± 1.0364.36 ± 1.0864.83 ± 1.31+0.7%
CFM10024.23 ± 0.1226.16 ± 0.1624.81 ± 0.22-5.2%
FlowPINN-P1209.6 ± 3.0213.2 ± 2.9213.0 ± 2.2-0.1%
FlowPINN-L124.29 ± 0.7540.88 ± 1.2334.96 ± 0.75-14.5%
FlowPINN-C131.06 ± 1.3047.16 ± 1.5943.20 ± 2.70-8.4%
FlowPINN-2T126.14 ± 0.3039.70 ± 2.2532.16 ± 1.32-19.0%
FlowPINN-2T424.49 ± 0.5930.18 ± 0.7627.77 ± 0.32-8.0%
MeanFlow139.49 ± 3.1555.56 ± 2.3748.06 ± 0.95-13.5%
MeanFlow426.29 ± 0.5831.11 ± 1.1529.29 ± 0.50-5.8%
Shortcut144.33 ± 3.5972.36 ± 1.2662.38 ± 1.16-13.8%
Shortcut433.93 ± 3.1948.94 ± 0.3941.89 ± 0.94-14.4%

Q3: Thanks for your great suggestion! The following table (also new Table 8 in the revision) compares the energy distances of FlowPINN-L, our flagship model, under different conditions. As commented in the main text (line 254 in the revision), poitwise terminal condition (fourth column) collapses as source-indexed regression (last column).

Table 8: Terminal-loss ablation study

targetno terminal in Eq. (14) (λ_term = 0)+ distributional terminal (λ_term = 1)+ pointwise terminal φ₁(x₀) = x₁ (λ_term = 1)ref.: source-indexed map (pointwise only)
checkerboard (d=2)0.0024 ± 0.00070.0024 ± 0.00091.8041 ± 0.01241.9217 ± 0.0139
8-Gaussians (d=2)0.0039 ± 0.00200.0037 ± 0.00202.7234 ± 0.00692.8650 ± 0.0033
funnel (d=10)0.0039 ± 0.00250.0019 ± 0.00080.4827 ± 0.01330.5041 ± 0.0123
funnel (d=100)0.2774 ± 0.03640.1048 ± 0.00840.6301 ± 0.01720.5224 ± 0.0154

Q4: As commented in Q1, the results have been randomized with 5 different seeds.

We hope our answers help clarify the concerns and we will be happy to take any further questions.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of revision

We thank all three anonymous reviewers for their careful reading and constructive comments and suggestions on improving this paper! We also appreciate their acknowledgment on the novelty of a single-time flow-map for one-step generation, model effort and clarity of presentation.

This paper aims to learn a single-time flow-map directly parametrized by neural network trained in the manner of PINN and to replace repeated ODE integration in the sampling stage with a cheap network evaluation. Unlike most two-time formulations in the literature, the theoretic challenge lies in where and how the velocity field is anchored in the flow-map network. A two-time relaxation nicely connects to the recent MeanFlow model.

We acknowledge the issue of CIFAR-10 experiments brought up by all the reviewers, performance parity on backbones, and distance to and comparison with existing works. To address all these concerns, in the revision we have mainly:

  • re-run the CIFAR-10 experiments with no teacher, at the same budget and with 5 random seeds and have new results in Table 2 (old Table 6).
  • added new results (Table 7) with a deeper DiT to explain the dependence of models' performance on the capacity of backbones.
  • explained the distinction of the proposed FlowPINN with LSD (regarding the defining domains in Table 9 and the terminal condition in Table 10) and MeanFlow (regarding the algorithmic and implementation differences in Section 3.4 and Appendix B).
  • added new numerical results in comparison with LSD and more recent iMeanFlow (in Tables 1, 2, 4 and 5) and illustrated superiority and competitiveness.

All changes have been marked in red in the revision. We hope our clarification and added numerical evidences have strengthened the rigor of our proposed FlowPINN as a novel and useful single-time, as opposed to two-time, flow-map in the one-step generative models. We thank all the reviewers for the great suggestions that have significantly improved the quality of this work and welcome any follow-up questions.

Area_Chair_YaCcOpenReview ↗

Engage in discussion

The authors and reviewers are requested to engage in the discussion phase.

Authors: please respond to the review comments.

Reviewers: please review the authors’ rebuttal and consider revising your assessment/score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.