acceptodds
Under review as a conference paper at ICLR 2027

Physics-Informed One-Step Generation on Long-Short Flow Maps

Abstract

We propose Physics-Informed Long-Short Flow Maps with Direct Guidance (PI-LSFM), a one-step generative method for producing data-aligned and physically valid samples. Existing one-step physics-informed generative models must jointly control data fidelity and physical constraints without the iterative correction of multi-step samplers, making it difficult to coordinate indirect surrogate supervision with strong physics enforcement. PI-LSFM addresses this challenge by decomposing generation into a long flow map for global transport and a near-terminal short flow map, where closed-form data and physics potentials provide directly computable guidance targets instead of indirect regression surrogates. The short map guides the long map during training, yielding a single one-step generator that requires neither staged distillation nor inference-time correction. Across Darcy flow, Kolmogorov flow, dynamic stall, and topology optimization, PI-LSFM achieves a favorable trade-off among distributional fidelity, physical residuals, and inference efficiency while using only one function evaluation at inference.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Physics-informed generative modeling, Long-short flow maps, One-step generative models, Flow matching, Scientific machine learning

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

This paper extends the Drifting Model framework to physics-constrained generation tasks and interprets its training objective from a Long-Short Flow-Map perspective. The experiments cover multiple tasks and show that the method achieves competitive sample quality and physical consistency compared with multi-step methods while using one-step generation. Although concurrent work has already proposed a Long-Short Flow-Map interpretation of Drifting Models, I still consider this paper’s extension to physics-constrained generation practically valuable, with sufficiently extensive experiments supporting the method’s effectiveness.

Critical strengths
  1. A meaningful extension to physics-constrained generation. The paper extends the Drifting Model framework by incorporating physics-based corrections into the training targets, addressing the practical need to generate samples that are both distributionally accurate and physically consistent.

  2. Broad experimental coverage. The evaluation includes Darcy flow, Kolmogorov flow, dynamic stall, and topology optimization, covering unconditional generation, conditional generation, and inverse design. Comparisons with representative multi-step methods and additional one-step baselines on Darcy provide substantial evidence of the method’s effectiveness.

  3. Systematic design studies. The paper investigates different numerical schemes for the data and physics terms, physics-guidance weights and schedules, and feature encoders. These studies support the chosen configuration and provide useful insights into applying the framework to physical tasks.

  4. Competitive performance with efficient inference. The method achieves a favorable trade-off between distributional fidelity, physical consistency, and inference efficiency using a single network evaluation, without requiring a pretrained generative teacher or inference-time correction. Reporting wall-clock inference time alongside function evaluations makes the efficiency benefits more concrete.

Critical weaknesses

The comparison of physics-correction schemes is limited to the Darcy task. Although FE achieves better overall performance than BE and BE-Lim in this setting, it remains unclear whether this advantage holds for other physical systems considered in the paper. The results support the choice of FE for Darcy, but provide limited evidence about the relative performance of these schemes across tasks.

Questions for authors

Have the authors compared FE, BE, and BE-Lim on any of the other tasks? Such a comparison would help clarify whether the advantage of FE extends beyond Darcy or depends on the particular physical system.

The potential-based long-short flow-map formulation appears closely related to the kernel-based vector-field construction in Drifting Models. Could the authors clarify this relationship by expressing the resulting data and physics guidance in a common vector-field formulation? In particular, it would be helpful to explain which components correspond to kernel-based attraction and repulsion, and whether the long-short flow-map perspective introduces distinct update rules or primarily provides an alternative derivation and interpretation. This connection could make the method easier to understand.

1 comment
AuthorsOpenReview ↗

Response to Reviewer nLsw

Thank you for the encouraging review. Both questions helped us make the paper clearer.

Q2: A common vector-field formulation; relation to the kernel-based attraction and repulsion of Drifting Models.

Following this suggestion, the revised paper now writes the target in this form (Sec. 3.3; Proposition C.2 in App. C.4). With x=ψ0→1θ(x0)x=\psi^\theta_{0\to1}(x_0) and qθ=[ψ0→1θ]♯p0q_\theta=[\psi^\theta_{0\to1}]_\sharp p_0, the target we use (TR for data, FE for physics) is

T(x0)=12(Ex1∼p1[x1k1ϵ(x1,x)]Ex1∼p1[k1ϵ(x1,x)]−x)⏟attraction to data+12(x−Ex~∼qθ[x~k1ϵ(x,x~)]Ex~∼qθ[k1ϵ(x,x~)])⏟repulsion from generated samples−λˉλJR⊤R(x)⏟physics.\mathcal T(x_0)=\underbrace{\tfrac12\Big(\frac{\mathbb E_{x_1\sim p_1}[x_1k_1^\epsilon(x_1,x)]}{\mathbb E_{x_1\sim p_1}[k_1^\epsilon(x_1,x)]}-x\Big)}_{\text{attraction to data}}+\underbrace{\tfrac12\Big(x-\frac{\mathbb E_{\tilde x\sim q_\theta}[\tilde xk_1^\epsilon(x,\tilde x)]}{\mathbb E_{\tilde x\sim q_\theta}[k_1^\epsilon(x,\tilde x)]}\Big)}_{\text{repulsion from generated samples}}-\underbrace{\bar\lambda\lambda J_{\mathcal R}^\top\mathcal R(x)}_{\text{physics}}.

The attraction term is the Forward Euler step of the data potential and the repulsion term is its Backward Euler step. The trapezoidal rule averages the two, which gives the attraction–repulsion field of a drifting-style update. For the physics term, Forward Euler is the residual gradient, and because the target sits under a stop-gradient it contributes exactly the gradient of the terminal loss λˉλ∥R(x)∥2\bar\lambda\lambda\|\mathcal R(x)\|^2.

In raw space this configuration is therefore equivalent to a drifting-style objective plus a terminal physics loss, in the sense that the two have the same training gradient. When the encoder is used, the physics step enters through ϕ(x+λˉTphys)\phi(x+\bar\lambda\mathcal T_{\rm phys}) (Eq. 13).

The long-short perspective also introduces distinct update rules. We tested them on Darcy, each at its best λˉ\bar\lambda or schedule (Fig. 4c, 4d; App. D):

Data rule (physics = FE), 100K stepsFieldRE ↓WD·10² ↓
FEattraction only8.1254.244
BErepulsion only1.8534.857
TRattraction + repulsion0.1170.213
Physics rule (data = TR), 400K stepsFieldRE ↓WD·10² ↓
FEresidual gradient0.0550.108
BEGauss–Newton step −(J⊤J+γE)−1J⊤R-(J^\top J+\gamma E)^{-1}J^\top\mathcal R0.4360.122
BE-Limprojection −J⊤(JJ⊤)−1R-J^\top(JJ^\top)^{-1}\mathcal R2.5400.184

Here J=JR(x)J=J_{\mathcal R}(x) and γ=1/(2λ)\gamma=1/(2\lambda), and the two bold rows are the same configuration at 100K and 400K steps.

The rule that coincides with attraction–repulsion plus the residual gradient is the best of the family, and the perspective accounts for it: the data term needs the second-order rule and the physics term the explicit one. The perspective also gives a direct data target, which is what separates PI-LSFM from physics-informed consistency, MeanFlow and shortcut training (RE 0.055 against 0.240, 4603.573 and 0.435 with the same terminal physics loss; Table 4). In these two respects it is more than an alternative derivation.

W1, Q1: FE, BE and BE-Lim on other tasks.

We compared the three rules on Darcy only and kept FE on the other three benchmarks without a per-task search (App. G.4). The residuals there are of different types, for example a divergence constraint on Kolmogorov flow and point-wise algebraic relations on dynamic stall. With FE, PI-LSFM has the lowest physical residual on both: RE·10¹ is 0.404 against 1.362 for the best 20-step baseline on Kolmogorov flow, and RE·10⁶ is 0.00015 against 0.143 on dynamic stall. The rule selected on Darcy thus remains effective on other physical systems.

FE also has a computational advantage that does not depend on the system. It needs only one vector–Jacobian product, whereas BE and BE-Lim require a linear solve with JR⊤JRJ_{\mathcal R}^\top J_{\mathcal R} or JRJR⊤J_{\mathcal R}J_{\mathcal R}^\top, whose conditioning depends on the residual. On Darcy this makes FE the cheapest rule, at 0.271 s per step against 0.388 for BE and 0.376 for BE-Lim.

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

The method decomposes the full flow map into a long flow map for global transport and a short flow map near the terminal time. The short flow map provides directly computable guidance from data-distribution and physics potentials, with the goal of alleviating the indirect supervision and optimization difficulties encountered when jointly enforcing distribution matching and physical constraints in one-step generative models.

Critical strengths
  1. One-step generation under physical constraints is challenging because the model must simultaneously produce samples that match the target distribution and satisfy physical requirements, without relying on iterative sampling-time corrections.

  2. The long-short flow-map decomposition provides a structured way to introduce data and physics guidance near the terminal time. The motivation for reducing indirect supervision is reasonable, and the proposed framework is worth investigating.

  3. The paper considers several distinct physical generation problems rather than relying on a single benchmark.

Critical weaknesses
  1. Physics-informed losses and the incorporation of physical constraints into generative modeling are already established research directions. Therefore, the main contribution of this paper does not appear to be the use of a physics loss itself, but rather the proposed long-short flow-map decomposition and the associated near-terminal direct guidance mechanism. The paper motivates this design by arguing that existing methods rely on indirect distributional supervision and may suffer from a mismatch when nonlinear physical residuals are applied to intermediate or reconstructed states. This is a reasonable motivation. However, I would like to see a clearer explanation of what is fundamentally new about the proposed formulation compared with existing physics-informed flow-map, consistency-model, and shortcut-based approaches.

  2. The reported Darcy flow results are promising: PI-LSFM achieves substantially better Wasserstein distance and Jensen–Shannon divergence than the evaluated one-step baselines, while also reducing the physical residual. However, the evidence is less conclusive when considering the experiments as a whole. In the topology optimization benchmark, for example, PI-LSFM is competitive with existing methods but does not outperform them on every reported metric. Thus, the results appear to support a favorable trade-off in some settings rather than uniform superiority across all tasks. I would encourage the authors to make the scope of their claims more precise and discuss the conditions under which PI-LSFM offers the greatest benefit.

  3. The proposed framework combines several design choices, including the long-short decomposition, direct near-terminal data guidance, physics-potential guidance, and numerical discretization schemes. The paper includes ablations on some of these choices, but it remains important to establish which components are responsible for the observed gains. In particular, I would like to see a clearer separation between the contribution of the overall framework and the contribution of the specific numerical implementation. Comparisons under matched training budgets and consistent hyperparameter-selection procedures would also help strengthen the conclusions.

Questions for authors

What is the key theoretical or methodological distinction between PI-LSFM and existing physics-informed consistency or shortcut-based approaches, beyond the use of a long-short decomposition and near-terminal guidance?

How sensitive are the results to the choice of (\Delta t) and the relative weight of the physics objective? Does the proposed mechanism provide consistent gains over a meaningful range of settings?

2 comments
AuthorsOpenReview ↗

Response to Reviewer mSsb (1/2)

Thank you for the constructive review. In short, physics-informed consistency and shortcut methods keep an indirect data objective, whereas PI-LSFM trains on a direct target. With the same terminal physics loss and matched settings, its RE is 0.055, against 0.240 for consistency and 0.435 for shortcut.

W1, Q1: What is new relative to physics-informed flow-map, consistency and shortcut approaches?

What differs is the signal that supervises the generator. The direct target is what lets the physics term and the data-distribution term work together, and this is the core of our method.

On the data side, consistency, MeanFlow and shortcut objectives supervise a long-range map through conditional-velocity regression or through self-consistency across times, because a long-range map has no consistent data-conditioned counterpart (Theorem F.1). PI-LSFM needs no time variable, conditional path or Jacobian-vector product: its target x+T(x0)x+\mathcal T(x_0) is a closed-form function of the data batch and the generated batch.

On the physics side, variable-endpoint methods in their original form apply the residual to an intermediate or reconstructed state, which leaves a Jensen gap. In PI-LSFM the Δt→0\Delta t\to0 limit places the residual on the generator output at t=1t=1, where the gap vanishes. Endpoint-fixed consistency models can also use a terminal physics loss, but their data objective remains indirect.

Table 4 isolates the data-side difference. All rows are one-step and share the terminal physics loss, generator architecture, data, batch size and number of generator-training steps:

Darcy, NFE = 1RE ↓WD·10² ↓JS·10¹ ↓
PI-Consistency (sCM)0.2403.8895.023
PI-MeanFlow4603.5734.2435.457
PI-Shortcut0.4352.3362.518
PI-LSFM0.0550.1080.171

Adding conflict-free updates to the three baselines does not close the gap (best RE 0.152, best WD·10² 2.184; Table 16), so resolving gradient conflicts between the two terms is not enough, and the gain is due to the direct target. The feature encoder is not what puts PI-LSFM ahead of the baselines. The encoder stabilizes training: without it the run eventually collapses (Fig. 11), yet before the collapse the raw-space model is already ahead of every baseline (RE 0.044, WD·10² 0.446; Sec. 3.3).

The 2D toy problem (App. E.1) shows the same thing in a setting that can be visualized. At any of the six physics weights tested, a physics-informed consistency model fails to balance the two terms: a small weight leaves the constraint violated and a large one collapses the modes (Fig. 9). Its residual stays more than 20× above ours at every batch size (Table 11), whereas PI-LSFM satisfies the constraint and keeps all three modes.

W3: Which components are responsible; framework vs. numerical implementation; matched budgets.

The framework defines the direct-target formulation, and the numerical rules determine how its data and physics components are computed. We compared these rules on Darcy, taking each rule at its best λˉ\bar\lambda or schedule (Fig. 4c, 4d; App. D):

Data term (physics = FE), 100K stepsRE ↓WD·10² ↓
FE8.1254.244
BE1.8534.857
TR0.1170.213
Physics term (data = TR), 400K stepsRE ↓WD·10² ↓
FE0.0550.108
BE0.4360.122
BE-Lim2.5400.184

The two bold rows are the same configuration at 100K and 400K steps. The data term needs the second-order rule, which combines attraction to the data with repulsion from the generated samples, while the physics term does best with the explicit rule, i.e. the residual gradient. In raw space this member (TR + FE) is equivalent to a drifting-style objective plus a terminal physics loss, in the sense that the two have the same training gradient; we have added this as Proposition C.2 (App. C.4). The implicit and projection rules that the framework also yields do not improve on it.

Regarding budgets and selection (App. G.1, G.4), the one-step baselines of Table 4 keep their original objectives under the matched settings listed above, including the same residual. The MAE pretraining is accounted for in Table 5: with it amortized, PI-LSFM costs 0.308 s per step, within the 0.223–0.328 s range of the three baselines. Our schemes are selected once on Darcy and kept for all four benchmarks, the bandwidth ϵ=0.1\epsilon=0.1 is shared, and the physics weights of the other three benchmarks are taken from prior work without a per-task search. On the toy problem the sCM baseline is swept over six physics weights from 0 to 100 (Table 11). Compared with the multi-step PBFM, PI-LSFM uses the same batch size, a matched parameter count, and 400K training steps instead of 1,248K.

AuthorsOpenReview ↗

Response to Reviewer mSsb (2/2)

Q2: Sensitivity to Δt\Delta t and to the physics weight.

Δt\Delta t is not a hyperparameter of the method: the loss is the Δt→0\Delta t\to0 limit (Eq. 6). As for the physics weight, with the selected schemes RE falls monotonically as λˉ\bar\lambda goes from 0.1 to 3.0, a 30× range, while WD and JS remain relatively stable (Fig. 6a, 400K steps). A separate 100K-step sweep shows that training degrades beyond this range, at λˉ=10\bar\lambda=10, where all three metrics rise (Fig. 7c). Three seeds give RE 0.125 ± 0.008, WD·10² 0.264 ± 0.047 and JS·10¹ 0.237 ± 0.025 (mean ± SD, n = 3, 100K steps; Table 17).

W2: Scope of the claims.

PI-LSFM reaches comparable quality much faster, and in many cases better quality. With 1 NFE instead of 20 or 100, inference is 2.8× to 130× faster, and on the two flow benchmarks PI-LSFM is first on 7 of the 10 metrics, including every residual and distributional metric, while the 20-step baselines keep a lower MMSE on both and a lower SMSE on dynamic stall. This is the accuracy–efficiency trade-off that Sec. 4 states as our claim. On topology optimization, PI-LSFM has the in-distribution compliance error closest to the reference (−0.00) and a lower volume-fraction error than PIDM in both settings (1.53 vs. 2.25 and 1.66 vs. 1.91), at 0.013 s per design against 1.701 s. PIDM keeps a lower residual and a lower out-of-distribution compliance error.

The benefit is largest where strong physics enforcement meets one-step training, which is the regime where indirect objectives become fragile. On Darcy, RE is 4× to 8× lower than for PI-Consistency and PI-Shortcut, and WD is more than 20× lower. On dynamic stall the residual is nearly three orders of magnitude below the best 20-step baseline (Table 3). The revised Secs. 4.2 and 4.3 report the results benchmark by benchmark.

Reviewer_TUPdOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces Physics-Informed Long-Short Flow Maps (PI-LSFM), which is a one-step generative model that tries to produce both data-aligned and physics-constrained samples. The method aims to alleviate the tension between satisfying data fidelity and physical constraints by addressing them directly during training. The authors evaluate on partial differential equation tasks, evaluating distribution fidelity, physical observables, and inference efficiency.

Critical strengths

S1. In general, I like the idea of addressing the conflict between enforcing physical residual and data-aligned loss in a one-step generative model. The authors do provide substantial mathematical derivations to arrive at their framework, mainly to derive short map targets, which are based on data-distribution physics residual corrections.

S2. The empirical results are dense, demonstrating promising physical consistency and distribution performance while maintaining the inference efficiency of the framework.

Critical weaknesses

W1. The authors do discuss the connection between PI-LSFM and Drifting Models [1], which I appreciate. However, I find the practical contribution of the individual components of the framework not clearly distinguished from existing methods.

The authors argue that the distinction comes from jointly derived physics guidance, alternative numerical discretizations, and feature-space training. While these components are interesting and motivated theoretically and empirically, it remains unclear how much of the improvement is specifically attributable to the proposed long-short formulation rather than combining a drifting-style distribution-matching objective with a physics correction.

Moreover, recent work [2] also provides a long-short flow-map interpretation of Drifting Models. A direct experimental comparison or ablation isolating these differences would materially strengthen the paper.

W2. The empirical results are strong but quite mixed. The method does not seem uniformly competitive in presented benchmarks. For example, on Kolmogorov flow, the MMSE is substantially worse than the multi-step baselines. Similarly, on dynamic stall, both MMSE and SMSE are worse than those achieved by PBFM and several other methods. For topology optimization, the method similarly has mixed results, where I can see some degradation in out-of-distribution performance.

W3. Moreover, I strongly feel the presentation of the paper could be improved. For example, there are frequently overflowing tables and figures in the margins. The presentation seems pretty dense to me, particularly the main derivation and the numerical target tables. I feel that the core algorithm could be communicated more directly, and particular emphasis on broader limitations related to the above weaknesses would benefit from explicit discussion of the practical and theoretical assumptions of the framework.

References
1: Deng, Mingyang, et al. "Generative modeling via drifting." arXiv preprint arXiv:2602.04770 (2026).
2: Li, Zhiqi, and Bo Zhu. "A long-short flow-map perspective for drifting models." arXiv preprint arXiv:2602.20463 (2026).

Questions for authors

See the weaknesses above, but here are some additional questions:

Q1. Given the relationship to Drifting Models, can the authors more clearly isolate the practical contribution of the long-short formulation and the additional physics correction? Would a drifting-style objective with a terminal physics loss achieve comparable results?

Q2. The authors do mention PIDDM as a potential one-step generative model. Why is it not included in the quantitative comparison?

Q3. The authors discuss the accuracy–efficiency trade-off, which is reasonable, but it remains unclear whether these differences arise from the one-step formulation, the physics correction, or the choice of training objective. Could authors provide some detailed discussion or some kind of ablation to help me understand these trade-offs better? I think this would materially strengthen the paper and the conclusions.

I would be happy to increase my score once these concerns are addressed adequately.

2 comments
AuthorsOpenReview ↗

Response to Reviewer TUPd (1/2)

Thank you for the careful review. In short, the configuration we selected is equivalent in raw space to a drifting-style objective with a terminal physics loss, and the gain over existing one-step methods is due to the direct target (RE 0.055 against 0.240–0.435 with the terminal physics loss held fixed).

W1, Q1: Contribution of the long-short formulation; would a drifting-style objective with a terminal physics loss be comparable?

Yes, it would, because in raw space that objective is the configuration we selected. With the trapezoidal rule (TR) for the data term and Forward Euler (FE) for the physics term, the target of Algorithm 1 is

T=12(mp1(x)−x)+12(x−mqθ(x))−λˉλJR⊤R(x),x=ψ0→1θ(x0),\mathcal T=\tfrac12\big(m_{p_1}(x)-x\big)+\tfrac12\big(x-m_{q_\theta}(x)\big)-\bar\lambda\lambda J_{\mathcal R}^\top\mathcal R(x),\quad x=\psi^\theta_{0\to1}(x_0),

where mp1m_{p_1} and mqθm_{q_\theta} are the kernel-weighted means of the data and of the generated samples. The first two terms are the attraction and repulsion of a drifting-style update. The target sits under a stop-gradient, so the third term contributes exactly the gradient of λˉλ∥R(x)∥2\bar\lambda\lambda\|\mathcal R(x)\|^2. In raw space (dynamic stall, topology optimization) the two objectives are therefore equivalent: they have the same training gradient. Darcy and Kolmogorov are trained in feature space, where the data term is the same field and the physics step enters through ϕ(x+λˉTphys)\phi(x+\bar\lambda\mathcal T_{\rm phys}) (Eq. 13). We have added this result as Proposition C.2 (App. C.4) and refer to it from Sec. 3.3.

Beyond this equivalence, the long-short formulation contributes four things.

First, it explains why both terms act on the generator output. Taking Δt→0\Delta t\to0 in the short map moves the residual to t=1t=1, where the Jensen gap vanishes. It also sharpens the kernel enough for the data target to be estimated from a batch, and it leaves a loss that involves only the single map ψ0→1\psi_{0\to1} (Sec. 3.3). This is how the two supervisions become one direct target.

Second, it yields a family of numerical rules and the evidence for choosing among them. The same short flow also gives single-sided data rules and implicit physics rules, which the standard drifting-plus-terminal-loss formulation does not specify. We tested them on Darcy, each at its best λˉ\bar\lambda or schedule (Fig. 4c, 4d; App. D):

Data term (physics = FE), 100K stepsRE ↓WD·10² ↓
FE (attraction only)8.1254.244
BE (repulsion only)1.8534.857
TR (both)0.1170.213
Physics term (data = TR), 400K stepsRE ↓WD·10² ↓
FE (residual gradient)0.0550.108
BE (Gauss–Newton)0.4360.122
BE-Lim (projection)2.5400.184

The two bold rows are the same configuration (TR + FE) at 100K and 400K steps. The member equivalent to drifting plus a terminal physics loss is thus the best one, with an RE nearly 8× lower than the next physics rule, and the framework accounts for it as the second-order data rule combined with the explicit physics rule.

Third, it extends the long-short interpretation of Drifting Models in [2] from the data term to physics. We write the data term as a potential, which removes the singularity of the velocity field at t=1t=1 and lets a physics potential enter the same short flow as a second force. Both potentials can then be discretized within the same numerical framework.

Fourth, and most importantly, it reveals the role of the direct target in physics-informed one-step generation, a setting in which the indirect objectives fail to balance the two terms. On the 2D toy problem a physics-informed consistency model cannot balance them at any of the six physics weights tested, and on Darcy the three indirect baselines stay far behind with the same terminal physics loss (see Q3).

W3: Presentation.

Thank you for pointing these out. In the revised paper, Tables 1 and 2 no longer collide with the surrounding text, and Tables 1, 2 and 8–10, the subtables of Fig. 4, the panels of Fig. 6 and the appendix equations that ran past the text width have been rescaled to fit. Sec. 3.3 now writes the whole target as a single vector field next to Algorithm 1 (the equation in our answer to W1), so the algorithm can be read without going through the derivations behind Tables 1 and 2. Finally, Secs. 4.2 and 4.3 name the metrics on which multi-step baselines are better, App. G.4 says that the schemes are selected on Darcy, and the conclusion discusses training-time-only constraint enforcement and the limited set of numerical schemes as limitations.

AuthorsOpenReview ↗

Response to Reviewer TUPd (2/2)

Q3: Do the differences come from the one-step formulation, the physics correction, or the training objective?

They come from the training objective. More precisely, the direct target is what lets the physics term and the data-distribution term work together, and this is the core of the method. Both terms act on the same quantity, the generator output, and the data term is a closed-form target and not a regression surrogate, so over a wide range of physics weights a strong physics term does not disturb distribution matching. With an indirect data objective the two are hard to coordinate (Sec. 2), and this is where one-step physics-informed training becomes fragile.

In Table 4 we train PI-Consistency (sCM), PI-MeanFlow and PI-Shortcut with the same terminal physics loss, architecture, data, batch size and training budget as PI-LSFM, all at NFE = 1:

Darcy, NFE = 1RE ↓WD·10² ↓JS·10¹ ↓
PI-Consistency0.2403.8895.023
PI-MeanFlow4603.5734.2435.457
PI-Shortcut0.4352.3362.518
PI-LSFM without the feature encoder (best checkpoint before collapse, Sec. 3.3)0.0440.4460.889
PI-LSFM0.0550.1080.171

Adding conflict-free updates to the baselines does not close the gap (best RE 0.152, best WD·10² 2.184; Table 16), so resolving gradient conflicts between the two terms is not enough, and the gain is due to the direct target. The feature encoder is not what puts PI-LSFM ahead of the baselines. The encoder stabilizes training: without it the run eventually collapses (Fig. 11), yet before the collapse the raw-space model is already ahead of every baseline, with a WD five times lower than the best of them. In raw space the physics term has the same gradient as the terminal loss used by the baselines (Proposition C.2), so in that row the data objective is the only difference.

Two further observations show the two terms cooperating under the direct target. As λˉ\bar\lambda increases from 0.1 to 3.0, RE falls monotonically while WD and JS remain relatively stable (Fig. 6a, 400K steps), so the physics weight can be raised 30× without paying for it in the distribution. A separate 100K-step sweep shows that training degrades beyond this range, at λˉ=10\bar\lambda=10 (Fig. 7c). The consistency baseline on the 2D toy problem (App. E.1) shows the opposite: a small weight leaves the constraint violated and a large one collapses the modes (Fig. 9), and its residual stays more than 20× above ours at every batch size (Table 11), whereas PI-LSFM satisfies the constraint and keeps all three modes.

As for the other two factors, the physics correction sets the residual level through λˉ\bar\lambda, and the one-step formulation gives the efficiency, 0.012 s per sample against 0.101 s for PBFM at 20 NFE.

W2: Results across benchmarks.

PI-LSFM reaches comparable quality much faster, and in many cases better quality. With 1 NFE instead of 20 or 100, inference is 2.8× to 130× faster. On the two flow benchmarks PI-LSFM is first on 7 of the 10 metrics, including every residual and distributional metric, and on topology optimization it is competitive with the 100-step baselines. This is the accuracy–efficiency trade-off that Sec. 4 states as our claim.

Table 3 (vs. best 20-step baseline)PI-LSFM first onBaselines better on
KolmogorovRE 0.404 vs. 1.362; WD 0.140 vs. 1.222; JS 1.621 vs. 7.440; SMSE 1.765 vs. 2.484MMSE 25.83 vs. 4.188
Dynamic stallRE 0.00015 vs. 0.143; WD 1.633 vs. 1.814; JS 0.403 vs. 0.680MMSE 5.218 vs. 1.490; SMSE 4.894 vs. 0.874

The 20-step baselines keep an advantage on some conditional statistics, which is consistent with the benefit of iterative refinement noted in Sec. 4.

Table 7PI-LSFMPIDM (100 NFE)Best baseline
In-distribution: RMAE\mathcal R_{\rm MAE} / CE % / VFE %1.66e−3 / −0.00 / 1.531.24e−3 / 0.06 / 2.251.24e−3 / 0.06 / 1.49
Out-of-distribution: RMAE\mathcal R_{\rm MAE} / CE % / VFE %1.95e−3 / 1.86 / 1.661.29e−3 / 0.56 / 1.911.29e−3 / 0.30 / 1.59
Time per design0.013 s (130× faster)1.701 s

In distribution, the compliance error of PI-LSFM is the closest to the reference among all methods. Out of distribution, its volume-fraction error remains below PIDM's, while its residual and compliance error are behind the 100-step baselines.

Q2: PIDDM.

We left PIDDM out of the quantitative comparison because it distills a pretrained multi-step diffusion teacher, which would add a pretraining stage and a different training budget to the matched comparison of Table 4. That table is therefore restricted to methods trained from scratch, and PIDDM is discussed in App. A.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General response and summary of the revision (1/2)

We thank all three reviewers for their time and suggestions. The reviews describe the extension to physics-constrained generation as "practically valuable, with sufficiently extensive experiments" (nLsw), the framework as "worth investigating" (mSsb), and the results as "dense, demonstrating promising physical consistency and distribution performance" (TUPd). Four questions come up in more than one review, so we answer them here and give the details in the individual threads. Changes in the PDF are marked in red.

Novelty: a direct target for physics-informed one-step generation (TUPd, mSsb). Existing physics-informed one-step methods supervise the generator indirectly. Their data objective is a regression surrogate or a self-consistency condition across times, and a strong physics loss is hard to coordinate with it (Sec. 2). PI-LSFM replaces this with a direct target. In the limit Δt→0\Delta t\to0 of the long-short decomposition, the data target becomes a closed-form function of the batch and the residual sits on the generator output at t=1t=1, where the Jensen gap vanishes. The direct target is what lets the physics term and the data-distribution term work together, and this is the core of our method. We analyze it in three ways.

First, Table 4 and App. G.1 compare direct and indirect objectives with generator architecture, data, batch size, number of generator-training steps and terminal physics loss all matched:

Darcy, NFE = 1RE ↓WD·10² ↓JS·10¹ ↓
PI-Consistency (sCM)0.2403.8895.023
PI-MeanFlow4603.5734.2435.457
PI-Shortcut0.4352.3362.518
best per metric across the three, with conflict-free updates (Table 16)0.1522.1842.363
PI-LSFM (direct target)0.0550.1080.171

Adding conflict-free updates does not close the gap, so resolving gradient conflicts between the two terms is not enough, and the gain is due to the direct target (the TUPd thread gives the encoder-free control).

Second, under the direct target the physics weight can be raised 30× (from 0.1 to 3.0) with RE falling monotonically while WD and JS remain relatively stable (Fig. 6a).

Third, the 2D toy problem shows the failure of the indirect objective directly. At any of the six physics weights tested, a physics-informed consistency model fails to balance the two terms: a small weight leaves the constraint violated and a large one collapses the modes (Fig. 9), and its residual stays more than 20× above ours (Table 11).

Relation to Drifting Models (TUPd, nLsw). The direct target also makes the relation to Drifting Models exact. The configuration we selected uses the trapezoidal rule (TR) for the data term and Forward Euler (FE) for the physics term. In raw space it is equivalent to a drifting-style objective plus a terminal physics loss, in the sense that the two have the same training gradient (Proposition C.2, App. C.4). The target of Algorithm 1 is

T(x0)=12(mp1(x)−x)⏟attraction (FE half)+12(x−mqθ(x))⏟repulsion (BE half)−λˉλJR⊤R(x),x=ψ0→1θ(x0),\mathcal T(x_0)=\underbrace{\tfrac12\big(m_{p_1}(x)-x\big)}_{\text{attraction (FE half)}}+\underbrace{\tfrac12\big(x-m_{q_\theta}(x)\big)}_{\text{repulsion (BE half)}}-\bar\lambda\lambda J_{\mathcal R}^\top\mathcal R(x),\qquad x=\psi^\theta_{0\to1}(x_0),

where mp1m_{p_1}, mqθm_{q_\theta} are the kernel-weighted means of the data and generated samples. Because the target sits under a stop-gradient, the gradient of ∥x−sg(x+T)∥2\|x-\mathrm{sg}(x+\mathcal T)\|^2 is that of the drifting-style loss plus that of λˉλ∥R(x)∥2\bar\lambda\lambda\|\mathcal R(x)\|^2. The long-short view of the data term was given for drifting models in the concurrent work of Li & Zhu (2026); this paper builds physics-informed one-step generation on it.

AuthorsOpenReview ↗

General response and summary of the revision (2/2)

Which numerical scheme is best (TUPd, mSsb, nLsw). The framework derives a family of update rules. We tested them on Darcy, each at its best λˉ\bar\lambda or schedule (Fig. 4c, 4d):

Data term (physics = FE), 100K stepsRE ↓WD·10² ↓
FE (attraction only)8.1254.244
BE (repulsion only)1.8534.857
TR (both)0.1170.213
Physics term (data = TR), 400K stepsRE ↓WD·10² ↓
FE (residual gradient)0.0550.108
BE (Gauss–Newton)0.4360.122
BE-Lim (projection)2.5400.184

The two bold rows are the same pair of rules, TR + FE, at 100K and 400K steps. The member that is equivalent to drifting plus a terminal physics loss is the best of the family, and the framework accounts for it as the second-order data rule combined with the explicit physics rule. The implicit and projection rules it also yields do not improve on it.

Results across benchmarks (TUPd, mSsb). PI-LSFM reaches comparable quality much faster, and in many cases better quality. With 1 NFE instead of 20 or 100, inference is 2.8× to 130× faster. This is the accuracy–efficiency trade-off that Sec. 4 states as our claim. On Kolmogorov flow and dynamic stall, PI-LSFM is first on 7 of the 10 metrics, including every residual and distributional metric; on dynamic stall its residual is nearly three orders of magnitude below the best baseline. The 20-step baselines keep a lower MMSE on both, and a lower SMSE on dynamic stall. On topology optimization, PI-LSFM is competitive with the 100-step baselines while generating a design 130× faster than PIDM, and has the in-distribution compliance error closest to the reference, but its out-of-distribution residual and compliance error are behind theirs.

Revision. Proposition C.2 is new; Sec. 3.3 now writes the target in this vector-field form next to Algorithm 1, and Related Work refers to it. Secs. 4.2 and 4.3 report the results benchmark by benchmark, including the metrics on which the multi-step baselines are better. App. D.3 and the caption of Fig. 11 state that the encoder-free metrics of Sec. 3.3 are from before the run collapses. Tables, figure panels and equations that ran past the text width now fit.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.