acceptodds
Under review as a conference paper at ICLR 2027

MintFlow: Minimal Trajectory Intervention for Constrained Flow Matching

Abstract

Flow matching models excel at generative modeling, and many downstream applications require their samples to satisfy prescribed constraints, such as observed measurements and physical laws. However, existing constrained samplers often face a trade-off: enforcing constraints can substantially displace samples from the pretrained data distribution. To address this trade-off, we introduce MintFlow, a training-free constrained sampling framework that formulates constraint enforcement as a minimal intervention on the pretrained flow trajectory. MintFlow seeks the minimal perturbation of an intermediate flow state such that its subsequent evolution under the pretrained flow field satisfies the target constraint. By minimally perturbing the flow state while keeping the pretrained flow field unchanged, MintFlow enforces the constraint while minimizing unnecessary deviation from the pretrained distribution. An adjoint formulation yields a closed-form expression for this perturbation, eliminating expensive iterative optimization. Furthermore, MintFlow adaptively selects the intervention time to balance the required perturbation magnitude with its amplification by the remaining flow. Across a range of tasks in generative vision and physical system modeling, MintFlow achieves competitive constraint satisfaction while preserving the pretrained generative distribution substantially better than state-of-the-art constrained methods.

Primary area:
generative models
Keywords:
Flow Matching, Constrained Generative Modeling, Distributional Fidelity

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_R9HLOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes MintFlow, a training-free method for imposing constraints on samples from a pretrained flow-matching model. It perturbs an intermediate state, then continues generation using the original velocity field. For a chosen intervention time, it linearizes the terminal constraint and computes the minimum-norm correction satisfying the linearized constraint via the adjoint method. MintFlow then selects an intervention time by balancing correction magnitude against remaining integration time, and optionally repeats the correction. On downstream tasks including image reconstruction, image editing, and physical systems, MintFlow consistently improves performance while maintaining low sampling cost.

Critical strengths
  • The paper provides a clear and useful formulation for general constrained generative modeling. The method directly addresses the contention between satisfying constraints and retaining plausible samples. Reusing a frozen model makes the approach practically attractive.
  • The proposed approach enjoys a coherent theoretical justification. Under some assumptions, the residual bound and Wasserstein stability bound appear sound. These explain why small, sensitivity-aware corrections can be useful.
  • Experiments covered in this paper span multiple domains with consistent reported gains. The evaluation covers image restoration, editing, learned image constraints, and two PDE systems. Cost comparison also makes this work competitive compared to existing gradient-based approaches.
Critical weaknesses

One practical assumption that the model adopts is that the base model is "rectified" enough for direct one-step extrapolation. However, this is not always available:

  • Even the most rectified model still requires tens of sampling steps for decent image generation quality. One-step generation remains harder unless the model is distilled.
  • Not all generative base models have been rectified. Additionally, diffusion models may adopt non-OT paths. The paper must consider accommodating these base models.

This will impact both the theoretical guarantees and practical performance.

Questions for authors
  • Is it possible to adopt the same approach to different flow/diffusion probability paths? Specifically, can you accommodate diffusion paths like those used in EDM that achieve strong image-generation performance?
  • Should one-step extrapolation still be used for these probability paths, how much will it affect the theoretical results?
  • The proposed approach applies to differentiable constraints pretty well. Can it be extended to non-differentiable rewards?
4 comments
AuthorsOpenReview ↗

Response to Reviewer R9HL

W: One practical assumption that the base model is "rectified".

We thank the reviewer for raising this concern. We would like to clarify that MintFlow does not require the base generative model to exhibit rectified or approximately linear trajectories. Both our methodological formulation and theoretical analysis are developed for general differentiable flow models, including nonlinear flows, and compute intervention sensitivities through the adjoint-based formulation without assuming linear transport.

A key advantage of MintFlow is its applicability across diverse flow models. This broader applicability is also demonstrated in our experiments (Table 4): we apply MintFlow to conventional flow matching models for image generation (Appendix D) and functional flow matching models for physics-informed generation (Section 4.3), neither of which relies on the one-step approximation.

When the model is rectified flow (RF) which were used in image inverse problems (Section 4.1) and image editing (Section 4.2), its linear structure provides an additional computational advantage, rather than restricting the applicability of MintFlow to rectified flows. The one-step endpoint approximation discussed in Appendix A.4 is a computational simplification specific to RF models, exploiting their approximately linear transport property. It is not an assumption required by the general MintFlow framework, and thus does not restrict its applicability to other flow models.

In the revised manuscript, we have elaborated on the RF-specific computational simplification in Appendix A.4 and Algorithm 3. We have also clarified which experiments use RF and which use other flow models at the beginning of Appendix B.

Q1: Is it possible to adopt the same approach to different flow/diffusion probability paths?

MintFlow is formulated for general ODE-based generative models and does not rely on a specific probability path or rectified transport. We have demonstrated its applicability across diverse flow models, including conventional flow matching, functional flow matching, rectified flow, and latent text-guided rectified flow. In particular, our experiments with conventional and functional flow matching demonstrate that MintFlow also applies effectively to nonlinear flow trajectories.

In principle, the same framework can be extended to diffusion models through their probability flow ODEs, including those associated with EDM, by computing the corresponding flow sensitivities using our adjoint-based formulation. While a systematic evaluation across additional diffusion backbones and probability paths is beyond the scope of this work, our extensive experiments across diverse models, datasets, and tasks (see Table 4) already demonstrate the broad applicability of MintFlow. We consider further extending its evaluation to a wider range of diffusion-based generative models an interesting direction for future work.

Q2: Should one-step extrapolation still be used for these probability paths, how much will it affect the theoretical results?

We would like to clarify that the methodology and theoretical analysis in the main manuscript are developed for general ODE-based flow models and do not assume one-step extrapolation. In particular, all theoretical guarantees are derived based on the general flow dynamics, without imposing any one-step approximation.

Q3: Can it be extended to non-differentiable rewards?

We thank the reviewer for this question. Our current theoretical formulation assumes that the constraint function HH is differentiable, as the minimum-norm correction is derived through a first-order Taylor expansion of the terminal constraint. However, the framework can also accommodate piecewise-differentiable constraints when the terminal state lies in a differentiable region. At points of nondifferentiability, using subgradients may provide a practical extension, but the validity of the resulting linearization and theoretical guarantees requires further investigation. More generally, techniques from constrained optimal control that accommodate nonsmooth constraints may provide another promising direction for extending our framework. A systematic extension to nonsmooth constraints is already identified as an important direction for future research in our Conclusion.

In this work, we focus on differentiable constraints, a setting that already encompasses a broad range of important applications, including image generation, image inverse problems, image editing, and physics-informed generation. We demonstrate the applicability of MintFlow across all these settings in our experiments.

I thank the authors for their response. One of my concerns remains.

For rectified paths, the authors' interpretation is incorrect. The OT path in the forward noising process never guarantees a straight path during sampling due to marginalization. Additionally, one-step extrapolation is not limited to flow models, as a curved diffusion path also enjoys an exact posterior. Therefore, my question is regarding the validity of one-step extrapolation, as there is no guarantee that such extrapolation will lead to decent predictions, especially in the early generation stages. I'd suggest the authors further justify the practical and theoretical validity in this sense.

Since the authors proposed an erroneous theoretical argument in their rebuttal that may undermine the overall correctness of the proposed method, I revised my score to 2 unless the authors justify the theoretical gap.

AuthorsOpenReview ↗

Author Response to Reviewer R9HL

We appreciate your prompt response and would like to address the remaining concerns point by point.

The OT path in the forward process never guarantees a straight path during sampling.

Dear Reviewer, thank you for your feedback. It appears there may be a slight misunderstanding regarding our proposed method. Our method does not assume that the RF sampling trajectory is exactly straight. Rather, the straight-path assumption is introduced only to simplify the computation of the minimal correction. Importantly, we do not use a single extrapolation to generate the final sample. Instead, the sampling process iteratively applies the estimated correction at each Euler step kk:

  1. Compute the minimal correction δk\delta_k under the single-step extrapolation approximation.
  2. Apply the correction to obtain ukcorr=uk+δku_k^{\mathrm{corr}} = u_k + \delta_k.
  3. Advance the corrected state by one Euler step with step size hh: uk+1=ukcorr+hfθ(ukcorr,tk)u_{k+1} = u_k^{\mathrm{corr}} + h f_\theta(u_k^{\mathrm{corr}}, t_k).
  4. Repeat the procedure using uk+1u_{k+1} as the reference state for computing the next correction.

Thus, the single-step extrapolation is used only to approximate the correction δk\delta_k at each step, not to approximate the entire sampling trajectory as a straight path.

One-step extrapolation is not limited to flow models:

We do not claim that one-step extrapolation is a mathematical property exclusive to RF. MintFlow is applicable to general differentiable flow models, and its applicability is not restricted to RF or straight trajectories. In the RF setting, we utilize the one-step extrapolation solely to efficiently approximate the minimum-norm correction δk\delta_k (see step 1 above). The same framework can also be applied to diffusion paths.

Validity of one-step extrapolation:

We agree that one-step extrapolation does not necessarily yield an exact correction for the terminal constraint. However, our algorithm does not use this approximation to directly generate the final sample. Instead, it uses the extrapolated endpoint only to estimate the correction δk\delta_k (Step 2), applies this correction to the current state, and advances one Euler step before recomputing the correction from the updated state. Thus, the approximation is refined throughout the sampling process rather than being used only once.

Crucially, this iterative procedure can compensate for the error introduced by the correction δk\delta_k obtained via one-step extrapolation. Since δk+1\delta_{k+1} is computed from the updated state uk+1u_{k+1}, it can account for the error left by δk\delta_k. Moreover, the extrapolation becomes increasingly accurate as the remaining sampling time decreases, and coincides exactly with the final discrete Euler transition at the last step. We further characterize the propagation of approximation errors through this iterative procedure in Theorem A.4 of the revised manuscript. Under appropriate regularity and local convergence assumptions, we establish a bound on the final constraint residual, showing that finer Euler discretization (smaller hh) reduces terminal constraint violation. This theoretical result is also supported by our experiments in Appendix H.3, where increasing the number of RF sampling steps improves image editing performance.

We have clarified these distinctions and strengthen the discussion of the theoretical and practical validity of one-step extrapolation in the revised manuscript.

Theoretical gap:

We thank the reviewer for raising this concern. However, we are not entirely sure which specific theoretical argument in our previous rebuttal is considered erroneous. Based on the preceding comments, we understood the main concern to be the validity of one-step extrapolation.

To address this concern, we have further developed our theoretical analysis and introduced a new convergence result in Theorem A.4. This result explicitly characterizes the propagation of approximation errors through successive corrections and establishes a bound on the final constraint residual, without assuming straight marginal sampling trajectories.

We acknowledge that a more comprehensive theoretical justification of our RF-specific formulation remains an important direction for future work. Nevertheless, our theoretical analysis, together with the experimental results, provides evidence that using one-step extrapolation to estimate perturbations can lead to an effective iterative correction procedure in practice, despite the approximation involved.

We hope this additional analysis addresses the reviewer's concern. If a different theoretical gap was intended, we would greatly appreciate further clarification so that we can address it directly.

Reviewer_R1emOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

MintFlow adds some constraints and works by correcting intermediate states of a pretrained flow while keeping its velocity field unchanged. The method uses adjoint sensitivies to compute a minimum norm correction to the linearized termnal constrain and chooses when to intervene and then resumes generation. Experiments cover image inverse problems, editing and physical systems.

Critical strengths
  • I like the formulation. For instance, it accounts for how the remaining flow transforms an intermediate correction, which gives a clear motivation for intervening along the trajectory. The connection to a minimum norm Newton step is straightforward, and the local error and stability bounds help explain the method.

  • The experimental coverage is a strength. The evaluation includes different pretrained models, measurement constraints, nonlinear learned constraints, and physical systems. Using the same frozen models and source noises across methods also makes the comparisons easier to interpret.

  • The main results are encouraging. Table 1 shows improvements in both measurement consistency and reconstruction quality. I also find the physics results useful: ECI and PCFM satisfy the explicit constraints more tightly, but MintFlow achieves a better balance with distributional agreement and dynamical consistency.

Critical weaknesses
  1. I think the distribution preservation claim needs more careful wording. The Wasserstein bound controls deviation from the original generator; it does not establish sampling from the correct conditional distribution or show that the selected intervention minimizes distributional distortion. The convergence result also covers a scalar constraint at a fixed intervention time, so its scope is narrower than the full adaptive procedure.

  2. The contribution of adaptive timing is not fully isolated. The ablations vary the integration resolution, time-selection weight, and number of interventions, but do not directly compare the same correction against a well-chosen fixed intervention time at comparable cost. This leaves some uncertainty about how much of the improvement comes from selecting the time adaptively.

Questions for authors
  • Appendix E.2 uses 10-step RF sampling but refers to Eq. (15), which is derived for a single Euler step. Do the experiments differentiate through all remaining steps as in Eq. (19), or use the one-step endpoint approximation? Also, how is the pseudoinverse correction computed for the vector-valued measurement constraints?

  • Are FID and W2 in Appendix D computed using different features or normalizations? With the same Gaussian feature distributions, FID should equal squared W2, but Table 7 reports FID 0.221 and W2 0.0048 for MintFlow. Please clarify the definitions or check these entries.

  • What acceptance tolerance is used to collect the 2,000 constraint-satisfying reference samples in Appendix D, and how many unconditional samples are needed to obtain them?

  • Could the authors provide the exact MintFlow hyperparameters used for the image experiments, including candidate intervention times, KK, λ\lambda, refinement count/stopping tolerance, and the numerical procedure used to compute As†rA_s^\dagger r for high-dimensional vector constraints?

3 comments
AuthorsOpenReview ↗

Responses to Reviewer R1em's Weaknesses

W1: I think the distribution preservation claim needs more careful wording.

We thank the reviewer for this insightful comment. In response, we have carefully revised the theoretical claims and their presentation throughout the manuscript, including the Introduction, Section 3.2, the Conclusion, and Appendix. In particular, we now distinguish distributional stability from conditional distribution recovery and clarify that our minimum-norm correction minimizes the derived Wasserstein upper bound, rather than the actual distributional distortion. We have also clarified that our theoretical results on the convergence of adaptive iterative interventions are limited to the scalar-valued constraint setting. Finally, further extending our theoretical analysis remains an important direction for future work.

W2: The contribution of adaptive timing is not fully isolated.

We thank the reviewer for this suggestion. To investigate the effect of intervention timing, we partitioned the generation time interval into three distinct regions:

  • Early: [0.01t, 0.33t]

  • Middle: [0.34t, 0.66t]

  • Late: [0.67t, 0.99t]

To ensure a fair and computationally comparable evaluation, we uniformly sampled 10 candidate intervention times within each region. For each candidate time, we generated 1,000 samples by applying MintFlow's correction mechanism at the corresponding fixed timestamp. We report the mean and standard deviation across the 10 fixed intervention times for each region. The results for the heat equation are summarized in the table below.

SelectionFID ↓H-Pass (%) ↑Mean H (×10−6\times 10^{-6}) ↓Sim. rel. L2 ↓Total NFE (×105\times 10^{5}) ↓
Adaptive2.91E-596.713.5710.012119.610
Early4.49E-5 ± 0.0004976.98 ± 5.558.409 ± 1.6400.01277 ± 0.007111.01 ± 0.36
Middle2.82E-5 ± 0.0000695.74 ± 4.163.370 ± 0.7350.00980 ± 0.000110.12 ± 0.213
Late3.14E-5 ± 0.0003998.09 ± 0.8082.375 ± 0.6620.01075 ± 0.00209.563 ± 0.231

Here, H-pass denotes the percentage of generated samples whose constraint violation is at most 1E-5.

The results reveal a trade-off across evaluation metrics: middle intervention achieves the best FID and simulation accuracy, whereas late intervention yields the highest constraint satisfaction and lowest computational cost. In contrast, early intervention performs relatively poorly in both distributional quality and constraint satisfaction. Notably, our adaptive strategy offers a favorable balance, requiring fewer NFEs than fixed middle and early interventions while maintaining strong generation quality and constraint satisfaction.

These results demonstrate that no single fixed intervention time is optimal across all metrics. Although a carefully chosen fixed intervention time can outperform adaptive selection on individual metrics, identifying such a time a priori is challenging because the optimal choice may vary across samples, pretrained flow models, and constraint operators. These findings motivate our efficiency-aware adaptive criterion, which balances correction magnitude against remaining integration time without relying on a manually specified intervention time. We have added these experimental results and the corresponding discussion to Appendix H.4 of the revised manuscript.

AuthorsOpenReview ↗

Responses to Reviewer R1em's Questions

Q1-1: For multi-step RF sampling,do the experiments differentiate through all remaining steps as in Eq. (19)?

We thank the reviewer for pointing out this ambiguity. Following the corresponding baselines, our RF-based experiments use the full 10-step Euler discretization for RF sampling. For computational efficiency, we exploit the approximately linear transport property of RFs to approximate the remaining trajectory using a one-step endpoint estimate. This allows us to compute the intervention sensitivity directly from the local Jacobian, without solving the adjoint ODE. Importantly, the one-step approximation is used only to compute the minimum-norm correction. Following the baseline protocol, we apply this correction at each of the 10 sampling steps and then take an Euler step to advance the corrected state to the next sampling time. We have clarified this in the revised Appendix A.4, and summarized the procedure in Algorithm 3.

Q1-2: How is the pseudoinverse correction computed for the vector-valued measurement constraints?

For the constraint-by-state sensitivity matrix AsA_s computed via the adjoint equation, we apply As†A_s^\dagger using a compact SVD, discarding singular values below a relative threshold. If the SVD fails, we fall back to a least-squares solver. Importantly, the pseudoinverse As†A_s^\dagger is never explicitly formed; instead, we directly compute its product with the constraint residual. We provide these implementation details in Appendix G.

Q2: Clarify FID and W2.

We thank the reviewer for catching this and apologize for the confusion caused by errors. FID and W2 are computed from the same feature representations and Gaussian approximations, so W2=FIDW_2=\sqrt{\text{FID}}. The discrepancies came from typos when transcribing the W2 entries into Tables 6-9. The implementations and the FID values are correct. We have corrected all affected W2 values. We checked all other reported metrics and found no errors. Since W2=FIDW_2=\sqrt{\text{FID}} is monotone, the corrected W2 values have the same ranking as FID. Thus, our comparison and conclusions are unchanged. We thank the reviewer again for bringing this issue to our attention, which has helped us improve the accuracy of the manuscript.

Q3: What acceptance tolerance is used to collect the constraint-satisfying reference samples in Appendix D, and how many unconditional samples are needed to obtain them?

We use an acceptance tolerance of 1E-4 to collect the 2,000 constraint-satisfying reference samples. We did not record the total number of unconditional samples generated during this collection process. The number of samples required depends on the acceptance rate, which can be substantially lower when the constraint corresponds to a rare event in the original distribution. We have clarified the acceptance threshold in Appendix D.3.

Q4: Could the authors provide the exact MintFlow hyperparameters used for the image experiments?

We thank the reviewer for this question. The detailed hyperparameter settings for the image generation experiments are provided in Appendix D.2.

For image inverse problems and image editing, we use rectified flow (RF) models and adopt the simplified implementation described in Algorithm 3. Since RF sampling uses a small number of steps, we do not employ adaptive intervention time selection. Instead, following the baselines' protocol for a fair comparison, we apply iterative corrections at each of the 10 RF sampling steps. Thus, the RF-based experiments use 10 iterative correction steps.

For the numerical procedure used to compute the pseudoinverse for vector-valued constraints, we refer the reviewer to our response to Q1-2.

thanks to the authors for the detailed response and updating the manuscript. Most of my concerns were addressed and i will be keeping my score. Just something to consider: there seems to be an inconsistency Q4 the authors say they used 10 steps while the actual submission sec H.3 says 20.

Reviewer_cpw2OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces MintFlow, which is a framework for training-free constrained sampling in generative models, formulating constraint enforcement as a minimal intervention on the pretrained flow trajectory. One of the key ideas in MintFlow is an adjoint formulation to enforce the given constraints and to adaptively select the intervention time to balance the required perturbation with its effect through the remaining flow.

Critical strengths

S1. The idea of enforcing the constraints through minimal intervention is well motivated and clear. I like the idea of balancing constraint consistency and distributional fidelity, and its key mathematical ideas, such as using adjoint formulations and choosing the intervention time, seem sound.

S2. In particular, I think the empirical evaluation is fairly broad and comprehensive, spanning different generative tasks, constraint types, and domains. Some of the reported results, particularly on the inverse problems, seem promising.

Critical weaknesses

W1. I have some significant concerns regarding the interpretation and consistency of the reported metrics. For example, in Appendix D.3, the authors state that they use the Wasserstein-2 distance, which is computed using a Gaussian approximation of the feature distribution. The square of the (W2W_2) distance should be equal to the corresponding Fréchet Inception Distance (FID) under the same Gaussian feature-space evaluation protocol. However, several reported values do not satisfy this relationship, raising the concern that these metrics are not used uniformly throughout the paper.

For example, in Table 7, the paper reports FID = 0.221 and (W2=0.0048W_2 = 0.0048) for MintFlow, whereas (FID≈0.470\sqrt{\mathrm{FID}}\approx 0.470). Interestingly, the PCFM row in the same table (FID = 10.225, W2=3.1977W_2=3.1977) satisfies this relationship. Similar discrepancies also occur in Tables 6, 8, and 9. These particular inconsistencies dilute the contribution of the paper because it relies on a central claim of distributional fidelity.

W2. Another key concern that I have in the paper is that the presentation of the core algorithm and its practical implementation is unclear. I do not understand which exact algorithm they eventually use in the experiments. The paper introduces Algorithms 1–3, with Algorithm 3 providing iterative refinement, along with a separate rectified-flow specialization.

If I understand this correctly, the general method requires forward integration, a backward adjoint solve, selection of an intervention point, and subsequent forward integration. However, Appendix A.4 confuses my understanding by introducing a local Jacobian simplification (Eq. 15), especially for rectified flows, and it is unclear how this leads to the multi-step sampling implementation (Eq. 19). In particular, Appendix E.2 mentions 10-step rectified-flow sampling while referring to the simplified formulation.

Moreover, there is very little discussion of how the pseudoinverse of the constraint Jacobian composed with the downstream flow sensitivity (Eq. 6) is computed in practice. For example, for vector-valued constraints, which we typically see in PDEs with hundreds of constraints (199 for Heat and 425 for RD1D in Appendix G.2), I do not see sufficient discussion of how this operator is implemented efficiently. I feel this is an important omission, given that efficiency is presented as a central advantage of the proposed method.

W3. While some of the image-generation results seem promising to me, I have several concerns about the performance of the PDE-specific tasks. For example, the authors state that they follow the PCFM benchmark protocol for their evaluation (Section 4.3), but it is quite unclear what exact implementations they are using.

The manuscript, for example, states in Appendix D.2 (line 1250) that "ECI and PCFM enforce the constraint through a final projection step." A similar characterization appears in Section 4.3, where these are referred to as post-hoc projection methods. However, PCFM and ECI are not simply terminal-projection methods. Both modify the sampling trajectory through intermediate corrections, with terminal correction applied when necessary. Interestingly, Appendix C discusses these mechanisms more accurately. I find this inconsistent and it is unclear to me which versions of the baselines are actually being evaluated.

Another point that I find surprising is Appendix G.3 (lines 2562–2563), where the authors state that standard ECI uses one mixing step and PCFM uses 10 Newton updates per Heat flow step and one per RD1D flow step. Since the Heat constraints are linear, one exact Newton/Gauss–Newton step should be sufficient to project a terminal state onto the constraint set. I do not understand why 10 Newton updates are needed for each Heat flow step, particularly since the original PCFM implementation uses one Newton correction per sampling step.

Another big issue I see from the results is that the reported performance of these baselines is quite different from what was reported in the original papers. For example, the PCFM paper (Table 3) reports FPD values of 1.22 for PCFM and 1.34 for ECI, compared with 1.77 for Vanilla FFM on the Heat equation. In contrast, MintFlow (Table 3) reports Heat FID of (1.851×10−11.851\times10^{-1}) for PCFM compared with (1.255×10−31.255\times10^{-3}) for Vanilla FFM. Similarly, the simulation error increases from 0.0557 for Vanilla to 0.445 for PCFM.

I understand that these are different distributional metrics and that the evaluation protocols may not be identical. However, the substantial difference in relative performance is surprising to me. What exactly has changed in the evaluation protocol or baseline implementation that leads to such different results?

Similarly, I find the computational comparison quite puzzling for simple linear constraints. Linear constraints are the simplest ones because they admit a direct projection. For example, the MNIST average-intensity constraint directly admits a closed-form projection. Yet Table 6 reports an inference time of 18.37 s for PCFM compared with 5.07 s for unconstrained sampling. If the baseline is implemented using only a terminal projection, this overhead seems difficult to justify. If the full PCFM method is being used, then the description of the baseline as a post-hoc projection method is unclear.

W4. The computational efficiency claims, I don't think, are sufficiently supported yet. As I referred to in W2, the MintFlow algorithm is mainly composed of a forward solve, an adjoint calculation, a correction solve, and forward reintegration. For high-dimensional constraints, computing the minimum-norm correction will introduce additional costs beyond the adjoint calculation itself. While the authors report favorable inference times in some image experiments, the corresponding general adjoint implementation and its computational breakdown remain quite unclear. In particular, Tables 3 and 10 do not report inference times for the PDE experiments.

I also think the comparison with existing literature could be improved. While Appendix C discusses closely related methods such as HardFlow and TOCFlow, I do not see direct comparisons that help distinguish their practical contributions. Other relevant works such as PBFM [1] and MS-Flow [2] could also be discussed. Moreover, the ablations in Appendix H mainly focus on sampling steps, regularization, and intervention counts. I feel that the individual contributions of the downstream Jacobian correction and adaptive intervention timing are not sufficiently distinguished.

References
[1] Baldan, Giacomo, et al. "Physics vs distributions: Pareto optimal flow matching with physics constraints." International Conference on Learning Representations. Vol. 2026. 2026.
[2] Denker, Alexander, et al. "Trajectory Stitching for Solving Inverse Problems with Flow-Based Models." arXiv preprint arXiv:2602.08538 (2026).

Questions for authors

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

Q1. Can the authors provide a detailed computational breakdown of their algorithm?

Q2. How does the cost scale for pseudoinverse correction and the adjoint methods, mainly for multiple (vector-valued) constraints? How do the authors keep the computational overhead tractable?

Q3. What version of the MintFlow algorithm is used for each experiment?

6 comments
AuthorsOpenReview ↗

Responses to Reviewer cpw2's Weaknesses 1 and 2

W1: FID and W2 are not consistent.

We thank the reviewer for catching this and apologize for the error. FID and W2 are computed from the same feature representations and Gaussian approximations, so W2=FIDW_2=\sqrt{\text{FID}}. The discrepancies came from typos when transcribing the W2 entries into Tables 6-9. The implementations and the FID values are correct. We have corrected all affected W2 values. We checked all other reported metrics and found no errors. Since W2=FIDW_2=\sqrt{\text{FID}} is monotone, the corrected W2 values have the same ranking as FID. Thus, our comparison and conclusions are unchanged. We thank the reviewer again for bringing this issue to our attention, which has helped us improve the accuracy of the manuscript.

W2-1: Clarification on the algorithm, its implementation, and how it relates to rectified flow.

We thank the reviewer for pointing out the lack of clarity regarding the relationship between our general algorithm, its rectified-flow (RF) specialization, and their practical implementations. We have revised the manuscript to address these issues.

General flow matching (FM) Our general method is based on the adjoint-based formulation, which is presented in two complementary forms: Algorithm 1 provides a concise overview, while Algorithm 2 details the full computational procedure, including forward integration, backward adjoint computation, intervention-time selection, and forward re-integration. This formulation is used for conventional FM in the image generation experiments (Appendix D) and functional FM in the physics-informed generation experiments (Section 4.3).

Rectified flow (RF) For RF models, we exploit their approximately linear transport property to simplify the sensitivity computation. As detailed in the revised Appendix A.4, the terminal state is approximated by a single Euler step from the intermediate state. Thus, a single VJP gives the sensitivity needed for the mininum-norm correction, without solving the adjoint ODE.

For our experiments on image inverse problems (Section 4.1) and image editing (Section 4.2) using rectified flow (RF) models, we largely follow the experimental setup of FlowChef (Patel et al., 2025), which employs multi-step RF sampling. In FlowChef, all compared methods perform their respective corrections at each sampling step. Following the same protocol, we apply our minimal-intervention correction at every sampling step using an RF-specific simplified formulation, followed by an Euler step to advance the corrected state to the next sampling time. The complete procedure is presented in the new Algorithm 3.

The revised appendix now states which algorithm each experiment uses. We hope this clarifies the distinction between the general FM formulation and its simplified RF implementation.

W2-2: Details on the implementation of the pseudoinverse.

For the constraint-by-state sensitivity matrix AsA_s computed via the adjoint equation, we apply As†A_s^\dagger using a compact SVD, discarding singular values below a relative threshold. If the SVD fails, we fall back to a least-squares solver. Importantly, the pseudoinverse As†A_s^\dagger is never explicitly formed; instead, we directly compute its product with the constraint residual. We provide these implementation details in Appendix G. (Link to the code https://anonymous.4open.science/r/MintFlow_PDE_Application-4EB0/)

For the computational efficiency of MintFlow with high-dimensional vector-valued constraints, please refer to our response to W4.

AuthorsOpenReview ↗

Responses to Reviewer cpw2's Weakness 3

W3-1: What is the exact benchmark protocol?

We thank the reviewer for requesting this clarification. We adapted the PCFM protocol to better expose the trade-off between preserving pretrained priors and enforcing constraints.

  • Constraints. For heat, we retain PCFM’s IC and mass-balance constraints and encode periodicity structurally. For reaction–diffusion, we additionally enforce each prescribed boundary flux. This is necessary because global mass balance alone does not guarantee satisfaction of individual boundary flux conditions: errors at different boundaries can cancel in the net flux, even when the prescribed flux at each boundary is violated.
  • Evaluation. PCFM uses fixed-IC ensembles with varying diffusivity or boundary fluxes. We instead use sample-specific ICs and physical conditions derived from a shared Vanilla FFM bank, with identical targets across methods. This compares all methods under the same prescribed physical conditions at each sample index.

W3-2: Confusion cased by the sentence “ECI and PCFM enforce the constraint through a final projection step.”

We apologize for the confusion. We agree that ECI and PCFM are not post-hoc projection methods. Indeed, in our experiments, we used the official implementations of both methods in their entirety, rather than applying only their final projection steps. We classify them as projection-based methods because their core operations involve projecting the predicted terminal states onto the constraint set. However, this does not mean that they simply apply a post-hoc projection to unconstrained samples after generation. We have revised the relevant descriptions throughout the paper to clarify this distinction and avoid further ambiguity.

W3-3: The discrepancy between the ECI/PCFM performance reported in the original PCFM study and our paper.

As previously noted, variations in specific constraints and evaluation protocols yield differing baseline performance numbers. Notably, our definition of simulation error differs from that of PCFM: PCFM computes simulation error utilizing ensemble mean fields, whereas our framework evaluates per-sample simulation error.

The reviewer raises an insightful question regarding why Vanilla FFM outperforms PCFM and ECI. We think Vanilla FFM can outperform PCFM and ECI on physical consistency because satisfying the initial condition (IC), boundary conditions (BC), and mass conservation does not necessarily guarantee adherence to the underlying PDE dynamics. For example, in the heat equation, a time-varying, zero-mean periodic perturbation that vanishes at the initial time can preserve these constraints while violating the governing PDE. Consequently, repeated constraint corrections may introduce distortions to the learned dynamics without necessarily improving PDE consistency. This observation underscores the primary objective of MintFlow, which is to achieve a balance between distribution fidelity and constraint fulfillment.

W3-4: Newton step 10 for heat.

Thank you for bringing this to our attention. We double-checked our code and confirmed that this was misreported. For both the heat and reaction-diffusion equations, the number of Newton steps for PCFM correction is set to 1.

AuthorsOpenReview ↗

Responses to Reviewer cpw2's Weakness 4 and Questions

W4-1: What is the computational cost for multiple (vector-valued) constraints.

We thank the reviewer for raising this important question. We agree that handling multiple vector-valued constraints introduces additional computational costs in MintFlow. The following table reports the sampling time and peak GPU memory across all methods for our PDE experiments.

Sampling methodHeat: Time/sample (s)Heat: Peak GPU memory (GiB)Reaction Diffusion: Time/sample (s)Reaction Diffusion: Peak GPU memory (GiB)
PCFM2.1460.25671.7680.24
ECI3.2690.63891.4130.5536
DiffusionPDE0.086690.63770.40890.552
D-Flow2.5341.5846.2391.743
MintFlow7.2674.69775.696.357

To better understand this overhead, we profiled the individual components of MintFlow.

ComponentHeat timeRD timeHeat peak GPU (GiB)RD peak GPU (GiB)
Initial forward flow0.831 s0.578 s0.1630.169
Terminal constraint and Jacobian0.027 s0.034 s0.1540.193
Adjoint solves32.350 s76.490 s2.1862.853
Ten candidate pseudoinverse solves0.201 s0.309 s0.2370.423
Candidate selection0.010 s0.005 s0.214 GiB0.360
Corrected forward suffix0.159 s0.011 s0.2430.393
Final projection, after sampling0.058 s0.632 s0.1700.296

The main bottleneck is not the pseudoinverse computation but the backward adjoint integration, particularly the batched vector–Jacobian products (VJPs) required for high-dimensional constraint outputs (Table 13 in the revised manuscript). Specifically, at each Euler step of the adjoint integration $ A_{t_i}=A_{t_{i+1}}\left(I+h_iJ_f(u_{t_i},t_i)\right), $ where hi=ti+1−tih_i=t_{i+1}-t_i, each matrix-Jacobian product Ati+1JfA_{t_{i+1}}J_f is computed using chunked vector-Jacobian products (VJPs) via torch.autograd.grad with is_grads_batched=True. The reaction–diffusion problem involves 425 constraint rows, requiring seven 64-row chunks per reverse step, compared with four chunks for the 199-row heat constraint.

To remedy this, promising future directions include hardware-level optimizations such as CUDA profiling (e.g., chunk-size sweeps to balance time and memory) and streaming candidate storage to prevent cloning full sensitivity tensors at every step of the reverse pass. We provide a detailed computational analysis and discussion of these limitations and potential improvements in Appendix G.7.

Nevertheless, this overhead does not diminish the broader significance of MintFlow. First, MintFlow is highly efficient for scalar-valued constraints, which cover a broad range of applications, including image generation, inverse problems, and editing demonstrated in our paper. Second, despite the additional computational cost, MintFlow achieves favorable trade-offs between distributional fidelity and constraint satisfaction compared with existing methods in our PDE experiments. Ultimately, the central contribution of MintFlow lies in its minimal-intervention formulation and the derivation of a closed-form minimum-norm correction under local linearization, together with theoretical support on constraint satisfaction and distributional preservation, rather than computational efficiency alone. Accordingly, we have revised the paper to clarify its computational limitations, particularly for high-dimensional vector-valued constraints.

W4-2: The comparison with existing literature could be improved.

We thank the reviewer for the helpful suggestions and for pointing us to these relevant works. We have substantially revised the Related Work section to provide a more comprehensive discussion, including the suggested papers and clearer comparisons with closely related methods. In particular, we focused on more explicitly distinguishing the underlying mechanisms and practical contributions of existing approaches from those of MintFlow.

W4-3: The individual contributions of the downstream Jacobian correction and adaptive intervention timing are not sufficiently distinguished.

Regarding the term downstream Jacobian correction, could the reviewer please clarify what this refers to in the context of our method?

Regarding the contribution of adaptive intervention timing, please see our detailed response to Weakness 2 raised by Reviewer R1em.

Q1: Can the authors provide a detailed computational breakdown of their algorithm?

We refer the reviewer to our response to W2.

Q2: How does the cost scale for pseudoinverse correction and the adjoint methods?

We refer the reviewer for our response to W4-1.

Q3: What version of the MintFlow algorithm is used for each experiment?

Please refer to our response to W2-1, where we explicitly clarify which version of MintFlow is used for each experiment.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to the Area Chair and Reviewers

Dear Reviewers and AC,

We sincerely appreciate for your time, effort, and constructive feedback. We particularly appreciate the reviewers' thoughtful questions and suggestions, which have helped us clarify and improve our manuscript.

In response to the feedback, we have carefully revised the manuscript, with all changes highlighted in blue for ease of reference. The major revisions and additions are summarized below:

  • Clarification of the algorithm: We provide additional explanations and details of the simplified algorithm for rectified flow models in Appendix A.4 and Algorithm 3. We also specify which algorithm is used for each experiment in Appendix B.

  • Implementation details for vector-valued constraints: We clarify how the minimum-norm correction is computed for vector-valued constraints, including the pseudoinverse implementation, in Appendix G.

  • Correction of W2W_2 values: We correct typographical errors in the reported Wasserstein-2 (W2W_2) values in Tables 6–9. Since W2W_2=√FID and the FID values are correct, rankings and conclusions are unchanged.

  • Additional experimental details: We provide further details on the benchmark protocols and experimental hyperparameters to improve reproducibility (Appendix G).

  • Additional ablation on adaptive versus fixed intervention times: We conduct additional experiments comparing adaptive and fixed intervention time selection (Appendix H.4).

  • Clarification of theoretical interpretations: We revise the wording and interpretations of our theoretical results to more precisely reflect the scope and implications of the guarantees.

  • Strengthening of related work: We expand the discussion of related work to better contextualize MintFlow and clarify its distinctions from existing approaches.

We provide detailed, point-by-point responses to each reviewer's questions and comments in the individual responses below. We hope that these revisions adequately address the reviewers' concerns and make the contributions, methodology, and theoretical implications of MintFlow clearer.

Best,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.