acceptodds
Under review as a conference paper at ICLR 2027

Solving Inverse Problems of Distributions with Wasserstein Flow Maps

Abstract

Many scientific datasets consist of distributions rather than individual points, such as point clouds, cell populations or spatial tissue neighborhoods. Recovering populations from partial or aggregate observations, and steering their generation toward desired properties, call for generative priors that can adapt to different measurement operators and guidance objectives. Wasserstein flow matching provides such priors, but obtaining terminal populations requires numerical integration, making repeated endpoint prediction for inference-time guidance computationally expensive. We introduce Wasserstein flow maps, a framework for learning finite-time transport maps between distribution-valued states, enabling direct prediction of terminal populations. We formulate these maps on Wasserstein space and derive semigroup, Eulerian, and Lagrangian characterizations. The resulting maps support deterministic reward lookahead and flow map trajectory tilting, allowing one pretrained prior per domain to address multiple inverse problems and guidance objectives without task-specific retraining. Across ShapeNet, single-cell populations, and spatial transcriptomic niches, Wasserstein flow maps improve generation metrics over Wasserstein flow matching baselines using 8 rather than 100 function evaluations, and outperform inverse problems methods on inpainting, super-resolution, deconvolution, and guided generation. On 100 held-out patients, lookahead guidance reduces reconstruction MMD by 35% for deconvolution from mean single-cell RNAseq embeddings and masked MMD by 37% for cell-type inpainting, relative to task-specific fine-tuned Wasserstein flow matching.

Primary area:
generative models
Keywords:
flow map, inverse problem, generative model, AI for biology

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_CDD5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2(+1)
Presentation
3(+1)
Soundness
3(+1)
Read review
Paper summary

This paper introduces Wasserstein flow maps for direct terminal population prediction, supporting few-step generation and inverse-problem guidance. The problem is practically relevant, but the contribution currently appears to be mainly an extension of existing frameworks. The value of the additional technical developments remains somewhat unclear. The experiments show improvements, yet do not fully separate the effects of flow-map learning, training choices, and auxiliary components. Overall, I have some reservations about the strength of the contribution and the support for the conclusions in the current version.

Critical strengths

1.The method addresses the cost of repeated terminal integration, with a clear connection between the motivation and the proposed design.

2.Comparing direct map prediction with velocity-field integration using the same network provides convincing evidence for few-step generation. Experiments across several domains and observation types also demonstrate the framework’s practical value.

Critical weaknesses

1.The contribution of the new design remains unclear. The approach builds on existing flow-map and guidance frameworks, while the comparisons also change the prior, training setup, and auxiliary composition predictors. I would appreciate a clearer account of the new technical contributions and comparisons under matched conditions.

2.The connection between theory and implementation needs further validation. Lagrangian training uses a surrogate objective, and FMTT omits some correction terms. The authors acknowledge these approximations, but direct checks of flow consistency and sampling bias are still missing.

3.The voxel deconvolution objectives are not equivalent. The implemented pointwise loss additionally penalizes within-voxel variance and therefore does not share the same set of minimizers as the stated mean-based loss. Please correct this claim and examine how the additional constraint affects reconstruction.

4.The efficiency comparison should account for the full cost. Reported NFE does not adequately capture gradient computation, score queries, and multiple FMTT walkers. Inference time and peak memory would provide a clearer picture of the practical advantage.

Questions for authors

1.Do the main gains persist when the prior, training setup, and composition predictors are matched?

2.How large are the flow-consistency errors and sampling biases introduced by the surrogate training and simplified correction?

3.Do the reconstruction and within-voxel structure results hold when using only the voxel-mean loss?

4.What are the runtime and memory costs per output when all gradient computations and walkers are included?

4 comments

We thank the reviewer for the thoughtful and constructive assessment. We are glad that the reviewer recognizes the “clear connection between the motivation and the proposed design,” finds our same-network comparison to provide “convincing evidence for few-step generation,” and appreciates the framework's practical value across different domains and observation types.

We particularly appreciate the opportunity to clarify our technical contributions and strengthen the empirical evidence. Our goal is to develop a reusable generative framework for scientific inverse problems where the unknown is an entire population rather than an individual sample. By learning finite-time transport on probability measures, our approach enables direct endpoint prediction and observation-dependent guidance without retraining the generative backbone. This is particularly relevant to scientific settings involving partial observations, aggregate measurements, or population-level constraints.

To address the reviewer's concerns, we conducted additional matched comparisons, flow-consistency and sampling-bias evaluations, corrected voxel-deconvolution experiments, and end-to-end computational benchmarks, as detailed below.

Weakness / Question 1 — Do the gains persist with matched training, priors, and conditioning?

A1. We agree that the original comparisons involve differences in prior training and auxiliary conditioning. As detailed in our response to Reviewer vj6R (W/Q 1, W/Q 2), we conducted matched-training and fixed-checkpoint comparisons to separate prior quality from inference-time guidance. These experiments show that lookahead improves reconstruction on fixed priors, while the effect of training choices varies across inverse tasks. Additional guided-versus-unguided controls using identical inferred composition tokens confirm that the single-cell improvements are not solely attributable to auxiliary predictors. We include these comparisons and clarify the respective contributions of prior training, flow-map learning, and guidance in the revision (App I.3, I.5).

W/Q 2 — How large are the flow-consistency errors and FMTT sampling bias?

A2. We agree that the theoretical flow identities alone do not establish the accuracy of learned maps or the simplified FMTT correction. We therefore evaluated both approximations directly.

Lagrangian surrogate and flow consistency.

Our Lagrangian objective follows [1], using a back-predicted input, conditional-velocity supervision, and cycle regularization. To quantify flow-consistency errors, we evaluated Lagrangian-trained maps in all three domains along their sampling trajectories, measuring cycle reconstruction, semigroup composition, and the forward (Lagrangian) and endpoint (Eulerian) residuals in Eqs. 3.9–3.10.

DomainCycle ↓Semigroup composition ↓Forward ↓Endpoint ↓
ShapeNet0.0060.0440.0170.033
Single cell0.0370.0270.0210.103
Spatial niche0.1580.0840.0800.180

Normalized relative errors (lower is better). Cycle and composition errors are normalized by displacement; differential residuals by the sum of their constituent terms.

The residuals are generally small on ShapeNet, with larger deviations on spatial niches. To isolate the effect of cycle regularization, we additionally retrained the niche model with and without the cycle term (3 paired seeds, 50k updates).

Metric ↓Without cycleWith cycle
Cycle error0.4530.127
Forward / endpoint residual0.117 / 0.2340.068 / 0.152
1-step generation MMD0.4590.378
8-step generation MMD0.3700.371
Guided reconstruction MMD0.2070.220

Cycle regularization substantially improves flow consistency and 1-step generation, but 8-step generation is essentially unchanged, and guided reconstruction does not improve. Thus, stronger consistency does not necessarily translate into better downstream reconstruction, consistent with RMF's observation that the benefits of cycle regularization can be task-dependent [1].

AuthorsOpenReview ↗

part 2 of the response

(cont'd A2) Simplified FMTT correction and sampling bias.

We additionally evaluated the effect of restoring the backward-flow residual in FMTT, both on a known-target Gaussian example and on spatial-niche enrichment. The controlled example confirms that while the omitted term can introduce notable bias for an inconsistent flow map, its empirical effect on spatial-niche enrichment is negligible. We provide the full setup and quantitative results in our response to vj6R (A4).

We observe that the omitted correction has negligible measurable impact on our target-error metric in this evaluated setting, despite imperfect flow consistency. We note that this does not establish exact distributional fidelity.We incorporate these consistency and sampling-bias evaluations into the revised manuscript and clarify the theoretical assumptions underlying the practical approximations (App E.4).


W/Q 3 — How do voxel-loss differences affect reconstruction?

A3. We agree that the implemented per-point penalty is not equivalent to the stated voxel-mean objective: it additionally penalizes within-voxel variance. To this end, we invite reviewer to our response to reviewer vj6R (A3). We remove the incorrect equivalence claim, and reran voxel deconvolution using the stated mean-based objective.

The revised results preserve competitive reconstruction performance of lookahead and FMTT. These results support the qualitative reconstruction findings while quantifying the effect of removing the additional variance penalty. The corrections are reflected in the updated manusript (App G.2, H.1).


W/Q 4 — What are the full inference-time and peak-memory costs?

A4. We agree that NFE alone does not fully capture the costs of gradient-based guidance, score evaluations, and FMTT walkers. We therefore profiled per-output inference latency and peak GPU memory, accounting for these computations under the inference settings used in the manuscript.

All measurements use a single NVIDIA B200 and generate one complete population per call: ShapeNet inpainting (2,048 points), bulk-to-single-cell deconvolution (1,024 cells), or spatial-niche super-resolution (128 cells).

TaskMethodSteps (walkers)Time (ms) ↓Peak memory (GB) ↓
ShapeNetLookahead3279.51.12
FMTT32 (4)276.44.09
FlowDPS100134.80.24
PnP-Flow100133.60.24
FT-WFM100132.90.24
Single cellLookahead1632.40.42
FMTT16 (4)59.61.34
FlowDPS100109.80.16
PnP-Flow10092.00.16
FT-WFM10091.10.16
Spatial nicheLookahead87.40.07
FMTT8 (4)9.90.11
FlowDPS10034.20.10
PnP-Flow10033.60.10
FT-WFM10032.70.10

Median of 12 synchronized warm inference calls at batch size 1, after compilation. Peak memory includes model parameters. FMTT's four walkers are evaluated together in a single batch. Training and compilation costs are excluded.

Lookahead is 1.6–4.6× faster than the WFM-based baselines across these benchmarks, including the cost of gradient computation. This supports the practical value of directly learned finite-time maps for repeated endpoint-based guidance. However, lookahead uses more peak memory than the WFM-based methods on ShapeNet and single-cell data, reflecting the cost of differentiation through the learned map. FMTT is slower and more memory-intensive, while these costs depend on the number of walkers and implementation: greater GPU parallelization could reduce FMTT's latency, potentially at the expense of additional memory. These results are included in the revised manuscript (App F.5).

[1] Woo, D., Skreta, M., Park, S., Neklyudov, K., and Ahn, S. Riemannian MeanFlow. ICML, 2026

We believe the new analyses and clarifications resolve the key issues raised by the reviewer, and we would greatly appreciate their positive reconsideration of our work.

Thanks for the detailed response and for running all these extra experiments—I really appreciate how candidly you checked the approximations and corrected the voxel-loss issue. The matched controls, consistency diagnostics, and runtime/memory benchmarks make the paper much clearer and more transparent.

Looking closely at the new results, though, I still have a couple of hesitations:

First, on novelty and matched comparisons: the method still feels mostly like lifting Riemannian MeanFlow and existing guidance techniques to Wasserstein space, and once the priors and training setups are matched, the gains are a bit mixed—for instance, 8-step WFM lookahead slightly edges out the learned map on super-resolution, and under the corrected voxel-mean loss, FT-WFM gets lower MMD and within-voxel error.

Second, the new diagnostics nicely illustrate the theory–practice gap: when flow consistency isn't tight, as on spatial niches, simplified FMTT can pick up noticeable sampling bias. The profiling also highlights real practical trade-offs: lookahead gives a nice speedup at the cost of higher peak memory, while FMTT ends up slower and much heavier on memory than WFM baselines on larger point clouds like ShapeNet.

All in all, your rebuttal addressed my concerns about experimental rigor and clarity, so I will raise my Soundness and Presentation scores accordingly. However, given the modest novelty and the mixed trade-offs above, I am leaning toward keeping my overall rating as is.

We thank the reviewer for the thoughtful follow-up and raised scores on Soundness and Presentation. We are glad that our new results addressed their concerns about experimental rigor, made paper much clearer and more transparent, and that the new diagnostics show the theory-pratice gap nicely.

We address the remaining concerns below.

[…] feels mostly like lifting Riemannian MeanFlow and existing guidance techniques to Wasserstein space

Our contribution is not a new finite-particle regression loss, but deriving finite-time flow maps from measure-valued dynamics, establishing their connection to a tractable particle realization, and applying them to inverse problems over distributions. We kindly invite the reviewer to our additional response to reviewer JR2s for a detailed discussion of this distinction.

[…] gains are a bit mixed […] 8-step WFM lookahead slightly edges out the learned map on super-resolution,

Our contribution is not predicated on uniformly outperforming numerical rollout across all tasks. While 8-step WFM gives better result on super-resolution, our learned map lookahead outperforms on two other tasks while avoiding repeated numerical integration for endpoint estimation. We also note that flow map with learned map endpoit prediction achieves 8.9× faster end-to-end inference compared to 8-step rollout with WFM in this comparison.

PriorLookahead endpointInference time (ms)
Wasserstein Flow Map (Ours)learned map79.5 ± 0.9
WFM1-step59.9 ± 0.3
WFM8-step rollout709.2 ± 7.5

Inference times are averaged across the three inverse tasks.

[...] under the corrected voxel-mean loss, FT-WFM gets lower MMD and within-voxel error.

Our central goal is to enable a single frozen generative prior to solve diverse inverse problems through inference-time guidance, without task-specific fine-tuning. FT-WFM is a strong baseline precisely because it is fine-tuned for each task, requiring additional training. Despite this distinction, our frozen flow map outperforms FT-WFM on several important inverse problems: e.g., single-cell bulk deconvolution (0.060 vs. 0.093 MMD) and cell-type inpainting (0.385 vs. 0.612 MMD).

[...] when flow consistency isn't tight, as on spatial niches, simplified FMTT can pick up noticeable sampling bias

We would like to clarify that the noticeable sampling bias was observed in our synthetic experiment with a deliberately inconsistent flow map, rather than on spatial-niche experiment. Although flow-consistency errors are larger on spatial niches, restoring the omitted FMTT residual had negligible effect on the measured target error (+0.0001, 95% CI [−0.0014, 0.0015]). Thus, while larger consistency residuals identify a potential source of bias, they do not themselves establish substantial sampling bias in this experiment.

lookahead gives a nice speedup […] while FMTT ends up slower and much heavier on memory than WFM baselines on larger point clouds like ShapeNet

FMTT is intended as a complementary inference method with different accuracy–compute trade-offs, rather than a universally preferable alternative. For instance, on spatial-niche enrichment, it improves target-composition accuracy over lookahead and even task-specific FT-WFM at medium and low target abundances, albeit at higher computational cost. Its practical value therefore depends on the inverse task and available compute budget.

We believe these clarifications address the reviewer's remaining concerns and further substantiate the theoretical and practical contributions of our work. We would greatly appreciate the reviewer reconsidering their overall assessment in light of the additional evidence.

Reviewer_JR2sOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper extends ODE flow map characterizations to flow models on Riemannian geometries where each sample is a probability measure, termed Wasserstein flow maps. The training objectives appear to display substantial overlap with Riemannian MeanFlow (RMF). The main difference to RMF enters via the set-transformer parameterisation and whole-population input state. The paper adapts reward guidance algorithms to solving population-valued inverse problems and demonstrates performance on point cloud and cellular domains.

Presentation

In Sections 3.2, 3.3 and 3.4, the flow map objectives are stated. The standard Lagrangian and Eulerian residuals (Eqs 3.9 and 3.10) are well established (Boffi et al. 2025) and this section aims to extend these conditions to the Wasserstein and Riemannian space. On inspection of the implemented objectives in Appendix D, the overlap with Riemannian MeanFlow (RMF) becomes clear. The Eulerian and Lagrangian objectives appear to apply RMFs construction on the product manifold. The semigoup loss is also discussed in RMFs comparison to Generalised Flow Maps, where it identifies a relationship between the two via a loss weighting. The relationship to RMFs is acknowledged in Section 2, but the present work does not make it clear what contribution is made beyond RMFs or if any empirical performance is gained by these changes.

Informativeness

The main contributions of the paper are to extend flow map objectives to Wasserstein flows. These are summarised in Proposition 3.1. However, it is unclear how this methodology extends beyond RMFs construction. Empirically, the use of reward guidance to solve inverse problems over populations is interesting. The results here are promising in comparison to the baselines shown and appear to be a useful general abstraction of the problem. Nevertheless, the comparison to RMFs is not provided and it is therefore challenging to assess the contribution.

Soundness

The mathematical demonstrations appear reasonable on inspection. My concern lies in the relationship of these objectives to Riemannian MeanFlows and the present papers contributions beyond this prior work, rather than technical correctness.

Critical strengths

The paper investigates a clear problem domain: reward guidance for inverse problems over populations. The paper presents an extension of standard ODE flow map objectives to Wasserstein spaces to model population-valued states. This extension is supported by promising empirical results for unconditional generation and inverse problem applications on point cloud and cellular domains.

Critical weaknesses

The main weakness of this work is that it appears to inherit the objectives from Riemannian MeanFlows without sufficient discussion of the present works contributions beyond these objectives. My understanding is that applying the Riemannian MeanFlow algorithm with a permutation-invariant population-dependent network would recover the current works implementation for Lagrangian and Eulerian conditions. The endpoint semigroup condition is also presented in Appendix C.2 of RMFs.

Questions for authors
  • What is the mathematical difference between the Riemannian MeanFlow Semigroup, Lagrangian, and Eulerian objectives and the objectives in Equations D.2, D.3, and D.4, respectively?
  • What is the implementation difference between the algorithm described and Riemannian MeanFlow objectives, architecture and interpolant, trained on whole-population network inputs?
  • If there are algorithmic differences to Riemannian MeanFlows with a population-valued input state, can they be isolated and empirically tested in one guidance experiment?
4 comments

We thank the reviewer for the thoughtful assessment. We are glad that the reviewer finds reward-guided inverse problems over populations “interesting,” the empirical results “promising,” and our formulation “a useful general abstraction of the problem.” We also appreciate the careful comparison with Riemannian MeanFlow (RMF), which gives us an opportunity to clarify the technical contributions beyond prior work.

While our finite-particle Eulerian and Lagrangian losses naturally coincide with RMF under a population-dependent product-manifold formulation, they arise from a different starting point: measure-valued dynamics on Wasserstein space. Our contribution lies in deriving finite-time flow-map identities, constructing a permutation-equivariant, population-coupled particle realization, and leveraging these maps for efficient guidance in distributional inverse problems.

While RMF learns flow maps on fine-dimensional Riemannian manifolds, generating one sample at a time, our Wasserstein Flow Maps operate on the infinite dimensional Wasserstein space of distributions defined on an arbitrary Riemannian manifolds, enabling our framework to generate complete distributions at once.

We detail the mathematical correspondence and new controlled comparisons below.

Weakness / Question 1 — How do the training objectives relate to RMF?

A1. We agree with the reviewer that our Eulerian (D.3) and Lagrangian (D.4) objectives coincide with RMF's Eqs. 10–12 when RMF is applied to the product manifold (MN\mathcal M^N) with population-dependent inputs. This equivalence arises naturally when we specialize our measure-level flow-map identities to empirical distributions and express them in average-velocity coordinates. The semigroup objectives differ in form: our Eq. D.2 uses the endpoint-consistency (G-PSD) loss of Generalised Flow Maps, whereas RMF's Eq. 13 regresses average velocity.

Our contribution lies not in new finite-particle loss forms, but in their derivation from measure-valued dynamics and their role in learning flow maps for distributional inverse problems, as detailed in our answer to question 2 below.


W/Q 2 — How does our implementation differ from RMF with whole-population inputs?

A2. We agree that RMF applied to MN\mathcal M^N with our population-dependent architecture can reproduce our finite-particle Eulerian and Lagrangian objectives. The interpolant and other training choices are also compatible with RMF. The distinction lies in our measure-level formulation and its connection to distribution-valued inverse problems. Rather than starting from a flow on a fixed-dimensional product manifold, we begin with measure-valued dynamics and their finite-time solution operator Ψs,t:P2(M)→P2(M)\Psi_{s,t}:\mathcal P_2(\mathcal M)\to\mathcal P_2(\mathcal M).

We derive the corresponding flow-map identities at the distribution level and then specialize them to empirical measures. Our shared, population-conditioned and permutation-equivariant particle map consequently induces a well-defined map on empirical distributions, Ψ^s,tθ(μX)=μΦs,tN,θ(X)\widehat\Psi_{s,t}^{\theta}(\mu_X) =\mu_{\Phi_{s,t}^{N,\theta}(X)}.

This connects finite-particle transport to the underlying measure-valued dynamics, rather than treating populations solely as labeled configurations. App. B.1 further establishes the connection to Wasserstein geometry through the permutation quotient, while App. C.1 provides well-posedness and stability results for the ideal population dynamics.

More importantly, this formulation provides the basis for our distributional inverse-problem framework. Observations and rewards act on entire populations, and Eq. E.4 connects particle-level reward tilting to target laws over empirical measures. We exploit this structure to develop lookahead and FMTT guidance for distribution-valued unknowns using a shared pretrained prior.

Thus, we do not claim a fundamentally different finite-particle RMF algorithm. Our contribution is the measure-level derivation, its compatible particle realization, and their integration into a framework for inference-time guidance in distributional inverse problems.

AuthorsOpenReview ↗

part 2 of the response

W/Q 3 — Can the differences from RMF be isolated in a controlled comparison?

A3. To assess the empirical differences, we conducted ablation on bulk-to-single-cell deconvolution, varying both pointwise versus population-dependent networks and RMF's average-velocity versus endpoint semigroup training.

The models share the same architecture apart from cross-particle interactions, as well as matched initialization, spherical source, independent coupling, time sampling, diagonal anchor, optimizer, and 250k training updates. All methods use identical lookahead guidance (16 steps), source draws, and a frozen composition inferrer, and are evaluated on 100 held-out patients across 3 paired training seeds.

NetworkResidualMMD ↓Composition r ↑
PointwiseRMF, Eq. 130.120 ± 0.0710.946
PopulationRMF, Eq. 130.087 ± 0.0070.935
PopulationEndpoint, D.20.051 ± 0.0000.967

MMD is reported as mean ± SD over 3 training seeds. Lower MMD and higher composition correlation are better.

The results establish two complementary findings. We see that population interactions provide a notable benefit, where MMD decrease from 0.120 to 0.087 when using population model. Additionally, endpoint semigroup training improves reconstruction under the same population architecture. MMD decreases from 0.087 with RMF's semigroup loss to 0.051 with endpoint consistency. This comparison isolates the training objective without changing the population network.

While we do not claim the novelty of endpoint objective itself, these controls demonstrate the practical benefit of combining population-dependent modeling with endpoint consistency for the evaluated inverse problem. We now include further discussions on RMF in Sec. 3.4 and App D.2, and the controlled comparisons in App I.2 of updated manuscript.

We hope that the clarified relationship to RMF and the additional controlled experiments address the reviewer's concerns regarding novelty and empirical contributions. We would greatly appreciate the reviewer reconsidering their assessment in light of these clarifications and results.

Thank you for the clarifications. We are agreed that the final loss implementation is the same as RMF and the contribution lies in the mathematical derivation of these losses for measure-valued maps (or, empirical populations in practice).

The table in A3 appears to conflate architecture and objective: the table compares the RMF eq 13 objective to the endpoint objective but the architectures are different ("apart from cross-particle interactions"). My understanding would be that under a shared architecture these results would be similar.

The added discussion on the comparison to RMF is an improvement, but the general framing of the contributions as "Wasserstein Flow Maps" is over-claiming given the implementations of WFMs and RMFs collapse to the same algorithm in practice for population-valued inputs.

Overall, the claims of the paper do not appear to reflect the contributions and I therefore maintain my rating.

We appreciate the reviewer's acknowledgment of our theoretical contribution in deriving flow-map objectives from measure-valued dynamics. We address the remaining concerns below.

table in A3 appears to conflate architecture and objective [...] under a shared architecture these results would be similar.

We apologize if our description of the A3 experiment was unclear. The architecture is identical between RMF and endpoint training, including cross-particle interactions; only the semigroup objective changes. Under this same population-dependent architecture, endpoint training achieves 0.051 vs. 0.087 MMD for RMF's semigroup objective. The ablation illustrates the practical importance of the objective choice in the population-valued inverse problem.

[...] framing of the contributions as "Wasserstein Flow Maps" is over-claiming given the implementations of WFMs and RMFs collapse to the same algorithm in practice for population-valued inputs.

We emphasize that equivalent final finite-particle losses do not necessarily imply equivalence of the mathematical contributions. Below, we clarify why we believe our work, despite sharing objectives with RMF, is a meaningful contribution.

1. The central contribution is not a new finite-particle training objective, but a framework for generative modeling and inference over distributions.

The fundamental objects in our work are not individual points (x∈M)(x\in\mathcal M), but entire populations (μ∈P2(M))(\mu\in\mathcal P_2(\mathcal M)). We model a probability law (P∈P2(P2(M)))(P\in\mathcal P_2(\mathcal P_2(\mathcal M))) over these populations. For instance, in single-cell modeling, one generated sample represents an entire patient-specific cell population, rather than an individual cell.

This distinction is essential to the scientific problems we address, because observations such as aggregate gene expression or partially observed tissue neighborhoods constrain whole populations rather than an individual point. Our goal is to reconstruct plausible populations given these observations while capturing variability between possible population-level solutions.

2. We agree that RMF provides the relevant finite-particle flow-map machinery, but this does not exhaust the contribution of our formulation.

As detailed in our initial repsonse (A2), our starting point is general population-dependent continuity equation on (P2(M))(\mathcal P_2(\mathcal M)), whose finite-time solution operator is (Ψs,t:P2(M)→P2(M))(\Psi_{s,t}:\mathcal P_2(\mathcal M)\to\mathcal P_2(\mathcal M)). Building upon this general framework, we then establish the measure-level flow identities and their relationship to empirical particle maps.

Importantly, our analysis establishes well-posedness and Wasserstein stability of the ideal population dynamics (App. C.1) and connects finite-particle representations to Wasserstein geometry through the permutation quotient (App. B.1).

These results provide a mathematical connection between measure-valued dynamics and their finite-particle realization. We agree that, once restricted to fixed-size empirical measures, the resulting Eulerian and Lagrangian losses coincide with RMF on the product manifold. This equivalence emerges from our measure-level derivation; it does not replace the theoretical results established by that derivation.

3. Our main practical contribution is to use these population flow maps to solve distribution-valued inverse problems with a reusable generative prior.

In Sec. 4, we formulate inverse problems through observation-conditioned laws over populations, dPy(μ)∝ery(μ)dPprior(μ).dP_y(\mu)\propto e^{r_y(\mu)}dP_{\mathrm{prior}}(\mu). We connect this measure-level inference target to particle-based implementations and adapt both deterministic lookahead and Feynman–Kac trajectory tilting to population-valued unknowns. Unlike RMF's focus, our framework explicitly addresses population-level observation operators and rewards using a shared frozen distribution-level generative prior. We demonstrate this across point-cloud, single-cell, and spatial-niche inverse problems.

4. A shared loss does not fully characterize a method's contribution

We believe it is important to distinguish equivalence of training objectives from equivalence of scientific contributions. Sharing a loss function does not make two modeling frameworks equivalent in their scope or purpose: many fundamentally different models, from linear regression to deep neural networks, are trained using MSE, while they can differ in the functions they represent and the problems they enable us to solve.

We recognize the algorithmic overlap with RMF. However, our novelty lies in the measure-level formulation, its theoretical connection to empirical populations, and framework for distributional inverse problems—not in the loss itself. We believe these contributions merit consideration alongside shared finite-particle objectives when assessing the paper's novelty.

Reviewer_vj6ROpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper studies inverse problems in which the unknown is a population or probability distribution. It proposes Wasserstein flow maps that directly predict finite-time transport between population-valued states, with training formulations based on semigroup, Eulerian, and Lagrangian characterizations. These maps support deterministic reward lookahead and flow map trajectory tilting (FMTT), allowing pretrained generators to be reused across observation operators. Experiments cover ShapeNet, single-cell populations, and spatial transcriptomic niches, including generation, inpainting, super-resolution, deconvolution, and guided generation. The paper reports improved reconstruction metrics and fewer function evaluations than several WFM-based baselines. The single-cell deconvolution task reconstructs populations from mean cell embeddings, rather than measured bulk gene expression.

Critical strengths

Direct endpoint prediction avoids repeated integration for terminal-objective evaluation, and the same-network comparison in Appendix I.1 provides evidence of competitive generation quality with substantially fewer evaluations than integrating the associated velocity field.

Critical weaknesses
  1. Table 2 changes both the generative prior and the inverse solver: FlowDPS and PnP-Flow use WFM/RWEFM, whereas lookahead and FMTT use the learned flow maps. Tables 5–6 and Appendix F.1 also document differences in conditioning architecture, optimization schedules, EMA, dropout, and transport coupling. Better reconstruction could therefore reflect a different training recipe or a better prior, rather than the proposed inference mechanism. The same-network ablation in Appendix I.1 supports the unconditional sampling-efficiency claim, but does not resolve this attribution problem for the inverse tasks. Controlled comparisons using the same trained prior and matched conditioning are needed to separate these effects.

  2. The single-cell pipeline includes separately trained composition inferrers and, for some inpainting configurations, a held-out-type predictor whose outputs supply per-cell conditioning tokens, as described in Appendix F.3. Although the generator is frozen, the complete pipeline therefore includes observation-specific supervised inference. The descriptions of the inferrer variants do not make their allocation across experimental arms sufficiently clear. An unguided baseline using exactly the same inferred composition and centroid tokens is needed to establish how much of the reported improvement comes from guidance rather than auxiliary inference.

  3. Appendix G.2 replaces the voxel-mean discrepancy in Eq. G.2 with a per-point penalty toward assigned voxel means and describes the objectives as equivalent. However, even with fixed, balanced assignments,

∑i∈v∣xi−yv∣2=nv∣mv(X)−yv∣2+∑i∈v∣xi−mv(X)∣2,\sum_{i\in v}\\|x_i-y_v\\|^2 = n_v\\|m_v(X)-y_v\\|^2 + \sum_{i\in v}\\|x_i-m_v(X)\\|^2,

where mv(X)m_v(X) is the mean of the nvn_v points assigned to voxel vv. The implemented penalty additionally suppresses within-voxel variation, which the stated observation does not constrain. Two points at yv+ay_v+a and yv−ay_v-a, for example, have zero mean mismatch but a positive per-point penalty. Balanced occupancy does not remove this discrepancy. The implementation therefore changes the reconstruction objective, and the voxel-deconvolution results do not directly validate the stated formulation. The conclusions need to be checked using the actual voxel-mean objective.

  1. Appendix E.4 omits the backward-flow residual from the implemented FMTT importance-weight rate, while Appendix D.4 uses an isotropic-source score formula in runs with a data-fitted anisotropic source, omitting the covariance-inverse factor. These limitations are disclosed, but their effects on target fidelity are not quantified. An incorrect reference score can change the reference marginals, rather than merely approximate the terminal reward. The exact corrector analysis therefore does not establish the behavior of the implemented sampler, and reconstruction metrics alone cannot close this gap. Restoring the covariance-correct score and full weight rate, together with a tractable known-target test, would clarify whether the implemented procedure samples the intended distribution.
Questions for authors
  1. Do the inverse-problem improvements persist when the methods use the same underlying trained prior and matched conditioning? Can you separate the effects of the training recipe, prior quality, and guidance mechanism rather than changing all three together?

  2. Which composition inferrer, masking-training regime, and held-out-type predictor is used in each single-cell experimental arm, including FT-WFM? How much improvement remains over unguided generation using exactly the same inferred composition and centroid tokens?

  3. How is the claimed equivalence between Eq. G.2 and the implemented per-point voxel penalty reconciled with the additional within-voxel variance term above? Do the reported conclusions hold when optimizing the stated voxel-mean objective?

  4. What is the effect of restoring the covariance-correct score and the full FMTT weight rate? Can target fidelity be evaluated on a tractable population inverse problem with a known tilted distribution, rather than only through reconstruction metrics?

6 comments

We thank the reviewer for the careful and constructive assessment. The questions have motivated several additional controlled experiments, and the new results have strengthened our empirical evidence and helped clarify where the benefits of our approach arise. Below, we present matched-prior and conditioning experiments, along with additional evaluations and corrections addressing the implementation concerns.

Weakness / Question 1 — Do the gains persist with matched training and a fixed prior?

A1. We agree that the original comparisons do not fully isolate the effects of prior training and inference-time guidance. We therefore retrained WFM with matched architectures and optimization settings as our flow map, and evaluated different inference strategies on each frozen prior. We additionally evaluate SetFM, which drops the OT coupling WFM uses, and instead uses the same independent coupling as our flow-map, providing a closer control for the training objective.

For lookahead, we compare three ways of estimating the terminal population: (i) a one-step estimate obtained by extrapolating the instantaneous velocity to terminal time, (ii) an 8-step numerical rollout of the same velocity field, and (iii) direct endpoint prediction using the learned flow map. The first two use the frozen prior's velocity field, while the third uses the learned finite-time map. Below we report reconstruction MMD for deconvolution, and masked MMD for inpainting and super-resolution tasks (lower the better).

PriorInference methodInpainting (planar)Super-resolution (M=128)Voxel deconv. (R=3)
Wasserstein Flow map (ours)PnPFlow0.3290.2370.037
FlowDPS0.3110.1870.028
Lookahead (1-step)0.2660.1320.028
Lookahead (8-step)0.2630.1300.028
Lookahead (learned map) (ours)0.2270.1310.024
WFMPnPFlow0.2310.1340.039
FlowDPS0.2370.1340.039
Lookahead (1-step)0.2420.1410.036
Lookahead (8-step)0.2310.1260.033
SetFMPnPFlow0.4760.2280.037
FlowDPS0.3050.1790.028
Lookahead (1-step)0.2600.1420.028
Lookahead (8-step)0.2550.1320.028

Matched controls show that our Wasserstein flow map lookahead improves inpainting and voxel deconvolution beyond training or prior-quality differences, while achieving 1.6–4.6× faster inference than WFM-based guidance (See our A4 to reviewer CDD5).


W/Q 2 — Which auxiliary predictors are used, and what does guidance add with identical conditioning?

A2.

Clarification of the composition inferrers and conditioning.

The key distinction is that FT-WFM is fine-tuned for each inverse task, whereas our pretrained flow map remains frozen and uses inference-time guidance. Our method obtains cell-type conditioning tokens from a lightweight, separately trained composition inferrer (0.43–0.46M parameters), using the observed cell-type histogram for inpainting/super-resolution or the observed mean embedding for bulk deconvolution. In contrast, FT-WFM jointly trains its generative backbone and composition inferrer on partially observed populations for inpainting/super-resolution. For bulk deconvolution, it fine-tunes the backbone conditioned directly on the observed mean embedding. The obvious drawback of such approach is that it is tailored to the specific class of inverse problems and don't generalize to arbitrary ones, unlike our framework.

Ablating the contributions of lookahead guidance and auxiliary inference.

The composition inferrer is only meant at generating higher-fidelity samples across all possible inverse problems on the single-cell RNAseq modality. Crucially, the contribution of our proposed framework for distributional inverse problem is orthogonal to the base model quality. To isolate the contribution of our guidance mechanism on inverse problem performance, we compared (1) guided and unguided generation on the same evaluation cohort and identical inferred composition/centroid tokens (Inferred, unguided vs lookahead); and (2) guided and unguided generation on a model without cell-type tokens conditioning (No cell type, unguided vs lookahead). Below we report reconstruction MMD for deconvolution, and masked MMD for inpainting and super-resolution tasks (lower the better).

TaskNo cell type, unguidedNo cell type, lookaheadInferred, unguidedInferred, lookahead
Inpainting0.9830.9010.4640.385
Super-resolution (M=256)0.4870.2020.1390.089
Bulk-to-single0.4120.1550.1400.060

Our results show that guidance consistently improves reconstruction with identical inferred tokens across both settings, directly establishing a contribution beyond auxiliary inference.

AuthorsOpenReview ↗

part 2 of the response

W/Q 3 — How do voxel-loss differences affect reconstruction?

A3. We thank the reviewer for identifying this discrepancy. We agree that the equivalence claim in Appendix G.2 was incorrect. The implemented per-point objective includes an additional within-voxel variance penalty beyond the stated voxel-mean matching loss, which was not accurately reflected in our description.

Results using the stated voxel-mean objective. To assess the effect of this discrepancy, we reran voxel deconvolution using the voxel-mean objective in Eq. G.2, keeping the frozen priors and evaluation protocol fixed. Each entry reports new value (old value), averaged over ShapeNet categories, for R=4. We observe that the corrected objective preserves the competitiveness of flow-map guidance:

MethodMMD ↓CD ↓Δwv ↓
Lookahead0.0160 (0.0133)0.0199 (0.0233)0.0061 (0.0080)
FMTT0.0157 (0.0113)0.0209 (0.0224)0.0062 (0.0102)
FlowDPS0.0180 (0.0150)0.0203 (0.0234)0.0069 (0.0091)
PnP-Flow0.1200 (0.1037)0.0520 (0.0527)0.0082 (0.0579)
FT-WFM0.0143 (0.0143)0.0231 (0.0231)0.0056 (0.0056)

Whole-cloud reconstruction metrics; FT-WFM values remain unchanged.

We note that the discrepancy only concerns the supplementary voxel-deconvolution experiments and does not affect any other inverse tasks or the main Table 2 results. The manuscript is updated reflecting these changes (App. G.2, App. H.1, Table 13).


W/Q 4 — How do the source score and simplified FMTT weights affect sampling fidelity?

A4. We thank the reviewer for highlighting these points.

Source distributions.

We mentioned data-fitted anisotropic Gaussian sources since such priors are sometimes used in WFM, but this description was imprecise in the context of our FMTT experiments. None of our reported FMTT experiments uses such a source: Euclidean experiments use isotropic Gaussian prior, while single-cell experiments use a uniform prior on the sphere. Thus, the covariance-inverse issue does not affect our reported results. We now clarify this in App. D.4.

Omitted backward-flow residual.

We agree that omitting the backward-flow residual can make simplified FMTT importance-weight rate inexact. To quantify the effect of this approximation, we conducted two additional experiments.

Known-target experiment. We constructed a hierarchical Gaussian population inverse problem with an analytically tractable posterior. Each population contains eight correlated particles observed through their population mean. We compared the simplified and full FMTT rates using an exact and a deliberately inconsistent flow map (32,768 walkers, 1,024 steps, 3 seeds).

Flow mapSimplified rate ↓Full rate ↓
Exact0.0090.009
Deliberately inconsistent0.2560.010

Maximum KS error between the sampled population-mean marginal and its analytic posterior marginal.

With an exact map, both rates closely match the target marginal. With an inconsistent map, omitting the residual introduces substantial bias, which the full rate largely removes. This directly demonstrates the importance of the correction when flow consistency fails.

Evaluation on our trained model. We also restored the full correction for spatial-niche enrichment using the trained flow map as in our experiments. Across 100 targets, 3 proposal seeds, and 8 walkers, the change in target error was +0.0001 [−0.0014, 0.0015], with consistent findings at 8 and 16 steps.

Together, these experiments demonstrate that the simplified rate is not generally exact, while showing that the omission has little measurable effect in our spatial-niche experiment. We include these additional results in App. E.4.

We believe that these controlled comparisons, additional ablations, and implementation corrections address all of the reviewer’s concerns and substantially strengthen the evidence for our main contributions. We therefore hope the reviewer will reconsider their assessment in light of these new results.

As the discussion period approaches its end, we wanted to follow up on our earlier response. We have provided substantial additional experiments and clarifications addressing all concerns raised in your review, and would greatly appreciate your feedback before the deadline. We believe the new evidence strengthens our contributions and merits reconsideration of the overall assessment. We would be happy to address any remaining questions or concerns. Thank you for your time and consideration!

Comments on the paper

2 comments

We are grateful to the reviewers for their comments and constructive feedback. We address reviewer-specific comments in our individual responses. We summarize the main revisions and new results addressing concerns shared across the reviews below.

  • Contributions and relationship to prior work - We propose a framework for distribution-valued inverse problems by learning a one-step measure-level generator combined with inference time-guidance. Our objectives lift Riemannian Mean Flows (RMF) to the space of distributions, realizing a Wasserstein manifold variant thereof. (App. D.2, I.2 in revision)
  • Separating guidance from prior quality and conditioning - Our additional experiments with matched training and fixed-prior controls demonstrate that flow-map lookahead improves inverse problem solving (App. I.3 in revision), and establish the benefits of lookahead guidance beyond cell-type conditioning (App. I.5 in revision).
  • Connecting theory and implementation - We add flow-consistency diagnostics, cycle-regularization ablations, and known-target FMTT evaluations to quantify practical approximations (App. D.5, E.4 in revision). We also update the voxel-loss description and corresponding results (App. G.2, H.1 in revision).
  • Practical efficiency - Wall-time end-to-end benchmarks show that flow-map lookahead achieves 1.6–4.6× faster inference than WFM-based guidance across the three evaluated domains. We also report peak memory and all FMTT walker costs, making the computational trade-offs explicit (App. F.5 in revision).
AuthorsOpenReview ↗

Final Summary to the Area Chair and Reviewers

We sincerely thank the reviewers for their constructive feedback. We are encouraged that all three reviewers have recognized meaningful contributions from our additional experiments and clarifications during the rebuttal. We thank the reviewer JR2s for acknowledging our theoretical contribution and the improved discussion of RMF, and reviewer vj6R and CDD5 for recognizing the additional empirical evidence and positive reassessment of our work. We would like to highlight three key takeaways below.


1. Central contribution and theoretical novelty

Our work is motivated by scientific inverse problems where the unknown is an entire population, such as reconstructing single-cell populations from bulk measurements, spatial tissue niches from partial observations, or incomplete 3D point clouds. Each state is a probability measure μ∈P2(M)\mu\in\mathcal P_2(\mathcal M), and the generative model learns a law over populations P∈P2(P2(M))P\in\mathcal P_2(\mathcal P_2(\mathcal M)).

Our central theoretical contribution is to formulate finite-time flow maps directly from measure-valued dynamics, rather than starting from a finite-particle RMF construction. We establish the population flow-map operator, derive its flow-map identities and empirical-particle realization (Secs. 3.3–3.4; App. C.2), and establish well-posedness and Wasserstein stability of the ideal dynamics (App. C.1), together with the empirical Wasserstein quotient connection (App. B.1).

As discussed extensively with JR2s, we acknowledge that our finite-particle objectives overlap with RMF. However, equivalence of finite-particle training objectives does not imply equivalence of mathematical contributions. This correspondence arises naturally from specializing our measure-level formulation to empirical populations, rather than being the starting point of our derivation. Our novelty lies in the measure-level formulation, its theoretical connection to empirical population transport, and its application to distributional inverse problems—not in introducing new finite-particle losses.


2. Practical contribution: efficient and reusable guidance for distributional inverse problems

Learned flow maps directly predict terminal populations without repeated numerical integration, making them particularly useful for inverse problems, where guidance repeatedly evaluates and differentiates observation-dependent rewards on predicted endpoints.

Across point-cloud reconstruction, single-cell deconvolution and inpainting, and spatial-niche reconstruction and enrichment, we demonstrate that one frozen generative prior per domain can be reused across different inverse tasks, without task-specific backbone fine-tuning. This reusability also comes with competitive reconstruction quality, outperforming task-specifically fine-tuned WFM on several single-cell tasks.

Concretely, learned-map lookahead achieves 1.6–4.6× faster end-to-end inference than the evaluated WFM-based guidance baselines across three tasks. In a controlled comparison against 8-step WFM endpoint rollout, it achieves an 8.9× speedup while maintaining competitive reconstruction quality.


3. Substantial additional evidence addressing the reviews

We conducted matched-prior, cross-solver, RMF, and identical-conditioning ablations, demonstrating that our design choices and inference-time guidance provide meaningful improvements beyond differences in prior quality or auxiliary conditioning alone.

We also evaluated FMTT approximation at distributional level. A controlled Gaussian experiment demonstrates the importance of the full correction for a deliberately inconsistent flow map, while our trained spatial-niche model shows close agreement between simplified and fully corrected output distributions. We emphasize that FMTT is a complementary guidance strategy, rather than our central contribution.

Additionally, we quantified flow-consistency errors, corrected and reran the voxel-deconvolution experiments, and provided end-to-end runtime and memory measurements, strengthening the empirical validation while clarifying the remaining computational and accuracy trade-offs.


We believe the substantial additional experiments, corrections, and clarifications directly address all major concerns raised during review and significantly strengthen the paper. Given the further evidence provided throughout the discussion, we hope the reviewers and Area Chair will consider both the measure-level theoretical contribution and the demonstrated practical value of efficient, reusable guidance in the final decision.

We sincerely thank again all reviewers and the Area Chair for their time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.