- 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
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).
| Task | Method | Steps (walkers) | Time (ms) ↓ | Peak memory (GB) ↓ |
|---|---|---|---|---|
| ShapeNet | Lookahead | 32 | 79.5 | 1.12 |
| FMTT | 32 (4) | 276.4 | 4.09 | |
| FlowDPS | 100 | 134.8 | 0.24 | |
| PnP-Flow | 100 | 133.6 | 0.24 | |
| FT-WFM | 100 | 132.9 | 0.24 | |
| Single cell | Lookahead | 16 | 32.4 | 0.42 |
| FMTT | 16 (4) | 59.6 | 1.34 | |
| FlowDPS | 100 | 109.8 | 0.16 | |
| PnP-Flow | 100 | 92.0 | 0.16 | |
| FT-WFM | 100 | 91.1 | 0.16 | |
| Spatial niche | Lookahead | 8 | 7.4 | 0.07 |
| FMTT | 8 (4) | 9.9 | 0.11 | |
| FlowDPS | 100 | 34.2 | 0.10 | |
| PnP-Flow | 100 | 33.6 | 0.10 | |
| FT-WFM | 100 | 32.7 | 0.10 |
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.
| Prior | Lookahead endpoint | Inference time (ms) |
|---|---|---|
| Wasserstein Flow Map (Ours) | learned map | 79.5 ± 0.9 |
| WFM | 1-step | 59.9 ± 0.3 |
| WFM | 8-step rollout | 709.2 ± 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.
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.
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).
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.
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).
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].