acceptodds
Withdrawn from ICLR 2027

Wasserstein-Fisher-Rao Flow Matching: Resolving Jacobian Stiffness via Dual-Action Probability Transport

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Deterministic continuous generative models, including Flow Matching and Continuous Normalizing Flows, restrict probability density transport to mass-conserving kinematic trajectories. When bridging disconnected topological supports, this strict conservation physicalizes as local velocity divergence singularities (), which induces severe spatial Jacobian stiffness and collapses explicit ODE solver step sizes. To resolve this geometric bottleneck, we introduce Wasserstein-Fisher-Rao Flow Matching (WFR-FM), a dual-action continuous generative framework parameterized by an unbalanced continuity equation. WFR-FM decouples spatial advection from topological probability weighting via an independent scalar reaction field, allowing trajectories to traverse zero-density corridors by dynamically suppressing and reconstituting mass rather than forcing extreme spatial curvature. Theoretically, we prove that the simulation-free WFR-FM regression strictly upper-bounds the macroscopic Unbalanced Hellinger Divergence. Furthermore, the decoupled objective satisfies Polyak-Łojasiewicz (PŁ) optimization bounds, ensuring exponential convergence and finite gradient variance up to high dimensions (), even under discrete minibatch sampling approximations. Empirically, across disconnected 2D manifolds, WFR-FM strictly bounds continuous integration stiffness, reducing inference Number of Function Evaluations (NFE) by up to compared to state-of-the-art mass-conserving baselines. On high-dimensional empirical benchmarks exhibiting intrinsic non-equilibrium dynamics, the decoupled scalar field physically isolates local density variations, achieving a Relative Mass Error (RME) of 0.0016 without adjoint simulations.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Flow Matching, Continuous Normalizing Flows, Unbalanced Optimal Transport, Wasserstein-Fisher-Rao, ODE Jacobian Stiffness, Non-Equilibrium Dynamics

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_YGAROpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The authors argue that deterministic continuous generative models incur an increased cumulative number of function evaluations (NFE) when dealing with target distributions that exhibit clear separation. To resolve this problem, they propose reducing NFE by utilizing local density source-sink variations (gϕg_\phi). The experiments evaluate NFE alongside distributional discrepancy and support leakage, linking the proposed reduction in integration difficulty to both computational cost and generation quality. The reported NFE reductions provide evidence for this direction.

Critical strengths

This paper investigates mass modulation to reduce ODE sampling costs. Evaluating NFE alongside distributional error and support leakage connects efficiency to generation quality, with reported reductions suggesting practical potential for faster sampling.

Critical weaknesses

The contribution relative to prior WFR-FM is not established. WFR-FM: Simulation-Free Dynamic Unbalanced Optimal Transport (Peng et al., ICLR 2026) already introduces joint velocity–growth regression, mass-weighted objectives, and simulation-free training. The manuscript neither cites nor directly compares against this closely related work. Although its focus on inference stiffness is potentially valuable, the paper does not establish which technical advances or empirical gains extend beyond the existing framework.

Questions for authors

1.How does the proposed method differ technically from WFR-FM (Peng et al., ICLR 2026), which already combines velocity–growth regression, mass-weighted objectives, and simulation-free training? Which differences constitute the new contribution, and what evidence attributes the reported NFE improvements to these differences rather than to the shared framework?

2.Table 1 shows that WFR-FM’s Void leakage is comparable to or higher than several baselines. How does this support the claim of preserving geometric support, and does the NFE advantage remain at comparable generation quality?

Reviewer_KWv8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors introduce WFR-FM, an unbalanced continuous generative framework to overcome the topological limitations of mass-conserving models such as Flow Matching and CNFs. On 8 Gaussians (R = 2, 5) and a checkerboard, WFR-FM is reported to need 20–85 NFE versus 670–32,581 for mass-conserving baselines, with comparable transport quality. A D = 64–512 stability test and an scRNA-seq benchmark follow.

Critical strengths

Fixed solver and tolerances across methods, and explicitly excluding fixed-step solvers , is a reasonable intent. Several baselines are included, with 5 seeds, and a limitations section is present.

Critical weaknesses
  1. The code in Appendix C suggests the reaction field cannot affect the reported results. In forward, d_x = v: spatial trajectories depend only on the velocity network, never on g. sample() returns z1[..., :dim], discarding the log-weights. All sample-based metrics (EMD, void leakage, kinetic action, NFE of the x-trajectories) are therefore computed on unweighted samples from a plain velocity flow. For both toy datasets, getitem sets mass_ratio = 1.0, so m_t = exp(t·10⁻¹⁰) ≈ 1. The velocity loss is then ordinary CFM MSE. If no log_density_fn is set, g_t = 0 and the scalar field is trained toward zero (the same “dormancy” the paper blames on VGFM). Net effect: on the toy benchmarks, WFR-FM’s velocity learning appears essentially identical to OT-CFM. Consistently, Table 1 shows WFR-FM ≈ OT-CFM/Struct-Coupling in kinetic action (0.2357 vs 0.2384) and EMD (0.3498 vs 0.3431). Yet the NFE differs by roughly 500×.

  2. Confound and likely baseline artifact. If the same v-regression gives 44 NFE in one model and 25,018 in another, something other than “decoupling” explains it (undertraining, different augmented state, e.g., log-det channels with tight tolerance for FFJORD-style baselines, different time embedding, a different solver call). Tens of thousands of NFE for dopri5 on a 2D, 8-mode problem is far above what is normally seen (typically tens to hundreds). VGFM, which the authors say has a dormant growth field, is also at 26–45 NFE with similar quality, which suggests low NFE is the normal behavior and the baselines are the anomaly.

  3. No stiffness is ever measured. There are no Jacobian eigenvalues, divergence along trajectories, step-size traces, or rejected-step counts. The central causal claim rests on NFE alone.

  4. Missing the decisive ablation. WFR-FM with g ≡ 0 (and m_t = 1), i.e., plain CFM with the same backbone, same training budget, same solver, is absent. Also missing: NFE at matched quality, fixed-step Euler/midpoint results (how FM is actually used), and a topology-matched prior such as a K-component Gaussian mixture prior, which sidesteps the obstruction trivially.

  5. Headline numbers are selectively presented and partly inconsistent. “236×” equals 10,381/44, the ratio to the cheapest R = 2 baseline. Against 2-Rectified Flow, it would be 740× (32,581/44). The intro’s “34×–236×” mixes datasets (670/20 = 34× is checkerboard). The “97× lower kinetic action” compares against FFJORD/OT-Flow (0.078), which aren’t among the four named baselines. OT-CFM (0.0009) and Div-Aligned (0.0000) are as low as or lower than WFR-FM (0.0008). On R = 2, WFR-FM’s Sinkhorn EMD (0.3498) is about 1.9× worse than FFJORD/OT-Flow/2-RF (0.19–0.20). On the checkerboard, it has the same void leakage as everyone (≈ 0.63), so all methods apparently fail there (leakage 0.60–0.78), contradicting “robust topological support alignment”. At R = 5, kinetic action is worse than OT-CFM/Struct/VGFM/Div-Aligned, and EMD differences are within one SD (n = 5). NFE SDs of exactly 0.0 (WFR-FM 44.0 ± 0.0, 20.0 ± 0.0; VGFM 26.0 ± 0.0) across five independently trained networks with an adaptive solver are surprising and need explanation.

  6. Table 1 vs. Table 4 contradict each other. Table 4 marks hyphens as “catastrophic solver failure (timeout/NaN)”, but Table 1 reports finite means for the same cells: OT-CFM R = 5 (27,366), Struct-Coupling R = 2 (20,932), Div-Aligned R = 5 (20,888), and finite leakage for many “failed” cells. Table 4 also says “>30,000” for 2-RF, whereas Table 1 says 30,802.

  7. Metrics are undefined. “Void leakage”, “kinetic action” (does it include the g term?), the Sinkhorn EMD configuration, and the checkerboard “void” are never defined in the main text. For a perfect model, leakage should be ≈ 0, so values near 0.6 for all methods are red flags about the metric or the models.

  8. D = 512 experiment is under-specified. The data distribution isn’t given, and the “KL-Ablation” is not defined. NaN at D ≥ 64 from log-ratios is surprising in float32 (log-densities are only hundreds), and WFR-FM itself exponentiates the ratio. Gradient variance is not a PŁ test. The text says “no gradient clipping” (Table 3, Fig. 3), but train_single calls clip_grad_norm_(…, train_config.grad_clip).

  9. Baseline fidelity and fairness. “Divergence-Aligned FM (Fan et al., 2025)” cites a NeurIPS paper titled Adaptive divergence regularized policy optimization for fine-tuning generative models, an RL fine-tuning method, which does not obviously match the described “penalize velocity divergence” baseline (to verify). 2-Rectified Flow is a two-stage method, which conflicts with the paper’s “single-stage only” scope. WFR-FM uses two networks (about 2× parameters, and both are evaluated at every NFE), versus one for baselines. Text says 4 layers; code defaults to 2 residual blocks. No per-method hyperparameter tuning; training is only 100 epochs on 10k points. “Optimal WFR coupling / Minibatch Unbalanced OT”: the code uses balanced entropic Sinkhorn with argmax rounding, off by default (coupling_mode="independent").

  10. Section 5.1 sensitivity claims don’t match data. “Uniform box prior reduces leakage to 0.31 on R = 5” is higher than the 0.0250 in Table 1. The γ/τ grid shows only “reaction loss ~10⁻³” (consistent with a field trained toward zero), and τ is never defined. NFE for the grid is not shown.

Questions for authors
  1. Please report NFE and metrics for WFR-FM with g ≡ 0, and for OT-CFM trained under your exact pipeline and solver call.
  2. What do ‖g‖ and the log-weights look like over time, and are weights used in any metric?
  3. Why are baseline NFEs so high, and what does Euler with 50–100 steps give each method?
Reviewer_7kw7OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces Wasserstein-Fisher-Rao Flow Matching (WFR-FM), a method designed to alleviate Jacobian stiffness and reduce the computational cost of ODE-based inference when transporting distributions with disconnected supports. Building on the unbalanced continuity equation, WFR-FM simultaneously models spatial transport through a velocity field and mass changes through a scalar growth field. The method defines conditional interpolation paths for both spatial displacement and mass evolution, allowing simulation-free training using a mass-weighted regression loss. On the theoretical side, the authors claim to derive an upper bound for the endpoint Hellinger error and provide convergence guarantees under the Polyak-Łojasiewicz (PŁ) condition. Experimental evaluations on synthetic data, high-dimensional benchmarks, and single-cell transcriptomics datasets are used to demonstrate gains in inference efficiency and numerical stability, together with the capability to capture non-conservative mass dynamics.

Critical strengths
  • The proposed framework jointly models spatial dynamics and population mass variations, making it suitable for studying systems involving proliferation and death, such as single-cell populations.

  • According to the experimental results reported in the paper, the proposed method can significantly reduce the number of function evaluations (NFE) during inference.

Critical weaknesses
  • Incorrect WFR geodesic: The single-particle WFR geodesic proposed in this paper is incorrect. The derivation of the single-particle WFR geodesic has already been presented in [1], and the resulting expression differs from the one used in this paper.

  • Inconsistency between the implementation and the proposed method: The code provided in the appendix is inconsistent with the method described in the main text. For example, on page 19, line 984, the authors use balanced OT rather than unbalanced OT. Moreover, the datasets defined on page 20 use mass_ratio = 1.0.

  • Inconsistent experimental results: Several experimental results reported in the paper are inconsistent. For example, Table 1 in the main text reports NFE and Void Leakage results for the 8-Gaussians and Checkerboard datasets, whereas the corresponding entries in Table 4 of the appendix are marked with dashes ("-").

[1] WFR-FM: Simulation-Free Dynamic Unbalanced Optimal Transport

Questions for authors
  • Regarding Weakness 1: Could the authors clarify the relationship between the single-particle WFR geodesic proposed in this paper and the result derived in [1]?

  • Regarding Weakness 2: To avoid potential misunderstandings, I encourage the authors to provide a carefully corrected implementation that is consistent with the method described in the paper.

  • Regarding Weakness 3: The authors should carefully verify the reported experimental results and provide reliable and internally consistent results.

[1] WFR-FM: Simulation-Free Dynamic Unbalanced Optimal Transport

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

Related papers

Discussion (0)

Sign in to comment.