acceptodds
Under review as a conference paper at ICLR 2027

Probabilistic Geodesic Flow Matching on Location-Scale Families

Abstract

Flow matching (FM) has recently emerged as a promising framework for generative modeling due to its conceptual simplicity and strong empirical performance. In FM, samples are transported along a vector field parameterized by a neural network, inducing a probability path that evolves from a simple noise distribution to the target data distribution, governed by an ordinary differential equations (ODE). However, existing FM approaches predominantly rely on probability paths derived from optimal transport (OT) between Gaussian distributions, which may be suboptimal for capturing complex data with inhomogeneous structures such as heavy tail or sharp contrast. In this work, we generalize FM to the broader class of location–scale families for handling data inhomogeneity and introduce a novel class of probability paths defined as geodesics on the manifold of probability distributions. We name this approach probabilistic geodesic flow matching to distinguish it from prior geodesic (Riemannian) FM methods defined in input space. We argue that Euclidean OT-based paths are not necessarily optimal in probability space and may limit modeling flexibility. Through synthetic benchmarks and scientific datasets at different scales, we demonstrate that the proposed method more effectively captures complex distributions, leading to improved or comparable generative performance compared with SOTA geometry-motivated models.

Primary area:
generative models
Keywords:
exponential power distributions, inhomogeneous data, manifold of probability paths, Poincare semi-circle

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_TSHEOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper extends flow matching methods from Gaussian distributions to exponential power distributions. They derive the conditional flows using Fisher-Rao geodesics rather than the distance defined in the input space. The method is evaluated on synthetic datasets, MNIST, and QM9.

Critical strengths

The idea of using the Fisher-Rao geodesic is actually interesting, but I'm not fully convinced why evaluate the method on these Euclidean datasets.

Critical weaknesses

Critical Weaknesses

  1. Theoretical link between conditional Fisher Rao optimality and marginal generation: Theorem 4.1 provides a very clean result in that the proposed schedule is in fact the Fisher Rao geodesic between the conditional endpoints. In practice however the PG path acts as an affine interpolant with a schedule that is sample dependent in that it is determined by ∥x1∥\|x_1\|. It's not at all clear from a theoretical point of view why Fisher Rao optimality at the conditional level should necessarily lead to a more learnable marginal vector field or better sampling.

  2. The empirical performance of the pure PG path: The geometric formulation of PG is by far the most significant conceptual contribution, but the pure PG path has yet to show a consistent empirical advantage over the standard OT baseline across the benchmarks reported. Specifically:

    • In Table 2 the isotropic PG variant is worse than OT on both ED and MMD across both qq settings.
    • In Tables 3 and B.4 pure PG performs worse than OT on all three evaluation metrics for both qq values.
    • In Table 4 pure PG gives higher FID and classifier FID scores than OT for both qq settings.
  3. Source of empirical gains: The majority of the empirical gains are attributable to the anisotropic version or to the hybrid (HB) approach which switches over to the standard OT path at tc=0.85t_c = 0.85. A number of the improvements are actually within one std of the baseline, so it's not easy to say that the gains are due to the Fisher Rao geodesic construction per se.

  4. Alignment between the text claims and the tabulated results: There are several instances where the summary statements in the main text don't match the numbers in the tables. Line 461 states that HB achieves the lowest W2W_2 and ED at q=1q=1, yet in Table 2 HB's W2W_2 is actually higher than a number of the baselines and the anisotropic PG variant has a lower ED. The generalisation that q=1q=1 almost universally outperforms q=2q=2 is also contradicted by multiple examples in Tables 2, 3, B.3 and B.4.

  5. Baseline consistency on Euclidean data: On Euclidean data with a flat metric RFM reduces exactly to standard OT flow matching. On Table 2 (at q=1q=1) there is a significant difference between the two methods (W2W_2 of 5.49 vs 5.66, with reported std devs of around 0.01). The difference between these two baselines which should be identical is on a similar order to the improvements seen from the proposed variants so that does raise some questions about variance and baseline calibration.

Questions for authors
  1. What result can show that minimizing Fisher-Rao energy can improve generation, rather than simply defining a different interpolation in another space?
  2. How are the two branches of HB joined at tct_c? Is there some matching step in the implementation, and how is the training target kept consistent when sampling with an ODE?
  3. How does the experimental part compare with RFM? What manifold / metrics and conditional path are used on these Euclidean datasets?
Reviewer_3HCoOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper extends conditional flow matching from Gaussian paths to location-scale families, focusing on the exponential power (EP) distribution Pq(μ,σ2I)\mathcal{P}_q(\boldsymbol{\mu}, \sigma^2\mathbf{I}), which reduces to the Gaussian at q=2q=2. It derives the Fisher-Rao metric on the isotropic EP manifold, ds2=(cμ∥dμ∥2+cσdσ2)/σ2ds^2 = (c_\mu\|d\boldsymbol{\mu}\|^2 + c_\sigma d\sigma^2)/\sigma^2, and uses the resulting closed-form geodesic between Pq(0,I)\mathcal{P}_q(\mathbf{0},\mathbf{I}) and Pq(x1,σmin⁡2I)\mathcal{P}_q(\mathbf{x}_1,\sigma_{\min}^2\mathbf{I}) as the conditional path. It also proposes an anisotropic variant and a hybrid (HB) path that switches to the OT interpolant at tc=0.85t_c = 0.85 to mitigate stiffness. Experiments range from 2D benchmarks to MNIST, scientific datasets, and QM9 molecule generation.

Critical strengths
  1. The Fisher-Rao derivation for the isotropic EP family appears correct and recovers the known Gaussian result (cμ=1c_\mu = 1, cσ=2dc_\sigma = 2d) at q=2q=2 (Remark 4). The Poincaré semicircle parametrization θ(t)=2arctan⁡(eLttan⁡(θ0/2))\theta(t) = 2\arctan\big(e^{Lt}\tan(\theta_0/2)\big) gives fully explicit constants, avoiding the implicit solve required by the sech⁡/tanh⁡\operatorname{sech}/\tanh form. Its O(d)\mathcal{O}(d) cost becomes small relative to training in higher dimensions, adding about 10% per epoch in the 100D funnel, although it roughly doubles the per-epoch time in 2D.

  2. Treating the source distribution as a tunable member of a location-scale family, with qq controlling concentration and tails, is a natural generalization of Gaussian flow matching, and the sampler via Proposition 2.1 is simple. Because μc\mu_c, RR, θ0\theta_0, and θ1\theta_1 depend on ∥x1∥\|\mathbf{x}_1\|, the PG path amounts to a data-dependent noise schedule, which is genuinely distinct from the data-independent schedulers of standard affine-path flow matching.

  3. The evaluation is broad, spanning synthetic, image, scientific, and molecular data. Most tables report mean ±\pm standard deviation over 10 seeds, and several report time per epoch. The qq-sweep (Table B.3) and the explicitly stated stiffness limitation are welcome.

Critical weaknesses
  1. Several claims overstate the results. Sec. 5.1.2 says q=1q=1 “almost unanimously” beats q=2q=2 in Table 1, but it wins only 11 of 18 cells, and isotropic PG trails OT on ED and MMD at both values of qq (e.g., ED 1.82×10−11.82\times10^{-1} vs. 1.43×10−11.43\times10^{-1} at q=1q=1). The same section credits PG-1 with the lowest ED and MMD in Table B.4, but these belong to PG-aniso; isotropic PG at q=1q=1 has the second-worst ED. On checkerboard, OT at q=2q=2 beats every PG configuration on all three metrics (Table B.2), and in the qq-sweep, q=2q=2 gives the best ED on all three 2D datasets (Table B.3), which weakens the heavy-tail motivation. Table B.6 is called “comparable,” yet at q=2q=2, HB’s MMD on alanine dipeptide is about 18× worse than OT’s (3.06×10−33.06\times10^{-3} vs. 1.71×10−41.71\times10^{-4}). On MNIST, pure PG is worse than OT on every metric (Table 2), and on QM9, PG’s validity (90.63%) falls below all four external baselines (94.4–98.9%), a gap described as “slightly lower” (Sec. 5.3). In the higher-dimensional benchmarks, the best results come from HB or PG-aniso rather than the isotropic method the paper is named after.

  2. The hybrid path is discontinuous. As defined in Sec. 4, the HB conditional flow switches at $t_c $ from μtc(x1)+σtc(x1)x\mu_{t_c}(x_1) + \sigma_{t_c}(x_1)x to tcx1+(1−tc)xt_c x_1 + (1-t_c) x, and the two do not generally coincide. For example, with q=2q=2, d=2d=2, ∥x1∥=1\|x_1\|=1, and σmin⁡=10−3\sigma_{\min} = 10^{-3}, Eq. (19) gives σ0.85≈0.003\sigma_{0.85} \approx 0.003, while the OT branch has scale 0.150.15. The probability path therefore jumps at tct_c, and no vector field can generate it through the continuity equation. Since HB yields several of the paper’s headline results (Tables 1, 2, and B.7), this needs a continuity-preserving construction or a justification.

  3. Path geometry is confounded with time parametrization. The constant-hyperbolic-speed parametrization front-loads the motion. In the same example, at t=0.2t=0.2 the PG mean has covered about 93% of the distance with σt≈0.30\sigma_t \approx 0.30, whereas OT has covered 20% with σt=0.80\sigma_t = 0.80. This likely explains the earlier emergence of structure in Figs. 2, 3, and B.1, which therefore cannot distinguish a better path from a faster schedule; final-sample quality at matched compute is what matters. Separately, the claim after Eq. (21) that PG recovers the sinusoidal path holds only for the curve’s trace. It requires λ=cσ/cμ=∥x1∥\lambda = \sqrt{c_\sigma /c_\mu} = \|x_1\|, although λ\lambda is fixed by (d,q)(d,q), and as σmin⁡→0\sigma_{\min}\to0 we get L→−∞L\to-\infty, so θ(t)→0\theta(t)\to0 for all t>0t>0 rather than θ(t)=π(1−t)/2\theta(t) = \pi(1-t)/2.

  4. The theory is incremental, and the motivation is heuristic. Theorem 3.1 extends Lipman et al.’s Theorem 3, as Remark 2 acknowledges, and the decomposition LCFM=LFM+CL_{\mathrm{CFM}} = L_{\mathrm{FM}} + C in Theorem A.1 is standard. The proof of Proposition 3.1 assumes unit variance per coordinate, omitting the factor sq,ds_{q,d} for q≠2q\neq2; this changes the reported emin⁡e_{\min}, though not the minimizer. Eq. (18) applies the parameter-space metric gg to the sample-space vector field vt(x)v_t(\mathbf{x}), so ∥vt(x)∥g\|v_t(x)\|_g is not well defined as written. Beyond the heuristic in Sec. 4, the paper does not link Fisher-Rao energy to sample quality, and training still uses an L2L_2 loss (Remark 8). Classical work on the Fisher-Rao geometry of Gaussians, and on noise-schedule design (VDM, EDM, timestep sampling), is not cited.

Questions for authors
  1. Is the hybrid path continuous at tct_c?

  2. Is the gain from the path geometry or from the time schedule? Can you report final sample quality versus the number of function evaluations, using both adaptive and fixed-step solvers?

  3. Is the gain from the EP source or from the PG path? Can you run a factorial ablation at matched settings: OT with P1\mathcal{P}_1 noise, PG with N(0,I)\mathcal{N}(\mathbf{0},\mathbf{I}) noise, and PG with P1\mathcal{P}_1 noise, including isotropic vs. anisotropic PG?

1 comment
AuthorsOpenReview ↗

Correction of two columns misplaced in Table 5

In response to your comment on "QM9, PG’s validity (90.63%) falls below all four external baselines (94.4–98.9%), a gap described as “slightly lower” (Sec. 5.3). ", we realized we misplaced results of OT/PG/HB in two columns Mols. Val. (\\%) and Mols. S. (\\%) in the submission and have fixed them by swapping them back (marked in red in the revision) -- note that the percentage of valid molecules (Mols Val) should be greater than or equal to that of stable molecules (Mols. S), i.e. Mols. Val. (\\%) ≥\geq Mols. S. (\\%).

Reviewer_6yinOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates the use of the exponential power distribution, a location-scale family that includes the Gaussian (q = 2), as the latent (source) distribution for flow matching models.

The choice of the interpolant is also investigated: the authors propose to use the one that minimizes the energy associated with the Fisher–Rao metric of this family. This yields an explicit solution, a semicircle in the Poincaré half-plane (I did not check the proof in detail). Because this path becomes numerically difficult near the data, the authors also propose a hybrid version that follows the geodesic and then switches to the usual linear path for t ≥ 0.85. The method is evaluated on 2D toy examples, higher-dimensional synthetic distributions, MNIST, several scientific datasets and QM9 molecule generation.

Critical strengths

-The role of the latent law in the performance of flow matching is an important question in generative modeling.

Critical weaknesses

My main concern is that this paper claim to extend the FM framework while from their introduction in 2023, FM methods have never been restricted to gaussian distribution. Overall (if we forget this overclaim / clumsy writing) the contribution is incremental : Authors presents the introduction of location-scale prior distribution as a extension of the flow matching framework (page 2) while in the seminal paper (lipman 2023) this issue is already solved. Regarding to this reading : "it generalizes FM" in page 2 is a bit problematic...

-One of the paper's motivation is to better modelize distribution's tails ... while nothing in the numerical experiment nothing allow to asses tail's behavorial specifically

I currently recommend rejection for this paper (1) but with a better positioning I could increase my score to weak reject

Questions for authors

In this work you introduce an hyperparameter, the distribution q but how should q be selected in practice?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.