Riemannian Flow Matching (RFM)

Updated 12 February 2026

Riemannian Flow Matching is a generative modeling framework that extends flow matching to arbitrary Riemannian manifolds using closed-form geodesics and manifold-native interpolation.
It leverages neural ODEs and tangent field regression to transform base distributions into target distributions without relying on stochastic simulations.
The framework has been successfully applied to uncertainty quantification, robot control, material design, and molecular docking, with theoretical guarantees on convergence and stability.

Riemannian Flow Matching (RFM) is a simulation-free generative modeling and transport framework that generalizes flow matching from Euclidean spaces to arbitrary Riemannian manifolds. It enables neural ODE-based normalizing flows, density estimation, policy synthesis, and generative modeling for data whose intrinsic geometry is non-Euclidean—such as points on spheres, tori, Lie groups, or more general curved spaces. RFM sidesteps the need for stochastic differential equations or explicit simulation, leveraging closed-form geodesics and tangent field regression to learn manifold-respecting dynamics. The approach is underpinned by differential geometric formalism, replacing standard vector calculus and interpolation with their manifold-native analogues, and has been successfully applied in uncertainty quantification, material and molecular design, robot control, and generative modeling for geometric or topological data.

1. Mathematical Foundations and Core Formulation

The central object in Riemannian Flow Matching is a time-dependent vector field $v_t(x)$ living in the tangent bundle of a Riemannian manifold $(\mathcal{M},g)$ , where $g$ is the metric tensor. For distributions $p_0$ (base) and $p_1$ (target) on $\mathcal{M}$ , RFM seeks a flow $\psi_t$ solving the ODE

$\dot x_t = v_t(x_t), \quad x_0 \sim p_0,$

such that the pushforward of $p_0$ at $t=1$ approximates $(\mathcal{M},g)$ 0. The evolution of the time-marginal densities $(\mathcal{M},g)$ 1 is governed by the continuity (Liouville) equation in manifold coordinates: $(\mathcal{M},g)$ 2

The training objective—generalizing the Conditional Flow Matching loss from Euclidean space—regresses the learnable vector field $(\mathcal{M},g)$ 3 towards a closed-form “ground truth” velocity field $(\mathcal{M},g)$ 4 that generates a prescribed family of paths (typically geodesics) connecting $(\mathcal{M},g)$ 5 and $(\mathcal{M},g)$ 6: $(\mathcal{M},g)$ 7 where $(\mathcal{M},g)$ 8 follows the geodesic interpolation and $(\mathcal{M},g)$ 9 is derived from the manifold’s geodesic velocity.

The Riemannian metric $g$ 0 admits a local inner product $g$ 1 on each tangent space, and all norm computations, divergences, and gradients are performed in this geometry. Geodesic interpolation exploits the exponential and logarithmic map structure, with closed forms available on spaces such as spheres, tori, and Lie groups (Chen et al., 2023, Braun et al., 2024, Davis et al., 24 Oct 2025).

2. Geometric Structure: Manifold Interpolation and Tangent Fields

Key ingredients distinguishing RFM from Euclidean flow matching are the use of geodesic interpolants and tangent field projections:

Geodesic interpolation: Given $g$ 2, the minimal geodesic is traced by $g$ 3.
Target field: The conditional velocity is $g$ 4, explicitly computed from the manifold’s geodesic structure. On the sphere $g$ 5, this reduces to Spherical Linear Interpolation (SLERP) and a velocity field that is always tangent to the sphere (Ju et al., 29 Jan 2026).
Tangent projection: In embedded manifolds, outputs of neural vector fields are projected to the tangent bundle using an explicit projection operator, e.g., for the sphere $g$ 6.

The Riemannian divergence operator $g$ 7 is essential for log-density estimation and uncertainty quantification (Ju et al., 29 Jan 2026). On embedded submanifolds, divergence and gradients are adapted using chart or projection formalism.

RFM further generalizes to non-standard geometries using spectral premetrics (e.g., Laplace–Beltrami eigenbasis distances) in settings with complex topology or discretized meshes to define interpolation and velocity fields in closed form (Chen et al., 2023).

3. Neural Parameterization and Algorithmic Implementation

Neural vector fields $g$ 8 are parameterized by deep architectures appropriate to the problem domain:

Standard settings use MLPs or residual blocks, with time and potentially modality or context encoded via sinusoidal embeddings and adaptive normalization (Ju et al., 29 Jan 2026).
For structured data (e.g., molecules, crystal graphs), equivariant graph neural networks respecting the permutation and symmetry group actions of the data space are used (Miller et al., 2024, Sriram et al., 2024).
For visual and sensorimotor policy learning, architectures combine CNN backbones for perception with feedforward or U-Net heads outputting tangent vectors, leveraging FiLM or AdaLN for context modulation (Ding et al., 2024, Braun et al., 2024).
Contexts may include vision (image patches), past states or actions (for temporal policies), and domain-specific discrete variables.

Training is simulation-free: each iteration samples $g$ 9, selects a random $p_0$ 0, computes the geodesic interpolant $p_0$ 1 and its target velocity $p_0$ 2 (analytically or numerically), feeds $p_0$ 3 to $p_0$ 4, evaluates the squared Riemannian norm loss, and performs backpropagation (Chen et al., 2023, Ju et al., 29 Jan 2026, Sriram et al., 2024).

At inference, ODE integrators (projected Euler or Runge–Kutta) evolve samples from $p_0$ 5 to $p_0$ 6 on the manifold. The ODE solution is periodically re-projected to the manifold as needed.

4. Theoretical Guarantees and Convergence

Recent analyses establish non-asymptotic convergence rates for RFM samplers under both learning and discretization error (Guan et al., 5 Feb 2026):

For a learned vector field $p_0$ 7 with uniform or mean-square tangent and Jacobian approximation guarantees, and Euler time step $p_0$ 8, the output marginal $p_0$ 9 at time $p_1$ 0 satisfies

$p_1$ 1

cleanly decomposing total variation error into ODE numerical and learning errors.

Constants $p_1$ 2 and $p_1$ 3 depend polynomially on manifold dimension, curvature, and regularity of $p_1$ 4, $p_1$ 5, and the score $p_1$ 6.
Explicit polynomial iteration complexity bounds are available for the hypersphere $p_1$ 7 and the SPD $p_1$ 8 manifold, with computational scaling depending on dimension and curvature parameters.
Under mild smoothness, uniqueness of vector field solutions for the regression loss and the validity of manifold continuity equations are guaranteed by differential geometric theory (Davis et al., 24 Oct 2025, Braun et al., 2024, Chen et al., 2023).

This analytical framework provides principled guidance for step-size selection, early stopping, and balancing learning accuracy with discretization constraints, especially as $p_1$ 9 where the marginal density and its score can become singular (Guan et al., 5 Feb 2026).

5. Specialized Methodologies and Generalizations

RFM incorporates several generalizations and methodological variants:

Generalised Flow Maps and Few-Step Sampling: Discrete-time analogues such as Generalised Flow Maps (GFM) permit direct mapping between times $\mathcal{M}$ 0 and $\mathcal{M}$ 1, utilizing self-distillation and higher-order geometric properties, reducing the number of required integration steps (Davis et al., 24 Oct 2025).
Pullback Geometry: When a global diffeomorphism (or learned isometric chart) $\mathcal{M}$ 2 exists, RFM can be pulled back to a latent space for efficient learning with manifold constraints preserved (Kruiff et al., 2024, Collas et al., 20 May 2025).
Discrete Data: Fisher Flow Matching reformulates RFM over the positive orthant of the sphere via the Fisher–Rao metric, using information geometry for categorical and sequence data (Davis et al., 2024).
Product and Structured Manifolds: Multi-stage and product manifold settings handle geometry on spaces such as $\mathcal{M}$ 3 for molecular docking (Matcha), or $\mathcal{M}$ 4 for crystal lattices (FlowMM, FlowLLM) (Frolova et al., 16 Oct 2025, Miller et al., 2024, Sriram et al., 2024).
Stability Enhancements: Stable RFM (SRFMP) leverages Lyapunov functions and invariance principles for time-robust control on Riemannian support (Ding et al., 2024).

6. Application Domains and Empirical Results

RFM and its extensions have demonstrated efficacy across a range of applications:

Uncertainty Quantification in Vision-LLMs: Negative log-density of RFM-computed embedding distributions is an effective epistemic uncertainty proxy, yielding near-perfect correlation with prediction error and outperforming prior baselines in OOD detection and data curation (Ju et al., 29 Jan 2026).
Robot Motion and Policy Synthesis: RFM-based policies (RFMP, SRFMP) achieve smoother, geometry-compliant action trajectories on rotation, pose, or torus manifolds, and deliver efficient, fast inference for visuomotor tasks, matching or exceeding diffusion or consistency models (Braun et al., 2024, Ding et al., 2024).
Crystal and Material Generation: RFM-powered models such as FlowMM and FlowLLM enforce all crystal symmetries and periodicities, yielding state-of-the-art stability, uniqueness, and efficiency in material discovery, with dramatic speedups in ODE steps compared to diffusion models (Miller et al., 2024, Sriram et al., 2024).
Molecular Docking: Multi-stage RFM captures pose refinement across translation, rotation, and torsion, achieving superior accuracy and physical plausibility at fractions of the computational cost of diffusion-based and co-folding baselines (Frolova et al., 16 Oct 2025).
Statistical and Brain Connectivity Data: RFM with pullback geometry efficiently generates valid SPD and correlation matrices, outperforming prior methods and enabling fast, exact and scalable sampling in high-dimensional matrix manifolds (Collas et al., 20 May 2025).

Empirical results consistently demonstrate that RFM stabilizes likelihoods, improves sample quality and realism, and preserves manifold constraints for both synthetic and real-world datasets.

7. Limitations, Open Problems, and Future Directions

While RFM provides a powerful, flexible generative modeling paradigm for manifold-valued data, certain limitations and research opportunities persist:

Scalability on Arbitrary Geometries: Closed-form geodesics are only available on certain manifolds. On general meshes or stratified spaces, efficient spectral or learned premetrics must be deployed, and fully simulation-free training remains an open challenge (Chen et al., 2023, Kruiff et al., 2024).
Boundary and Topological Constraints: RFM maintains manifold constraints via tangent projections and Neumann eigenfunctions, but more intricate or non-Riemannian structures may require further generalization.
Interaction with Base Distributions: Hybrid base distributions, e.g., combining neural LLMs with RFM for continuous refinement, offer strong empirical gains but introduce complexities in end-to-end differentiability and symmetry enforcement (Sriram et al., 2024).
Discretization-Driven Error Accumulation: Ensuring injectivity and accuracy under discretized ODE solvers, especially in high curvature or high-dimensional settings, requires adaptive step-size schemes and accuracy balancing (Guan et al., 5 Feb 2026).
Generalized Objective Variants: Multi-sample and self-distillation objectives, few-step GFMs, and stable autonomous fields represent ongoing research directions for sharpening sample quality and reducing computation (Davis et al., 24 Oct 2025, Ding et al., 2024).

Promising future directions include learning data-driven or biologically informed metrics, extending RFM to stratified or singular spaces, integrating composite objectives for task-stratified generation, and further optimization of hybrid latent-manifold architectures.

Reference Highlights:

REPVLM and epistemic uncertainty: (Ju et al., 29 Jan 2026)
RFMP and SRFMP for robot policy learning: (Braun et al., 2024, Ding et al., 2024)
Generalized manifold flow-matching and GFM: (Davis et al., 24 Oct 2025)
RFM on statistical manifolds/discrete data: (Davis et al., 2024)
RFM for crystals and materials: (Miller et al., 2024, Sriram et al., 2024)
Pullback geometry and applications: (Kruiff et al., 2024, Collas et al., 20 May 2025)
Theory of discretization and convergence: (Guan et al., 5 Feb 2026)

Riemannian Flow Matching thus constitutes a versatile, geometry-native toolkit for simulation-free flow-based modeling across the spectrum of data geometries encountered in contemporary scientific, engineering, and computational domains.