Riemannian Flow Matching

• Riemannian Flow Matching is a geometric framework that constructs generative models on manifolds by learning flows aligned with intrinsic curvature, geodesic distances, and topology.
• It employs closed-form geodesics and simulation-free training objectives, bypassing traditional score estimation and divergence computation for efficient probability transport.
• RFM underpins state-of-the-art applications in geometry, molecular design, and robotics, offering high fidelity, speed, and robust theoretical convergence guarantees.

Riemannian Flow Matching (RFM) is a geometric framework for constructing generative models that transport probability distributions on manifolds by learning flows consistent with the manifold's Riemannian structure. Unlike traditional flow-based models defined in Euclidean space, RFM captures intrinsic geometric properties—such as curvature, geodesic distance, and manifold topology—allowing simulation-free, divergence-free training objectives and efficient, high-fidelity generative modeling on a broad class of manifolds. The framework generalizes core concepts from continuous normalizing flows, optimal transport, and conditional flow matching, and underpins a growing class of state-of-the-art models in geometric deep learning, discrete data modeling, and generative tasks on structured domains.

1. Mathematical Foundation

Riemannian Flow Matching operates on a smooth, connected, complete Riemannian manifold $(\mathcal{M}, g)$ of dimension $d$, where $g$ is a smoothly varying metric tensor defining an inner product $\langle u, v \rangle_g = u^{\top} g(x) v$ for $u, v \in T_x \mathcal{M}$. At the heart of RFM is the flow ODE: $\frac{dx_t}{dt} = v_t(x_t),\quad x_0 \sim p_0$ where $v_t : \mathcal{M} \rightarrow T\mathcal{M}$ is a time-dependent vector field, and $x_0$ is sampled from a simple base distribution $p_0$ supported on $\mathcal{M}$. The temporal evolution pushes $d$0 forward to a target distribution $d$1 through the induced diffeomorphism $d$2.

Mass conservation along the flow is governed by the Riemannian continuity equation: $d$3 where $d$4 denotes the intrinsic divergence associated with the volume form $d$5.

RFM introduces a premetric $d$6, typically chosen as the geodesic distance $d$7 induced by $d$8, which admits the exponential and logarithm maps: $d$9 such that $g$0 iff $g$1. Geodesics $g$2 encode shortest paths between points.

The target velocity field for conditional flow matching is derived by differentiating along geodesics: $g$3 The training objective is then the squared-norm error in the Riemannian metric: $g$4 This formulation bypasses the need for score estimation and divergence computation, and is simulation-free for manifolds with closed-form geodesics (Chen et al., 2023).

2. Extensions and Key Variants

2.1 Variational Riemannian Flow Matching

RG-VFM (Riemannian Gaussian Variational Flow Matching) generalizes variational flow matching to Riemannian manifolds by optimizing a KL-based variational objective for endpoint matching. The posterior is modeled as a Riemannian Gaussian: $g$5 with loss

$g$6

On homogeneous spaces with closed-form geodesics, RG-VFM yields an unbiased geodesic-MSE loss and shares the computational advantages of RFM, but endpoint-matching is fundamentally distinct from the velocity-matching of RFM (Zaghen et al., 18 Feb 2025).

2.2 Flow Matching on Lie Groups

For matrix Lie groups $g$7, straight lines are replaced with exponential curves: $g$8 with corresponding tangent vectors

$g$9

This approach exploits only closed-form group operations and supports fast simulation-free generative modeling on, e.g., $\langle u, v \rangle_g = u^{\top} g(x) v$0 or $\langle u, v \rangle_g = u^{\top} g(x) v$1 (Sherry et al., 1 Apr 2025).

2.3 Statistical Manifolds and Discrete Data

On the manifold of categorical distributions (the simplex $\langle u, v \rangle_g = u^{\top} g(x) v$2) endowed with the Fisher–Rao geometry, geodesic interpolation maps to the positive orthant of the sphere via $\langle u, v \rangle_g = u^{\top} g(x) v$3, with closed-form geodesics: $\langle u, v \rangle_g = u^{\top} g(x) v$4 Riemannian Flow Matching on these manifolds yields state-of-the-art performance on discrete generative modeling by leveraging optimal transport and exact likelihoods (Cheng et al., 2024, Davis et al., 2024).

3. Algorithmic and Practical Considerations

A key benefit of RFM is the simulation-free, closed-form computation of the supervision signals for many simple geometries. The general algorithm for RFM training is:

• For each batch, sample $\langle u, v \rangle_g = u^{\top} g(x) v$5, $\langle u, v \rangle_g = u^{\top} g(x) v$6, $\langle u, v \rangle_g = u^{\top} g(x) v$7.
• Construct geodesic interpolation $\langle u, v \rangle_g = u^{\top} g(x) v$8.
• Compute the analytic target velocity $\langle u, v \rangle_g = u^{\top} g(x) v$9.
• Predict $u, v \in T_x \mathcal{M}$0 with a neural network projecting into the appropriate tangent space.
• Minimize the squared Riemannian-norm loss $u, v \in T_x \mathcal{M}$1.

For manifolds lacking closed-form geodesics, RFM can use spectral approximations (eigenmaps and Laplace–Beltrami eigenfunctions) as surrogate premetrics (Chen et al., 2023, Huang et al., 2 Oct 2025).

Sampling requires integration of the trained vector field ODE: $u, v \in T_x \mathcal{M}$2 with geodesic projection at each step to respect manifold constraints.

4. Theoretical Guarantees and Convergence

RFM is consistent in the sense that as the learned field $u, v \in T_x \mathcal{M}$3 approaches the optimal conditional field, the pushforward law converges to the data law in distribution (Chen et al., 2023). Recent non-asymptotic analyses establish explicit total variation convergence rates of the form

$u, v \in T_x \mathcal{M}$4

where $u, v \in T_x \mathcal{M}$5 is the Euler step size and $u, v \in T_x \mathcal{M}$6 is the field approximation error (Guan et al., 5 Feb 2026). These results hold under smoothness and curvature conditions for compact and Hadamard manifolds, with explicit iteration complexities available for the hypersphere $u, v \in T_x \mathcal{M}$7 and SPD$u, v \in T_x \mathcal{M}$8 matrices.

5. Applications and Empirical Advances

Riemannian Flow Matching forms the backbone of generative models in a diverse array of geometric domains:

• Geometry and shape synthesis: Learning flows on spheres, tori, hyperbolic spaces, and general surfaces yields high-fidelity data generation and tractable likelihood estimation (Chen et al., 2023, Davis et al., 24 Oct 2025).
• Protein, molecule, and material design: RFM powers multi-stage pipelines for molecular docking (Matcha (Frolova et al., 16 Oct 2025)), crystal-structure prediction (FlowMM (Miller et al., 2024)), and disordered material generation (DMFlow (Wu et al., 4 Feb 2026)).
• Graph and matrix data: Spectral Geodesic Flow Matching for graphs (SFMG (Huang et al., 2 Oct 2025)) and pullback-based RFM for SPD and correlation matrices (DiffeoCFM (Collas et al., 20 May 2025)).
• Robotics and control: RFM-based policies enable geometrically aware trajectory inference and real-time visuomotor control with competitive or superior smoothness and inference efficiency compared to diffusion models (Braun et al., 2024, Ding et al., 2024).
• Discrete and statistical manifolds: Fisher–Rao RFM underlies SOTA models for categorical and biological sequence data generation (Davis et al., 2024, Cheng et al., 2024).

Across these settings, RFM has demonstrated strong sample quality, simulation-free inference, geometric faithfulness, and efficient training.

6. Limitations, Open Directions, and Relations

Current implementations of RFM are most efficient when the target manifold admits closed-form geodesics and tractable exponential/logarithm maps; generalization to arbitrary manifolds often requires eigenfunction-based spectral distances (Chen et al., 2023). Limitations include potential scalability issues for very high dimensions (in the spectral case), and the need for further advances in self-distillation and few-step generative flows on highly curved or singular manifolds (Davis et al., 24 Oct 2025).

RFM is fundamentally distinct from stochastic score-based generative models (diffusions), simulation-heavy continuous normalizing flows, or variational endpoint-matching frameworks (RG-VFM (Zaghen et al., 18 Feb 2025)). The velocity-matching loss in RFM is unbiased and directly expresses the conditional optimal transport flow on the manifold, in contrast to endpoint-based objectives which differ outside of Euclidean space.

7. Empirical Benchmarks and Comparative Performance

A wide range of empirical studies have confirmed RFM's advantages:

• On the hypersphere $u, v \in T_x \mathcal{M}$9, RFM and geometric variants achieve $\frac{dx_t}{dt} = v_t(x_t),\quad x_0 \sim p_0$0 compared to $\frac{dx_t}{dt} = v_t(x_t),\quad x_0 \sim p_0$1 for Euclidean flows;
• In molecular docking, RFM-based Matcha achieves $\frac{dx_t}{dt} = v_t(x_t),\quad x_0 \sim p_0$2\% success (RMSD$\frac{dx_t}{dt} = v_t(x_t),\quad x_0 \sim p_0$3) on Astex versus $\frac{dx_t}{dt} = v_t(x_t),\quad x_0 \sim p_0$4\% for AlphaFold 3, and is $\frac{dx_t}{dt} = v_t(x_t),\quad x_0 \sim p_0$5 faster than co-folding models (Frolova et al., 16 Oct 2025);
• In dense graph generation, SFMG matches state-of-the-art on degree, spectral, and clustering metrics with up to $\frac{dx_t}{dt} = v_t(x_t),\quad x_0 \sim p_0$6 speedup over diffusion models (Huang et al., 2 Oct 2025);
• For brain connectivity matrices, DiffeoCFM yields superior alignment to manifold constraints and F1-scores compared to post-projection diffusion and flow baselines (Collas et al., 20 May 2025);
• On real-world discrete-sequence benchmarks, RFM-based categorical models surpass Dirichlet and discrete-diffusion flows in NLL and domain-specific metrics (Davis et al., 2024, Cheng et al., 2024).

RFM thus defines a mathematically rigorous, computationally efficient, and empirically validated framework for manifold-aware generative modeling across continuous, discrete, and structured data domains.