Continuous & Discrete Flow Matching

• Continuous and discrete flow matching is a generative modeling framework that transports data distributions via learned velocity fields or CTMC generators along prescribed probability paths.
• It leverages ODEs, continuity equations, and Kolmogorov forward equations to define dynamics between latent and target distributions in both continuous and discrete settings.
• Recent advances extend the methodology to categorical, ordinal, manifold-valued, and variable-cardinality data, offering improved stability, scalability, and efficiency.

Flow matching denotes a family of generative modeling methods that transport a simple source or prior distribution to a target or data distribution along a probability path. In its continuous form, the learned object is a time-dependent velocity field whose ordinary differential equation transports samples through continuous normalizing flows; in discrete settings, closely related constructions replace vector fields by probability velocities, continuous-time Markov chain generators, or continuous reparameterizations of discrete distributions on structured statistical manifolds (Wald et al., 28 Jan 2025, Gat et al., 2024, Cheng et al., 14 Apr 2025). Recent work has expanded this family to discrete ordinal data, categorical data, mixed continuous–discrete representations, manifold-valued states, and variable-cardinality objects, while also sharpening the mathematical limits imposed by topology, stochasticity, and representation geometry (Shenfeld et al., 1 May 2026, Sha, 14 Dec 2025, Nordlinder et al., 12 Nov 2025).

1. General formulation and probability paths

In the continuous setting, flow matching learns velocity fields of curves connecting a latent distribution and a target distribution. A standard formulation specifies a curve of measures $\mu_t$, $t \in [0,1]$, together with a velocity field $v_t$ satisfying the continuity equation

$\partial_t \mu_t + \nabla_x \cdot (\mu_t v_t) = 0,$

and the sample-wise flow ODE

$$\partial_t \phi(t,x) = v_t(\phi(t,x)), \qquad \phi(0,x)=x,$$

with $\mu_t = \phi(t,\cdot)_{\#}\mu_0$ (Wald et al., 28 Jan 2025). The same review shows that such curves and their velocities can be characterized and learned via transport plans, Markov kernels, and stochastic processes, with the latter two strictly broader than the coupling approach (Wald et al., 28 Jan 2025).

In discrete settings, the corresponding evolution is not written through a Euclidean continuity equation. Instead, probability flow is governed by the Kolmogorov forward equation

$$\frac{d}{dt} p_t(y) = \sum_{x \in \mathcal{X}} u_t(y,x)\,p_t(x),$$

where $u_t(y,x)$ is the transition rate of a continuous-time Markov chain on the discrete state space $\mathcal{X}$ (Khan et al., 5 Feb 2026). Discrete Flow Matching further generalizes the construction by working with a general family of probability paths interpolating between source and target distributions, and by using learned posteriors such as probability denoiser ($x$-prediction) and noise-prediction ($t \in [0,1]$0-prediction) for sampling (Gat et al., 2024).

The main regimes are summarized below.

Regime	State evolution	Canonical object
Continuous flow matching	ODE and continuity equation	Velocity field $t \in [0,1]$1
Discrete-state flow matching	Kolmogorov forward equation	CTMC generator $t \in [0,1]$2
Continuous-state discrete flow matching	Geometry on probability simplex or sphere	Flow in probability representation

This suggests that “flow matching” now names a common transport-based viewpoint rather than a single parameterization. The unifying theme is a prescribed probability path from source to target together with a learned dynamical object that makes sampling feasible (Wald et al., 28 Jan 2025, Cheng et al., 14 Apr 2025).

2. Continuous flow matching and the problem of continuity

Flow matching emerged as a framework for generative modeling through continuous normalizing flows (Sha, 14 Dec 2025). A typical formulation connects a prior $t \in [0,1]$3 and a target $t \in [0,1]$4 through an ODE

$t \in [0,1]$5

with the associated continuity equation

$t \in [0,1]$6

For Gaussian conditionals, conditional flow matching yields the optimal velocity

$t \in [0,1]$7

(Sha, 14 Dec 2025).

A central theoretical development concerns topological mismatch between the prior and target. When a unimodal prior is transported to a multimodal target, the optimal velocity field under standard flow matching objectives may be spatially discontinuous (Sha, 14 Dec 2025). In the paper’s bimodal Gaussian mixture analysis, the intermediate distribution is a mixture of two overlapping Gaussians, and there is a decision-boundary hyperplane

$t \in [0,1]$8

across which the posterior destination flips sharply from one mode to the other (Sha, 14 Dec 2025). The resulting jump discontinuity has magnitude approaching infinity as $t \in [0,1]$9 (Sha, 14 Dec 2025).

The distinction between temporal and spatial continuity is therefore essential. Temporal continuity of trajectories with respect to $v_t$0 is maintained by ODEs, but spatial continuity with respect to $v_t$1 at fixed $v_t$2 is not guaranteed (Sha, 14 Dec 2025). Because neural networks used to parameterize $v_t$3 are by design spatially continuous, they can only learn a smoothed approximation when the true optimal velocity is discontinuous. The paper identifies mode averaging and mode collapse as resulting artifacts, and gives a lower-bounded approximation error that diverges as $v_t$4 (Sha, 14 Dec 2025).

A common misconception is that continuous sample paths imply a continuous velocity field everywhere in state space. Theoretical analysis in the bimodal setting shows that this is false: continuous trajectories can coexist with spatial jump discontinuities in the optimal velocity field (Sha, 14 Dec 2025). The same work further argues that the issue is not specific to $v_t$5 loss, but may be a consequence of topological mismatch between distributions, with implications for manifold flow matching and representation learning (Sha, 14 Dec 2025).

3. Discrete flow matching, stochastic paths, and generator learning

Discrete Flow Matching was introduced as a discrete flow paradigm designed specifically for generating discrete data, with four stated contributions: a general family of probability paths, a generic formula for sampling from these paths using learned posteriors, practical gains from specific schedulers, and strong scaling on code generation benchmarks (Gat et al., 2024). In its simple convex-interpolation form, the conditional path can be written as

$v_t$6

while the marginal path is induced through a coupling $v_t$7 (Gat et al., 2024). Sampling is then driven by a probability velocity, for example

$v_t$8

with an analogous $v_t$9-prediction form using $\partial_t \mu_t + \nabla_x \cdot (\mu_t v_t) = 0,$0 (Gat et al., 2024).

The discrete setting differs structurally from the continuous one because its sample paths are stochastic. For that reason, the rectification strategy used in continuous flow matching does not directly extend to the discrete one (Haxholli et al., 2024). To address this, recent work proposes a dynamic-optimal-transport-like minimization objective for discrete flows with convex interpolants and derives an equivalent Kantorovich formulation in which transport cost depends only on inter-state similarity; this cost can be optimized with a minibatch strategy (Haxholli et al., 2024). The same paper also derives an upper bound on perplexity, motivated by the fact that discrete models lack an instantaneous change-of-variables mechanism and therefore do not admit the same density evaluation tools as continuous flows (Haxholli et al., 2024).

A complementary line of work studies discrete flow through generator matching. “Error Analysis of Discrete Flow with Generator Matching” derives a novel Girsanov-type theorem for CTMCs and expresses the KL divergence between two path measures through a Bregman divergence between the transition rates (Wan et al., 26 Sep 2025). Building on generator matching and uniformization, it establishes non-asymptotic error bounds for distribution estimation and emphasizes a contrast with discrete diffusion: discrete flow incurs no truncation error caused by truncating the time horizon in the noising process (Wan et al., 26 Sep 2025).

A second common misconception is that discrete flow matching is merely a tokenized version of continuous flow matching. The literature instead presents a distinct stochastic theory: CTMC generators replace vector fields, rectification no longer transfers directly, and evaluation requires tools such as perplexity upper bounds, path-space KL analysis, or uniformization-based error bounds (Haxholli et al., 2024, Wan et al., 26 Sep 2025).

4. Geometry, manifolds, and continuous-state formulations for discrete data

Several approaches model discrete data by embedding probability distributions into continuous geometric spaces. Fisher-Flow treats categorical distributions as points on a statistical manifold equipped with the Fisher-Rao metric,

$\partial_t \mu_t + \nabla_x \cdot (\mu_t v_t) = 0,$1

and uses the sphere map $\partial_t \mu_t + \nabla_x \cdot (\mu_t v_t) = 0,$2 as an isometry from the simplex interior to the positive orthant of the hypersphere $\partial_t \mu_t + \nabla_x \cdot (\mu_t v_t) = 0,$3 (Davis et al., 2024). On this hypersphere, flows are defined along closed-form geodesics, can be bootstrapped by Riemannian optimal transport, and the induced gradient flow is proved optimal in reducing the forward KL divergence (Davis et al., 2024).

A broader unification is provided by $\partial_t \mu_t + \nabla_x \cdot (\mu_t v_t) = 0,$4-Flow, which presents a unified framework for Continuous-State Discrete Flow Matching models. Its key device is the $\partial_t \mu_t + \nabla_x \cdot (\mu_t v_t) = 0,$5-representation

$\partial_t \mu_t + \nabla_x \cdot (\mu_t v_t) = 0,$6

under which prior CS-DFM variants correspond to different $\partial_t \mu_t + \nabla_x \cdot (\mu_t v_t) = 0,$7-geometries (Cheng et al., 14 Apr 2025). The paper states that $\partial_t \mu_t + \nabla_x \cdot (\mu_t v_t) = 0,$8-Flow adheres to the canonical $\partial_t \mu_t + \nabla_x \cdot (\mu_t v_t) = 0,$9-geometry of the statistical manifold, is optimal in minimizing the generalized kinetic energy, and yields a unified variational bound for the discrete negative log-likelihood (Cheng et al., 14 Apr 2025).

A related construction uses assignment manifolds. “Generative Modeling of Discrete Joint Distributions by E-Geodesic Flow Matching on Assignment Manifolds” defines a continuous normalizing flow on the submanifold of factorizing discrete measures

$$\partial_t \phi(t,x) = v_t(\phi(t,x)), \qquad \phi(0,x)=x,$$0

embedded into the meta-simplex by

$$\partial_t \phi(t,x) = v_t(\phi(t,x)), \qquad \phi(0,x)=x,$$1

and trains the model by matching geodesic flows of factorizing discrete distributions (Boll et al., 2024). The construction gradually assigns categories and avoids issues of discretizing the latent continuous model such as rounding and sample truncation (Boll et al., 2024).

For discrete non-negative ordinal data, Binomial Flows provide another bridge. The forward conditional is a binomial thinning process,

$$\partial_t \phi(t,x) = v_t(\phi(t,x)), \qquad \phi(0,x)=x,$$2

the denoiser is $$\partial_t \phi(t,x) = v_t(\phi(t,x)), \qquad \phi(0,x)=x,$$3, and the central identity is a discrete Tweedie’s formula,

$$\partial_t \phi(t,x) = v_t(\phi(t,x)), \qquad \phi(0,x)=x,$$4

where $$\partial_t \phi(t,x) = v_t(\phi(t,x)), \qquad \phi(0,x)=x,$$5 is the intensity of a Poisson-Föllmer process used for sampling (Shenfeld et al., 1 May 2026). The same framework yields exact likelihoods and unbiased log-likelihood estimators for ordinal discrete data (Shenfeld et al., 1 May 2026).

These geometric constructions share a common premise: discrete generation need not be confined to jump processes on finite sets. A plausible implication is that continuous-state representations can recover analytical tools—geodesics, Riemannian metrics, conditional means, exact likelihood expressions—that are difficult to obtain in purely discrete-state models (Davis et al., 2024, Cheng et al., 14 Apr 2025, Shenfeld et al., 1 May 2026).

5. Hybrid and hierarchical bridges between continuous and discrete flow matching

Recent systems often combine discrete and continuous mechanisms rather than choosing one exclusively. PolyFlow addresses the incompatibility between mesh connectivity and standard continuous denoising by introducing a compact topology embedder that projects discrete mesh topology into continuous per-vertex embeddings; after pretraining and freezing this embedder, a Transformer-based flow-matching model denoises joint vertex states in parallel and decodes topology via spacetime distance thresholding (Wang et al., 25 Jun 2026). The state of each vertex is

$$\partial_t \phi(t,x) = v_t(\phi(t,x)), \qquad \phi(0,x)=x,$$6

and inference proceeds by solving

$$\partial_t \phi(t,x) = v_t(\phi(t,x)), \qquad \phi(0,x)=x,$$7

with an Euler ODE solver (Wang et al., 25 Jun 2026).

Flow6D uses a two-stage discrete-to-continuous hierarchy for category-level 6D pose estimation. Rotation and translation are first discretized into bins and localized with a discrete flow matching model; a continuous flow matching model then predicts local pose residuals to refine the estimate (Mei et al., 22 Jun 2026). This two-stage discrete latent space localization–continuous pose regression strategy is explicitly designed to reduce search complexity and then optimize accuracy (Mei et al., 22 Jun 2026).

Purrception targets vector-quantized image generation by combining explicit categorical supervision with continuous transport dynamics. It learns categorical posteriors over codebook indices while computing velocity fields in the continuous embedding space, with

$$\partial_t \phi(t,x) = v_t(\phi(t,x)), \qquad \phi(0,x)=x,$$8

where $$\partial_t \phi(t,x) = v_t(\phi(t,x)), \qquad \phi(0,x)=x,$$9 is the posterior barycenter over codebook embeddings (Matişan et al., 1 Oct 2025). The paper emphasizes uncertainty quantification over plausible codes and temperature-controlled generation, and reports faster convergence than both continuous flow matching and discrete flow matching baselines on ImageNet-1k $\mu_t = \phi(t,\cdot)_{\#}\mu_0$0 generation (Matişan et al., 1 Oct 2025).

Branching Flows generalize the bridge further by adding stochastic splits and deletions. They transport a simple distribution to the data distribution while allowing the number of elements in the state to vary over a forest of binary trees, and they compose with any flow matching base process on discrete sets, continuous Euclidean spaces, smooth manifolds, and multimodal product spaces (Nordlinder et al., 12 Nov 2025). This directly targets settings in which the number of elements is not known a priori (Nordlinder et al., 12 Nov 2025).

For heterogeneous tabular data, Cascaded Flow Matching adopts a low-resolution discrete stage followed by a high-resolution flow-matching stage. The first stage generates categorical features and coarse categorical versions of numerical features; the second stage then generates detailed numerical values with a guided conditional probability path and data-dependent coupling, and the paper proves that the cascade tightens the transport cost bound (Mueller et al., 30 Jan 2026).

6. Applications, empirical behavior, and open technical issues

The empirical scope of continuous and discrete flow matching is broad. Discrete Flow Matching scaled to 1.7B parameters reaches $\mu_t = \phi(t,\cdot)_{\#}\mu_0$1 Pass@1 and $\mu_t = \phi(t,\cdot)_{\#}\mu_0$2 Pass@10 on HumanEval, and $\mu_t = \phi(t,\cdot)_{\#}\mu_0$3 Pass@1 and $\mu_t = \phi(t,\cdot)_{\#}\mu_0$4 Pass@10 on 1-shot MBPP coding benchmarks (Gat et al., 2024). In discrete ordinal modeling, Binomial Flows report a CIFAR-10 FID score of $\mu_t = \phi(t,\cdot)_{\#}\mu_0$5 for 8-bit image generation (Shenfeld et al., 1 May 2026). PolyFlow reports improved Chamfer Distance and Hausdorff Distance on Toys4K mesh generation, while Flow6D reports real-time inference at $\mu_t = \phi(t,\cdot)_{\#}\mu_0$6 FPS for category-level 6D pose estimation (Wang et al., 25 Jun 2026, Mei et al., 22 Jun 2026).

Reinforcement learning supplies a distinct application class. DRIFT models a discrete-action policy as a CTMC generator, updates an offline pretrained policy with an advantage-weighted discrete flow matching loss, and adds a path-space penalty that regularizes the full CTMC trajectory distribution rather than only the final action distribution (Khan et al., 12 May 2026). A separate offline RL framework replaces continuous flows with CTMCs, uses a Q-weighted flow matching objective, extends to multi-agent settings through a factorized conditional path, and shows that under idealized conditions optimizing this objective recovers the optimal policy (Khan et al., 5 Feb 2026).

Molecular generation highlights both achievements and limitations. FlowMol-CTMC benchmarks several discrete flow matching methods for 3D de novo molecule generation and reports that CTMC-based flow matching dramatically improves molecular stability and validity relative to continuous surrogates (Dunn et al., 2024). The same study introduces metrics beyond local chemical valency constraints and finds that, even though basic constraints are satisfied, the models tend to produce unusual and potentially problematic functional groups outside of the training data distribution (Dunn et al., 2024). This provides an objective caution against treating sanitization or valency satisfaction as sufficient evaluation criteria.

Training efficiency and variance reduction remain active issues. Temporal Pair Consistency couples velocity predictions at paired timesteps along the same probability path, operates entirely at the estimator level without modifying the model architecture, probability path, or solver, and is shown theoretically to induce a quadratic, trajectory-coupled regularization that reduces gradient variance while preserving the underlying flow-matching objective (Maduabuchi et al., 4 Feb 2026). The method improves FID at identical or lower computational cost than prior methods on CIFAR-10 and ImageNet, and extends to score-based denoising and rectified flow pipelines (Maduabuchi et al., 4 Feb 2026).

Several open technical issues recur across the literature. Topological mismatch can force discontinuous optimal velocities in continuous flow matching (Sha, 14 Dec 2025). In discrete flow matching, stochastic paths prevent a direct transfer of continuous rectification and complicate exact evaluation (Haxholli et al., 2024). In continuous relaxations of discrete data, assignment lag can degrade coherence, as observed in molecular generation benchmarks (Dunn et al., 2024). And for variable-cardinality data, fixed-size flow assumptions require explicit mechanisms such as branching, deletion, or cascaded coarse-to-fine generation (Nordlinder et al., 12 Nov 2025, Mueller et al., 30 Jan 2026). Collectively, these results indicate that the central design question is not whether a problem is “continuous” or “discrete” in isolation, but which probability path, geometry, stochastic process, and representation make the target transport learnable.