Simulation-Free Score and Flow Matching

• The paper presents [SF]²M, a simulation-free framework that unifies score matching and flow matching to model continuous-time stochastic processes without path simulation.
• It uses analytic Brownian bridge distributions and optimal transport couplings to derive closed-form regression targets for both drift and score fields, ensuring scalable high-dimensional training.
• The approach achieves state-of-the-art performance in latent SDE modeling and multi-marginal interpolation, offering improved computational efficiency and accuracy.

Simulation-Free Score and Flow Matching ([SF]$^2$M) is a unified framework for learning continuous-time stochastic dynamics that achieves generative modeling, inference, and trajectory alignment without requiring simulation of sample paths during training. It generalizes both score matching—central to diffusion models—and flow matching—core to continuous normalizing flows—within the Schrödinger Bridge (SB) formulation and related variational SDE formulations. By substituting static, analytically tractable Brownian bridge distributions and entropy-regularized optimal transport (OT) couplings for numerical SDE or ODE simulations, [SF]$^2$M enables scalable, statistically efficient learning on high-dimensional data, including time series and snapshot measurements at irregular intervals (Tong et al., 2023, Bartosh et al., 4 Feb 2025, Lee et al., 6 Aug 2025).

1. Conceptual Foundations

The foundational insight of [SF]$^2$M is to express continuous-time stochastic generative modeling as a Schrödinger Bridge problem, seeking the most likely stochastic evolution (in relative entropy to reference Brownian motion) matching specified input and output distributions. For distributions $q_0$ and $q_1$, the SB interpolates them by minimizing $\mathrm{KL}(\mathbb{P} \| \mathbb{Q})$ over all path measures $\mathbb{P}$ with $p_0 = q_0$, $p_1 = q_1$, where $\mathbb{Q}$ is Brownian motion. The optimal $\mathbb{P}^*$ is the Markovization of a mixture of Brownian bridges, weighted by an entropic-OT plan $\pi^*_{2\sigma^2}(q_0, q_1)$.

The SDE dynamics are parameterized as

$dx_t = u_t(x_t)\,dt + g(t)\,dw_t$

with corresponding Fokker–Planck evolution for the interpolating marginals $p_t$. [SF]$^2$M learns both the drift $u_t(x_t)$ and the score $\nabla\log p_t(x_t)$ fields, unifying continuous normalizing flows and score-based generative models.

2. Methodology and Training Objective

Training is conducted by regressing neural approximations of analytic drift and score fields from Brownian bridge mixtures, without forward-simulating SDE trajectories. The approach uses two core ideas across three main methodologies:

• Static Conditional Regression: For pairs or tuples $(x_0, x_1,...,x_K)$ drawn from a minibatch approximation of the entropic-OT plan, and for interpolation time $t$, [SF]$^2$M samples from the analytic Brownian bridge distribution $p_t(x|z)$ and computes closed-form expressions for the drift and score,

$u_t^\circ(x|z), \qquad \nabla_x \log p_t(x|z).$

These become the regression targets for neural nets $v_\theta(t, x)$ and $s_\theta(t, x)$ respectively.

• Combined Conditional Loss: The loss for the two-marginal case is

$$\mathcal{L}_{[\mathrm{SF}]^2\mathrm{M}} = \mathbb{E}_{t,z,x}\left[ \|v_\theta(t,x) - u_t^\circ(x|z)\|^2 + \lambda(t)^2 \|s_\theta(t,x) - \nabla_x \log p_t(x|z)\|^2 \right],$$

where $\lambda(t)$ is a time-dependent weight that stabilizes training near $t=0,1$ (Tong et al., 2023, Lee et al., 6 Aug 2025). The theoretical guarantee is that, at global optima, the learned fields match the correct SB interpolants (Tong et al., 2023). The unconditional version of the loss matches the targeted marginal drift and score, and equivalence of gradients is formally established.

• Simulation-free SDE Training (Latent Dynamics): In the variational SDE context, [SF]$^2$M provides a simulation-free surrogate for the negative log-likelihood bound by expressing the pathwise KL via a Monte Carlo expectation over reparameterized samples, never requiring solution of ODE/SDEs during training (Bartosh et al., 4 Feb 2025). The loss decomposes as

$$L_{\text{total}} = L_{\text{prior}} + L'_{{\rm diff}} + L_{\text{rec}} + \alpha_s L_{\text{score}} + \alpha_f L_{\text{flow}},$$

where each term has a closed-form Monte Carlo estimator based on explicit reparameterizations—no adjoint method or numerical solver is required.

3. Multi-Marginal and Irregular Timepoint Extensions

[SF]$^2$M extends naturally to the multi-marginal case, enabling trajectory inference and generative modeling from snapshot data at arbitrary and irregular time points without dimensionality reduction (Lee et al., 6 Aug 2025). The method constructs measure-valued splines across overlapping time windows, approximates the multi-marginal OT plan via a first-order Markov factorization, and defines regression objectives on analytic bridge interpolations:

• For a window $[t_i, t_{i+k}]$, a conditional Gaussian bridge is constructed, and neural nets regress on its analytic drift and score.
• The aggregate loss is a sum over all windows, stratifying time sampling to ensure coverage.
• The resulting framework enforces mass conservation (continuity PDE constraints) and correct stochastic transport, while score matching regularizes high-dimensional learning, preventing overfitting.

4. Implementation and Optimization Details

Implementation of [SF]$^2$M in both Schrödinger bridge and latent SDE contexts leverages the following structures:

• Neural Architectures: For drift and score fields, 3-layer MLPs are commonly used; alternative architectures (e.g., UNet) are employed for image and high-dimensional gene data (Tong et al., 2023, Lee et al., 6 Aug 2025, Bartosh et al., 4 Feb 2025).
• Conditional Bridge Regression: All training samples are generated from static bridge distributions using OT coupling, ensuring analytic availability of regression targets.
• Memory and Time Complexity: Per-batch OT costs $O(m^2 d)$ (with entropic regularization), typically $<1\%$ of overall training cost. No SDE simulation is performed, so wall-clock time and memory scale as $O(1)$ per SGD step, contrasting with $O(L\log L)$ or more for solver-based adjoint methods (Bartosh et al., 4 Feb 2025).
• Optimization: AdamW (or Adam) with prescribed learning rates, batch sizes (e.g., 512), and carefully chosen time-dependent weights $\lambda(t)$ for score loss regularization (e.g., $\lambda(t) = 2\sigma\sqrt{t(1-t)}/\sigma^2$).
• OT Coupling: Exact discrete OT is feasible for batches $m\leq 10^4$; otherwise, Sinkhorn regularization is used, with the entropic penalty set to $2\sigma^2$ (Tong et al., 2023).

5. Theoretical Guarantees

The convergence properties of [SF]$^2$M are established under general assumptions (Tong et al., 2023, Bartosh et al., 4 Feb 2025, Lee et al., 6 Aug 2025):

• Equivalence of Loss Gradients: The conditional regression loss achieves the same minimizers as the (intractable) unconditional marginal loss, ensuring that the learned drift and score fields solve the corresponding SB or conditional generative modeling problem.
• Consistency: For sufficient network expressivity and optimization, the learned stochastic process exactly recovers the governing bridge or variational SDE, and the variational bound is tight.
• Preclusion of Overfitting: Inclusion of the score-matching term penalizes degenerate solutions in high-dimensional settings, matching all infinite-dimensional statistics encoded by the log-density gradient.

6. Empirical Performance and Applications

[SF]$^2$M demonstrates state-of-the-art performance across a range of synthetic and biological datasets:

• SB Interpolation (2D, High-d): Achieves lowest Wasserstein errors and path energies on 2D synthetic tasks (Gaussian $\to$ moons, S-curve) and tight KLs on high-dimensional Gaussian SB tasks ($d=5,20,50$) (Tong et al., 2023).
• Latent SDE Sequence Modeling: Matches or surpasses adjoint-based SDE training in test MSE on 50-dimensional motion capture data, with $\sim500\times$ speed-up compared to adjoint sensitivity and $100\times$ fewer SDE evaluations versus prior simulation-free ARCTA (Bartosh et al., 4 Feb 2025).
• Snapshot Cell Dynamics: Accurately interpolates cell population densities in high-dimensional gene expression data, recovers smooth Waddington potential landscapes, and enables network inference (AUC-ROC $0.72$–$0.79$ on synthetic gene regulatory networks) (Tong et al., 2023).
• Multi-Marginal and Irregular Snapshot Problems: Consistently outperforms competing approaches (e.g., MIOFlow) on real and synthetic irregular snapshot interpolation, delivering improved held-out marginal fitting and generative smoothness (Lee et al., 6 Aug 2025).

A summary of empirical settings is provided below.

Task	Key Result	Reference
Gaussian $\to$ Moons	Lowest $W_2$ error/path energy vs OT-CFM, DSB	(Tong et al., 2023)
50-D mocap, latent SDE	MSE $4.50\pm0.32$ (vs $4.03\pm0.20$ for adjoint); $5\times$ faster	(Bartosh et al., 4 Feb 2025)
High-dimensional gene	Interpolates/recovers gene networks at $d=1000$	(Tong et al., 2023)
Multi-marginal (images)	Triplet SF$^2$M yields smoother/accurate interpolation	(Lee et al., 6 Aug 2025)

7. Practical Recommendations and Limitations

Best practices for effective application of [SF]$^2$M include:

• Time Sampling and Weighting: Uniform sampling of $t$ is preferred for simplicity; importance weighting can reduce variance.
• Bridge Priors: Choice of Euclidean versus geodesic OT cost affects interpolation on structured manifolds; the latter can yield improved fit on curved data.
• Regularization: Single Monte Carlo samples per update suffice; divergence terms simplify for diagonal noise.
• No Simulation Requirements: At no point is backpropagation through a solver necessary. All gradients flow through static analytic expressions, maximizing hardware efficiency.
• Limitations: OT computation, though negligible relative to network training for moderately sized batches, can present a bottleneck for extremely large sample sets. Edge cases for mini-batch OT in bifurcating structures may present challenges, as observed in high-dimensional single-cell bifurcation experiments (Lee et al., 6 Aug 2025).

Overall, [SF]$^2$M delivers a consistent, simulation-free pipeline for training continuous-time stochastic models in both generative and inference settings, scaling from low-dimensional trajectories to complex multi-marginal and high-dimensional data domains without resorting to single trajectory simulation at training time (Tong et al., 2023, Bartosh et al., 4 Feb 2025, Lee et al., 6 Aug 2025).