Recursive Flow Matching

Abstract: Generative models have emerged as a powerful paradigm for solving physics systems and modeling complex spatiotemporal dynamics. However, achieving high physical accuracy without incurring high computational cost remains a fundamental challenge, as existing approaches face a critical speed-fidelity trade-off. In this work, we introduce Recursive Flow Matching (RecFM), a generative framework for forecasting complex spatiotemporal dynamics. RecFM enforces self-consistency to align trajectories across discretization scales, reducing discretization errors and improving performance across metrics for physics-based tasks. To our knowledge, this is the first method to achieve high-fidelity one- and few-step (2-4 step) dynamic generation for scientific systems with performance comparable to state-of-the-art multi-step solvers. Across challenging scientific benchmarks, RecFM achieves up to a 20$\times$ speedup over leading diffusion-based emulators while improving predictive accuracy. Furthermore, RecFM reduces mean squared error by over 15% compared to vanilla flow matching, offering a scalable and efficient solution for real-time scientific emulation.

• The paper introduces Recursive Flow Matching, a method that enforces cross-scale trajectory consistency to achieve high-fidelity generation with only 1-2 inference steps.
• It recursively scales velocity fields with multi-scale supervision to minimize truncation errors and accelerate convergence by up to 20× compared to diffusion models.
• Experimental results demonstrate significant improvements in MSE and physical fidelity across benchmarks like Navier-Stokes and SST, validating its efficiency and accuracy.

Recursive Flow Matching: A Consistency-Based Generative Framework for Efficient Scientific Emulation

Introduction and Motivation

The paper "Recursive Flow Matching" (2605.26535) addresses the fundamental challenge of achieving high-fidelity, efficient generative modeling for spatiotemporal scientific systems, particularly those governed by nonlinear dynamics or partial differential equations (PDEs). Classical numerical solvers provide strong accuracy but are computationally prohibitive for real-time tasks and large-scale deployments. Existing deep generative frameworks such as diffusion models and continuous normalizing flows mitigate some of these challenges but face a speed-fidelity trade-off: most require tens to hundreds of iterative denoising or ODE steps, leading to high sampling costs and error accumulation, especially for long-term rollouts.

The core contribution of this work is the introduction of Recursive Flow Matching (RecFM), a novel generative framework that enforces cross-scale trajectory consistency, enabling few-step and even single-step high-fidelity generation for scientific systems. RecFM unifies and extends principles from flow matching, self-consistency modeling, and multi-scale supervision, breaking key limitations of standard flow-based and diffusion-based solvers. The approach is validated across a suite of challenging benchmarks, demonstrating up to 20× acceleration and significant improvements in predictive accuracy over strong diffusion and flow matching baselines.

Methodology: Recursive Flow Matching

Flow Matching and Consistency

Traditional flow matching (FM) methods configure a time-dependent vector field to transport samples from a target data distribution to a reference noise distribution. FM aims for efficiency but, when inference step counts are reduced, fidelity suffers due to integration errors and insufficient regularization of the underlying ODE trajectory. Recent work has applied self-consistency constraints to enforce semigroup properties of the learned flow, but these approaches provide limited supervision and do not exploit the full spectrum of geometric possibilities at each state.

Recursive Formulation and Multi-Scale Supervision

RecFM advances over FM by defining a family of recursively scaled trajectories, each passing through shared spatial states at different scales (parameterized by $\alpha$). The learned velocity network is conditioned both on time and recursion scale, and trained with a loss function that aggregates direct supervision of velocities at each scale with strong cross-scale consistency constraints.

The primary trajectory represents the standard FM path, while secondary trajectories are time-rescaled and scaled in velocity. The consistency loss enforces that velocities at different scales $\hat{v}^{(i+1)}$ and $\alpha\hat{v}^{(i)}$ match at their shared spatial coordinates, analogous to multi-scale, pointwise data augmentation in the conditioning space. This recursive supervision is theoretically demonstrated to reduce trajectory curvature (acceleration), minimize truncation errors, and accelerate convergence, leading to high-quality generation with a minimal number of integration steps.

Figure 1: Flow matching (left) vs. recursive flow matching (right), illustrating cross-scale coupling and trajectory alignment.

Inference and Theoretical Analysis

During inference, RecFM employs either single-step or multi-step Euler integration over the learned velocity field. Cross-scale consistency regularizes the field to enable accurate mapping even with coarse discretization, addressing the bottleneck present in conventional flow matching and diffusion models.

Theoretical results guarantee that the marginal distribution of recursively constructed secondary trajectories matches that of the canonical linear interpolation path, and that the cross-scale consistency loss directly penalizes trajectory curvature, thereby tightening the global discretization error bound. This enables RecFM to perform well even in the few-step regime, a property unattainable for standard FM and most diffusion-based models.

Experimental Results

Benchmarks and Setup

RecFM is evaluated on three representative spatiotemporal datasets:

• Sea Surface Temperatures (SST): Real-world climate modeling, characterized by complex, coupled ocean-atmosphere dynamics.
• Navier-Stokes Flow: Incompressible turbulent flow with pronounced nonlinearity and nontrivial spatial correlations.
• Helmholtz Staircase Equation: Analytical acoustic scattering problem capturing sharp, periodic wave propagation.

For each dataset, RecFM is compared to stochastic, diffusion-based, and state-of-the-art video-based transformer solvers, including VideoPDE, MCVD, DDPM, and DYffusion, as well as standard and multi-fidelity flow matching baselines. Evaluation metrics include MSE (deterministic accuracy), CRPS (probabilistic calibration), and SSR (ensemble spread-skill).

Quantitative and Visual Results

RecFM delivers notably improved accuracy and efficiency: on all benchmarks, RecFM achieves strong numerical improvements (up to $15\%$ lower MSE) with only 1-2 inference steps, compared to 5-20 steps for vanilla FM and 50-100 steps for autoregressive or diffusion-based baselines. A critical result is the up-to-20× speedup over VideoPDE on the Helmholtz and Navier-Stokes tasks without sacrificing predictive quality.

Crucially, increasing step count beyond 1-2 in RecFM does not yield further improvements, and often incurs error accumulation, confirming that cross-scale trajectory alignment is a more effective regularizer than simply increasing inference resolution.

Figure 2: Roll-out results and error maps for the Helmholtz Staircase equation, comparing RecFM and VideoPDE at key timesteps.

Figure 3: MSE as a function of inference steps on Navier-Stokes; RecFM achieves best performance with one or two steps.

Figure 4: Validation MSE versus number of function evaluations (NFE), demonstrating RecFM's faster convergence relative to vanilla FM.

Further, RecFM maintains greater physical consistency over extended trajectories. For Navier-Stokes, RecFM more accurately preserves kinetic energy statistics and maintains lower PDE residuals compared to both classical flow matching and physics-constrained PBFM.

Figure 5: Sample Navier-Stokes rollout, illustrating spatiotemporal fidelity of RecFM predictions.

Figure 6: Average kinetic energy over time for different methods, with RecFM more closely tracking ground-truth dynamics.

Additional Ablations

Ablation studies show the optimal performance is robust to moderate changes in consistency loss weight, and increasing recursion depth beyond two yields diminishing returns. Even in configurations without explicit consistency loss, the secondary trajectory terms alone suffice to surpass vanilla FM, highlighting the power of multi-scale supervision. Furthermore, when compared to shortcut consistency models, RecFM achieves better results on physics-governed tasks for one-step inference.

Implications and Future Directions

This work advances the landscape of generative modeling for dynamical systems by demonstrating that explicit multi-scale, cross-trajectory alignment is more effective for efficient, high-fidelity forecasting than simply accelerating or distilling existing diffusion or flow-based models. The results challenge the notion that sampling efficiency and physical fidelity are necessarily at odds, showing that appropriate regularization of geometric properties can reduce both computational cost and error.

Practically, RecFM's architecture-agnostic design (e.g., integration with VideoPDE's DiT backbone) and minimal computational overhead make it directly applicable to a broad spectrum of scientific emulation applications, including climate modeling, fluid dynamics, and other real-time PDE-based tasks. The theoretical guarantees and empirical evidence support its role as a scalable foundation for future AI-driven scientific simulators.

However, extension of RecFM to high-dimensional, unconstrained visual domains (such as natural video) remains challenging. The framework's reliance on structured, physically motivated trajectory coupling may require adaptation to handle semantics-driven dynamics and heterogeneous temporal correlations.

Conclusion

Recursive Flow Matching establishes a new paradigm in the efficient generative modeling of scientific and spatiotemporal systems, unifying trajectory flow matching with recursive, cross-scale consistency constraints. The method achieves strong numerical and physical accuracy in the few-step regime, offering notable improvements in speed and robustness over prior approaches. It opens promising avenues for further advances in foundation models for scientific computing, physical simulation, and beyond.