Unlocking the Duality between Flow and Field Matching

Abstract: Conditional Flow Matching (CFM) unifies conventional generative paradigms such as diffusion models and flow matching. Interaction Field Matching (IFM) is a newer framework that generalizes Electrostatic Field Matching (EFM) rooted in Poisson Flow Generative Models (PFGM). While both frameworks define generative dynamics, they start from different objects: CFM specifies a conditional probability path in data space, whereas IFM specifies a physics-inspired interaction field in an augmented data space. This raises a basic question: are CFM and IFM genuinely different, or are they two descriptions of the same underlying dynamics? We show that they coincide for a natural subclass of IFM that we call forward-only IFM. Specifically, we construct a bijection between CFM and forward-only IFM. We further show that general IFM is strictly more expressive: it includes EFM and other interaction fields that cannot be realized within the standard CFM formulation. Finally, we highlight how this duality can benefit both frameworks: it provides a probabilistic interpretation of forward-only IFM and yields novel, IFM-driven techniques for CFM.

• The paper establishes the formal duality between CFM and forward-only IFM, unifying probabilistic and physics-inspired generative approaches.
• It provides explicit mappings that show equivalence of dynamics while highlighting IFM's broader expressiveness over CFM.
• The work introduces a multi-sample CFM velocity estimator that improves performance metrics and offers insights into optimal volume coverage.

Summary of "Unlocking the Duality between Flow and Field Matching" (2602.02261)

This paper rigorously analyzes the relationship between Conditional Flow Matching (CFM) and the recently proposed Interaction Field Matching (IFM) frameworks for generative modeling. The authors show that, despite originating from distinct modeling principles—probabilistic path-based (CFM) versus geometric, physics-inspired field-based (IFM)—these frameworks share a precise formal duality. They provide explicit mappings that establish the equivalence of CFM and a restricted, forward-only class of IFM, while also proving that general IFM is strictly more expressive than CFM. The work additionally yields practical advances, including improved estimators for learning generative dynamics and insights into volume coverage distributions.

Theoretical Frameworks: CFM and IFM

CFM constructs generative ODEs by specifying a conditional probability path and compatible velocity field in data space, marginalizing over latent pairs or paths to yield the full generative drift. This framework encompasses both stochastic interpolant methods and denoising diffusion models, as in (Albergo et al., 2023) and (Song et al., 2020). The CFM objective is single-sample regression on conditional velocities.

By contrast, IFM formulates generative transport in an augmented space, typically $\mathbb{R}^{D+1}$, where interaction fields—vector fields satisfying connection, divergence-free, and flux constraints—are defined between sample pairs or between a sample and prior hyperplane. The induced global field is computed via superposition, and data is transported along its field lines. Field-based models, including Electrostatic Field Matching (EFM) (Kolesov et al., 4 Feb 2025) and Poisson Flow Generative Models (PFGM) (Zhang et al., 2022), fit this paradigm.

The expressive power of IFM arises from its capacity to realize arbitrary field line patterns, including (for EFM) backward-oriented lines, whereas CFM is constrained to 'forward-only' dynamics due to its time-parameterized ODE construction.

Main Results: Duality, Expressiveness, and Mappings

The authors prove a bidirectional equivalence between CFM and forward-only IFM. For any conditional flow in CFM, they construct an explicit forward-only interaction field whose induced generative dynamics (via field lines) matches the CFM ODE, identifying time in CFM with the auxiliary dimension ('z') in IFM. Conversely, given any forward-only interaction field, they extract a corresponding CFM conditional flow and velocity.

• Theorem 3.1: Any CFM generative process admits a forward-only IFM representation via a constructive mapping. Specifically, for conditional velocity $v_{t, x_0, x_T}(x)$ and path density $p_{t, x_0, x_T}(x)$, the associated interaction field is $$(v_{t, x_0, x_T}(x)p_{t, x_0, x_T}(x),\ p_{t, x_0, x_T}(x))$$ in $(x, t)$ space. The global field is the marginal of these objects over the joint endpoint distribution.
• Theorem 3.3: Any forward-only IFM field can be decomposed to yield a conditional velocity and path density satisfying the CFM requirements, restoring the original ODE-based dynamics.
• Any generative transport achievable by CFM is thus representable via forward-only IFM, but not vice versa: backward-oriented IFM fields induce dynamics inaccessible to standard CFM.

This duality is made explicit through the use of the superposition principle (for field aggregation) and the construction of marginal distributions and flows from local conditional objects.

Conceptual and Practical Implications

The duality unlocks several key insights:

• Probabilistic Interpretation of Forward-Only IFM: Via the CFM connection, forward-only IFM inherits a natural interpretation as evolving marginal distributions, facilitating probabilistic analysis and integration with standard generative modeling pipelines.
• Expressiveness: General IFM encompasses a broader family of generative dynamics; specifically, EFM and PFGM admit field lines and transports not realizable by any CFM construction, highlighting modeling advantages for certain applications.
• Unified Framework: Existing algorithms, such as stochastic interpolants, denoising diffusion, flow matching, and various field-matching methods, admit dual representations, suggesting a unified approach capable of leveraging both perspectives.

Multi-Sample Velocity Estimation and Volume Coverage

A significant practical contribution of the paper is the introduction of a multi-sample CFM velocity estimator, derived from the IFM global field formulas. The estimator combines velocities from multiple conditional pairs, weighted by their relevance, and is shown to provide performance benefits for certain models (notably EDM and PFGM++, with lower FID scores on CIFAR-10 for $N > 1$ samples).

Analysis reveals, however, that for two-sided stochastic interpolants, a single pair typically dominates the estimate, indicating redundancy in multi-sample targets for these specific models. The weight distribution of the estimator—measured via the Gini coefficient—demonstrates concentration for two-sided frameworks, but increasing dispersion for one-sided models as $N$ grows.

Additionally, the IFM viewpoint motivates exploration of alternative volume-coverage distributions for training, though the optimal choice remains unresolved and presents a research direction for reducing estimator variance and improving coverage of error-prone regions.

Limitations and Extensibility

• CFM Limitation: The inability of CFM to express backward field lines can limit the distributional transports realizable within this paradigm.
• Heuristic Training in IFM: IFM frameworks typically require ad hoc volume coverage distributions for practical field regression, which may hinder efficiency and optimality.

Extensions beyond the forward-only setting may enable new classes of invertible, non-monotonic generative models. Relaxing or generalizing the forward-only constraint could result in novel generative architectures with improved expressiveness and controllability.

Conclusion

This work settles the formal relationship between flow-based (CFM) and field-based (IFM) generative modeling frameworks, establishing their equivalence in the forward-only regime and delineating the additional expressive power of general IFM. Explicit constructive mappings provide a unified theoretical foundation and suggest practical algorithmic improvements, including multi-sample estimators and more flexible training distributions. These findings have direct implications for the design of future generative models, especially those exploiting physics-inspired principles, and illuminate promising directions for research at the intersection of flow, field, and probabilistic generative modeling.