Conditional Flow Matching (CFM) unifies conventional generative paradigms such as diffusion and flow-based models. Interaction Field Matching (IFM) is a recently proposed 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, which we call forward-only IFM. Specifically, we give explicit constructions in both directions and show that CFM and forward-only IFM induce the same family of data-space dynamics. General IFM, however, is strictly more expressive: already in EFM, its sampling rule yields transport maps that reverse orientation with positive probability, something no CFM flow can do. Our findings suggest using the probabilistic and geometric views jointly when developing generative models, and exploring field-based transport beyond the forward-only regime.
Figures & tables
Figure 1: Overview of the flow–field duality. We establish that CFM and forward-only IFM induce the same data-space generative dynamics (§ 3 ): CFM≡forward-onlyIFM⊊IFM . Panel (a) summarizes the resulting taxonomy. While conventional CFM (b, § 2.1 ) admits an equivalent forward-only field representation (c, § 2.2 ), general IFM is strictly more expressive: it includes EFM (d, § 2.2 ), which admits returning trajectories whose transport maps reverse orientation and therefore cannot be realized by any CFM flow (§ 4 ).
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
Aspect
Conditional Flow Matching
Interaction Field Matching
Elementary object
Conditional Flow (§ 2.1 ) ({{\mathbf{v}}_{t}^{\mathbf{x}_{0}{\color[rgb]{0.1211,0.4648,0.707},\mathbf{x}_{T}}}}(\mathbf{x}),\,{p_{t}^{\mathbf{x}_{0}{\color[rgb]{0.1211,0.4648,0.707},\mathbf{x}_{T}}}}(\mathbf{x})) defined in the data space D
Interaction Field ( 8 ) Ex0,xL(x~) defined in the extended space D+1
Global object
Conditional averaging of conditional velocities: \displaystyle\begin{aligned} \!\!\!\!{\mathbf{v}}_{t}(\mathbf{x})&=\!\int{{\mathbf{v}}_{t}^{\mathbf{x}_{0}{\color[rgb]{0.1211,0.4648,0.707},\mathbf{x}_{T}}}}(\mathbf{x})p_{t}({\mathbf{x}_{0}{\color[rgb]{0.1211,0.4648,0.707},\mathbf{x}_{T}}}\!\mid\!\mathbf{x}_{t}\!=\!\mathbf{x})d\mathbf{x}_{0}d{\color[rgb]{0.1211,0.4648,0.707}\mathbf{x}_{T}}\\ &\neq\!\int{{\mathbf{v}}_{t}^{\mathbf{x}_{0}{\color[rgb]{0.1211,0.4648,0.707},\mathbf{x}_{T}}}}(\mathbf{x})\,\pi_{0,T}({\mathbf{x}_{0}{\color[rgb]{0.1211,0.4648,0.707},\mathbf{x}_{T}}})\,d\mathbf{x}_{0}\,{\color[rgb]{0.1211,0.4648,0.707}d\mathbf{x}_{T}}\end{aligned}
Generalized superposition of pairwise fields: E(x~)=∫Ex0,xL(x~)π0,L(x0,xL)dx0dxL
Generative dynamics
ODE-based dynamics in the data space : dtdx=vt(x),xT∼πT
Field-line transport in the extended space : dτdx~=E(x~),x~∼πL×δz=L Forward-only fields dynamics in the data space : dzdx=E(x~)zE(x~)x,xL∼πL
Marginal distributions
Explicitly defined via conditional distributions: p_{t}(\mathbf{x})=\int{p_{t}^{\mathbf{x}_{0}{\color[rgb]{0.1211,0.4648,0.707},\mathbf{x}_{T}}}}(\mathbf{x})\,\pi_{0,T}({\mathbf{x}_{0}{\color[rgb]{0.1211,0.4648,0.707},\mathbf{x}_{T}}})\,d\mathbf{x}_{0}\,{\color[rgb]{0.1211,0.4648,0.707}d\mathbf{x}_{T}}
No explicit marginal density path: defined implicitly via field-line transport ( 10 )
Multi-sample global field regression: Ex~∼pvol∥fθ(x~)−E(x~)∥22 Normalized field in the forward-only case: Ex~∼pvolfθ(x~)−∥E(x~)∥2E(x~)22
Training volume
Explicitly defined by conditional distributions: {\mathbf{x}_{0}{\color[rgb]{0.1211,0.4648,0.707},\mathbf{x}_{T}}}\sim\pi_{0{\color[rgb]{0.1211,0.4648,0.707},T}},\ t\sim\mathcal{U}[0,T],\ \mathbf{x}_{t}\sim{p_{t}^{\mathbf{x}_{0}{\color[rgb]{0.1211,0.4648,0.707},\mathbf{x}_{T}}}}
Ad hoc heuristic distribution: x~∼pvol
Appendix
Table 1: CFM vs. IFM : side-by-side comparison of the Conditional Flow Matching and Interaction Field Matching frameworks.
Ex0,xT(x,t)x=vtx0,xT(x)ptx0,xT(x)Ex0,xT(x,t)t=ptx0,xT(x) OR originally defined as in § 2.2
Appendix
Table 2: Flow-field dual constructions . This table shows how a common framework can be formulated using both CFM and IFM.
Figure 2: Two-dimensional toy Gaussian dataset used for the empirical validation. Blue points denote samples from the source distribution πT , a narrow Gaussian centered near the origin. Orange points denote samples from the target distribution π0 , a symmetric four-component Gaussian mixture with modes arranged along the coordinate axes.
Type
Comparison
MSE ( ↓ )
RelMSE ( ↓ )
Matched
v1 vs. u1
2.05×10−7
1.08×10−6
Matched
v2 vs. u2
2.84×10−7
1.34×10−5
Mismatched
v1 vs. v2
4.93×10−1
0.34
Mismatched
u1 vs. u2
4.93×10−1
0.34
Appendix
Table 3: MSE and RelMSE between the CFM drift and the IFM-induced drift, for pairs built from the same interpolant (matched) and from different interpolants (mismatched). Here ui=Ei(x,t)x/Ei(x,t)t and RelMSE is defined in ( 41 ).
Figure 3: Probabilistic interpretation of a forward-only interaction field. Top: particles evolved under the dynamics induced by the trained IFM model. Bottom: heatmaps of the last component E1(x,t)t of the global field, estimated by Monte Carlo averaging the conditional fields over sampled endpoint pairs.
Table 4: Ablation of the volume-coverage distribution. For each field Ei , the normalized-field objective ( 13 ) is trained with pvol set to the matched path p_{i,t}^{{\mathbf{x}_{0}{\color[rgb]{0.1211,0.4648,0.707},\mathbf{x}_{T}}}} or to the mismatched path p_{j,t}^{{\mathbf{x}_{0}{\color[rgb]{0.1211,0.4648,0.707},\mathbf{x}_{T}}}} , j=i ; the last column reports the final W2 distance. Reported as mean ± standard deviation over 3 repeated evaluations of each trained model, with randomly resampled source subsets.
Figure 4: FID vs. training progress for the two-sided interpolant at different noise scales s .
N
FID ( ↓ )
Two-sided Linear
EDM
PFGM++ (d=128)
PFGM++ (d=2048)
1
3.03±0.03
2.29±0.04
2.40±0.04
2.40±0.05
256
3.04±0.03
2.21±0.03
2.24±0.03
2.14±0.04
2048
3.05±0.04
2.12±0.02
2.28±0.03
2.19±0.03
Appendix
Table 5: FID ( ↓ ) [ 26 ] on CIFAR-10 as a function of the number of samples N used for the multi-sample target. Best scores per column are bolded . Standard deviations are measured across 3 different sets of seeds for FID evaluation.
Figure 5: Training convergence on CIFAR-10 (FID vs. kimg) for different generative methods and multi-sample target sizes N . One kimg = one thousand images processed.
N
FID ( ↓ )
EDM
PFGM++ (d=128)
1
2.62±0.04
2.69±0.02
256
2.51±0.05
2.56±0.03
2048
2.63±0.04
2.61±0.03
Appendix
Table 6: FID ( ↓ ) at the final checkpoint on FFHQ-64 as a function of the multi-sample target size N . Best scores per column are bolded . Standard deviations are measured as in Table 5 .
Figure 6: Training convergence on FFHQ-64 (FID vs. kimg) for the single-sample target and multi-sample targets with N=256 and N=2048 .
Flow-based generative modeling learns probability transports from simple noise to data for fast sampling, density estimation, and controllable generation.
Flow Matching (FM) is a simulation-free method for learning a continuous, invertible flow that interpolates between two distributions, and in particular generates data from noise. Inspired by the variational nature of the diffusion process as a gradient flow, we introduce a stepwise FM model, Local Flow Matching (LFM), which sequentially learns a sequence of FM submodels, each matching a diffusion process up to the time-step size in the data-to-noise direction. In each step, the two distributions to be interpolated by the sub-flow model are closer than those in the full-flow matching model, which interpolates data to noise distributions, enabling smaller models with more efficient training. This variational perspective also allows us to prove a theoretical generation guarantee for the proposed flow model in terms of the χ2-divergence between the generated and true data distributions, leveraging the contraction property of the diffusion process. In practice, the stepwise structure of LFM is naturally amenable to model distillation, and various distillation techniques can be applied to accelerate generation. We empirically demonstrate that LFM achieves competitive generative performance compared to FM on unconditional generation of tabular and image datasets, and on conditional generation of robotic manipulation policies.
Chen Xu, Xiuyuan Cheng, Yao Xie
H. Milton Stewart School of Industrial and Systems Engineering, Georgia Institute of Technology. · Department of Mathematics, Duke University
Iterative generative models such as Flow Matching and Diffusion models have demonstrated strong test-time scaling behavior, where additional inference computation can improve generation quality. In contrast, Drift Models offer efficient one-step generation, but their direct generation paradigm limits such flexibility. In this work, we propose Drift Flow Matching (DFM), a framework that connects drifting generative modeling with flow-based iterative generation. DFM preserves the efficiency of direct transport maps while enabling generation to be refined through multiple inference steps when desired. This bridges the gap between one-step Drift Models and multi-step Flow Matching methods, and provides a novel generative paradigm that can adapt sampling computation to different quality--efficiency requirements. Extensive experiments across different tasks and datasets demonstrate the effectiveness and generality of the proposed framework.
Chenrui Ma, Xi Xiao, Lin Zhao +3
University of California, Irvine · University of Virginia · University of Alabama at Birmingham +1
We study Joint Flow Matching (JFM), a flow-based framework for learning a joint distribution over paired variables such as images and class labels. JFM transports the variables in opposite directions: an image is transported from noise to data while its label is transported from data to noise. This provides a single probabilistic model in which generation and classification are represented as complementary conditional inferences. We formulate the joint distribution induced by this construction and characterise the relationship between its forward and reverse conditionals. We investigate the resulting model for classification and conditional generation, including confidence calibration and consistency between generated samples and classifier predictions. Experiments on conditional datasets demonstrate competitive classification accuracy, well-calibrated confidence without post-hoc calibration, and classifier-consistent generation. We further explore the shared joint distribution as a basis for interpreting the relationship between discriminative and generative behaviour.
Hayden McAlister, Lech Szymanski
School of Computing, University of Otago, Dunedin, New Zealand