Flow Matching with Arbitrary Auxiliary Paths

Xin Peng, Ang Gao · 2026

Read Flow Matching with Arbitrary Auxiliary Paths by Xin Peng, Ang Gao (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-06364.txt.

Introduction

Xin Peng and Ang Gao generalize flow matching by letting its auxiliary variable follow any distribution at all. Flow matching trains generative models along probability paths, and prior work restricted the auxiliary part of those paths to Gaussian noise. Their framework, AuxPath-FM, draws the auxiliary variable from a chosen source, building trajectories from scaled data, scaled noise, and a weighted auxiliary term whose shape the designer now controls freely.

Theory shows the construction preserves the continuity equation and keeps the training objective consistent with the marginal form. Builders can pick Gaussian, uniform, Laplace, or even discrete coin flip style priors, each giving the flow different geometry. Structured semantic information can also ride in the auxiliary distribution for label guided generation. The authors present the framework as a general base for probability path design.

Catalog record

CREATORXin Peng, Ang Gao
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE46 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.06364

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Flow Matching with Arbitrary Auxiliary Paths Xin Peng, Ang Gao Abstract We introduce a new generative modeling framework, Flow Matching with Arbitrary Auxiliary Paths (AuxPath-FM) , which generalizes conditional flow matching by incorporating an auxiliary variable drawn from an arbitrary distribution into the probability path. Unlike prior methods that restrict auxiliary components to Gaussian noise, AuxPath-FM allows the variable η \eta to follow any distribution, producing trajectories of the form X t = a ⁡ ( t ) ​ X 1 + b ⁡ ( t ) ​ X 0 + c ⁡ ( t ) ​ η X_{t}=a(t)X_{1}+b(t)X_{0}+c(t)\eta . We theoretically demonstrate that this construction preserves the continuity equation and maintains a training objective consistent with the marginal formulation. This flexibility enables the design of diverse probability paths using various priors, including Gaussian, Uniform, Laplace, and discrete Rademacher distributions, each offering unique geometric properties for generative flows. Furthermore, our framework allows for specialized tasks such as label-guided generation by encoding structured semantic information into the auxiliary distribution. Overall, AuxPath-FM provides a principled and general foundation for probability path design, offering both theoretical generality and practical flexibility for diverse generative modeling tasks. 1 Introduction Continuous-time generative models, such as score-based diffusion models [ 38 , 18 ] and flow matching frameworks [ 27 , 29 , 43 ] , have emerged as powerful tools for modeling complex data distributions by learning dynamics that transport a simple base distribution to a target distribution. In conditional generative modeling, incorporating label or side information is crucial for controllable sample generation. However, most existing approaches introduce such information only through conditioning neural networks [ 19 , 12 ] , rather than modifying the underlying probability paths. While recent advances have expanded flow matching to multi-scale representations [ 47 ] , physical systems with hard constraints [ 44 ] , and complex temporal forecasting [ 33 ] , these models still primarily rely on network-side conditioning. Furthermore, recent theoretical analysis suggests that flow matching is naturally adaptive to the intrinsic manifold structure of data [ 25 ] , yet the influence of the path’s stochastic component remains under-explored in the context of trajectory-level control. Meanwhile, prior path-based formulations, including stochastic interpolants [ 1 ] and Schrödinger Bridge methods [ 6 , 42 , 21 ] , typically rely on Gaussian noise or Brownian dynamics for stochastic interpolation, which limits the flexibility of designing probability paths for structured guidance. Although recent research has introduced discrete state-space flows via edit operations [ 15 , 30 ] and geometric flows on statistical manifolds [ 10 ] , these methods often remain specialized to particular data types. Modern alignment techniques, such as policy gradient methods for flow matching [ 28 ] , further highlight the need for more expressive generative trajectories that can be steered via diverse signals. In this work, we introduce Flow Matching with Arbitrary Auxiliary Paths (AuxPath-FM) , a general framework that extends flow matching by incorporating an auxiliary random variable into the probability path. Specifically, we construct trajectories of the form X_{t}=a(t)X_{1}+b(t)X_{0}+c(t)\eta, (1) where X_{0}\sim p_{0} , X_{1}\sim p_{1} , and \eta is an auxiliary random variable drawn from an arbitrary distribution. Unlike prior interpolation schemes that restrict auxiliary terms to Gaussian noise [ 1 , 6 ] , our formulation allows flexible auxiliary designs such as condition-dependent sources [ 22 ] or structured transition kernels [ 37 ] , enabling the auxiliary variable to directly encode structured signals in the generative trajectory. Figure 1 : Generative trajectories of AuxPath-FM with diverse auxiliary distributions. The figure illustrates the interpolation paths X_{t}=a(t)X_{1}+b(t)X_{0}+c(t)\eta between two data samples X_{0} and X_{1} . By varying the distribution of the auxiliary variable \eta , we observe unique geometric properties: (a) Baseline ( \eta=0 ), (b) Gaussian \eta\sim\mathcal{N}(0,I) , (c) Uniform \eta\sim\mathcal{U}([-1,1]^{d}) , (d) Laplace \eta\sim\text{Laplace}(0,1) , (e) Rademacher \eta\in\{-1,1\}^{d} , and (f) Label-Guided \eta=F_{\phi}(Y) . Note how discrete and heavy-tailed priors impart distinct structural biases to the flow. We show that this construction preserves the continuity equation and yields a flow matching objective consistent with the marginal formulation [ 27 , 1 ] . A particularly important instance arises when \eta encodes semantic information such as class labels, enabling label-guided generation through the probability path itself. In contrast to conventional conditioning approaches and classifier-free guidance [ 19 , 12 ] , which operate through network-level modulation, our method embeds guidance directly into the trajectory design, leading to principled control of the generative dynamics. Overall, AuxPath-FM generalizes existing path-based frameworks such as stochastic interpolants and Schrödinger Bridge methods [ 1 , 6 , 42 ] , while retaining the standard flow matching training objective, thereby providing a unified and flexible perspective for structured continuous-time generative modeling. 2 Related Work Diffusion and Flow Matching. Continuous-time generative models have achieved remarkable success in modeling complex data distributions [ 38 , 18 , 23 ] . Score-based diffusion models learn the gradient of the log-density and simulate reverse-time stochastic differential equations to transform noise into data samples [ 38 ] . More recently, flow matching (FM) provides an alternative framework that directly learns the velocity field of a deterministic probability flow connecting a base distribution and a data distribution [ 27 , 29 ] . Instead of simulating stochastic processes, FM trains a neural network to match the velocity of a predefined probability path, enabling efficient training and flexible trajectory design [ 27 , 43 ] . Recent theoretical advances have further established that flow matching is naturally adaptive to manifold structures [ 25 ] , and generalizations such as Transition Matching [ 37 ] have unified flow models with discrete-time Markovian transitions to improve sampling efficiency and quality. Schrödinger Bridges and Stochastic Interpolants. Several works study generative modeling through stochastic interpolation between distributions [ 6 , 42 , 21 , 13 ] . Schrödinger Bridge methods formulate generative modeling as an entropy-regularized optimal transport problem, producing stochastic trajectories between endpoints [ 6 , 42 , 4 ] . More recently, stochastic interpolants provide a unifying framework connecting diffusion and flow-based methods [ 1 , 45 ] . In many such formulations, interpolation paths typically introduce auxiliary noise variables that follow Gaussian distributions [ 1 , 6 ] . However, recent extensions have begun to explore flow matching on non-Euclidean geometries, such as Riemannian manifolds [ 8 ] and periodic lattices for material discovery [ 31 ] , where the geometry of the space dictates the design of the interpolation path beyond simple Euclidean Gaussianity. Alternative Forward Transformations and Non-Gaussian Paths. Some works explore relaxing the standard Gaussian noise assumption. For example, Cold Diffusion replaces Gaussian noise with deterministic image transformations [ 3 ] . Other works investigate generalizing diffusion beyond Gaussian perturbations via structured noise models [ 38 , 23 ] . Notably, recent developments in discrete state-spaces have introduced Dirichlet Flow Matching [ 39 ] and Discrete Flow Matching [ 14 , 7 ] , which utilize probability paths on simplices or categorical spaces to handle discrete sequential data. These studies suggest that generative modeling can be viewed as learning to reverse a sequence of transformations. Our work further generalizes this perspective by allowing the auxiliary path components to follow any arbitrary

Read the complete text