Coupled Flow Matching

Wenxi Cai, Yuheng Wang, Naichen ShiarXiv 2025 · DOI · arXiv

Abstract

Coupled Flow Matching (CPFM) is a novel framework designed to unify dimensionality reduction with high-fidelity generative modeling. By integrating a kernelized optimal transport objective with bidirectional continuous flows, CPFM enables explicit control over semantic feature preservation in low-dimensional embeddings. Unlike traditional lossy compression methods, CPFM ensures that discarded information is recoverable, allowing for high-quality reconstruction from extremely low-dimensional latent spaces.

Introduction: The Compression-Reconstruction Trade-off

Machine learning practitioners often grapple with the tension between data compression and reconstruction fidelity. Classical dimensionality reduction (DR) techniques, such as PCA or t-SNE, provide intuitive visualizations but are inherently lossy; once data is projected into a 2D or 3D space, fine-grained details are irrecoverably lost. Conversely, generative models like Variational Autoencoders (VAEs) or Generative Adversarial Networks (GANs) excel at reconstruction but often rely on "black box" latent spaces where the geometry is an implicit, often uninterpretable, byproduct of training.

Coupled Flow Matching addresses this by introducing controllability. The framework allows users to explicitly define the semantic structure of the embedding space—for instance, requiring that specific labels cluster together—while delegating the storage of "nuisance" details to the weights of a flow network for later reconstruction.

Methodology: A Two-Stage Framework

CPFM achieves this controllable mapping through a two-stage process involving optimal transport and conditional flow matching.

Overview of the CPFM framework, showing the transition from Generalized Gromov-Wasserstein Optimal Transport to Dual-Conditional Flow Matching.

Stage 1: Generalized Gromov-Wasserstein Optimal Transport (GWOT)

The first stage establishes a probabilistic bridge between the high-dimensional data space ($\mathcal{X}$) and the target low-dimensional embedding space ($\mathcal{Y}$). The authors introduce a kernelized cost function that incorporates prior knowledge. For example, a label-sensitive kernel can be used to ensure that images of the same class are mapped to nearby regions in the latent space. To overcome the $O(n^3)$ complexity typical of quadratic optimal transport, the authors developed a variational reformulation that scales at $O(n^2)$, making the approach viable for large-scale datasets.

Stage 2: Dual-Conditional Flow Matching (DCFM)

With the optimal transport plan established, the model learns the bidirectional mapping. CPFM utilizes a single "shared drift" neural network, $u_\theta$, to perform both encoding ($p(y|x)$) and decoding ($p(x|y)$). By employing a binary "role flag," the network learns both directions simultaneously. This symmetry acts as a regularizer, ensuring mathematical consistency between the encoder and decoder while significantly reducing parameter overhead.

Key Insights: Controllability and Information Preservation

The core innovation of CPFM lies in how it handles information. While standard DR methods discard residual information, CPFM stores it within the vector fields of the flow network. This allows for high-fidelity reconstruction even from a 2D embedding.

Visualization of how different kernel functions in the GWOT stage allow users to "steer" the geometry of the latent space.

Furthermore, the kernel function acts as a "steering wheel" for the latent space. By adjusting the kernel, users can prioritize visual similarity, class labels, or physical properties (such as the dipole moment in molecular data), fundamentally altering what the embedding represents without sacrificing reconstruction quality.

Experimental Results

The authors evaluated CPFM on standard image benchmarks (MNIST, CIFAR-10, AFHQ) and scientific datasets (QM9).

  • Embedding Quality: CPFM achieved the lowest Wasserstein distance to target distributions compared to VAEs and Info-Diffusion, indicating a more precisely structured latent space.
  • Reconstruction Fidelity: Even at extreme compression levels (e.g., 2D embeddings for MNIST), CPFM achieved an FID score of 0.113, vastly outperforming VAEs (32.0) and DiffAE (4.96).
  • Semantic Clustering: Visualizations confirm that CPFM successfully clusters data by user-defined labels while maintaining smooth intra-cluster gradients for "style" or other secondary features.

Conclusion and Future Work

Coupled Flow Matching represents a significant step toward interpretable and controllable generative models. While the current implementation is computationally intensive—particularly the flow matching stage for high-resolution images—the $O(n^2)$ OT solver provides a scalable foundation. Future research will focus on multi-scale architectures to handle higher resolutions and exploring more complex target submanifolds beyond 2D planes.

Summary updated

BibTeX
@misc{cai2025coupled,
  author = {Wenxi Cai and Yuheng Wang and Naichen Shi},
  title = {{Coupled Flow Matching}},
  year = {2025},
  eprint = {2510.23015},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2510.23015},
}
Figure