Beyond Optimal Transport: Model-Aligned Coupling for Flow Matching
Yexiong Lin, Yu Yao, Tongliang Liu · 2025
Read Beyond Optimal Transport: Model-Aligned Coupling for Flow Matching by Yexiong Lin, Yu Yao, Tongliang Liu (2025) 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-2505-23346.txt.
Introduction
Yexiong Lin, Yu Yao, and Tongliang Liu improve flow matching in a 2025 paper. Flow matching trains a model to carry samples from a source spread to a target spread. Random pairings of source and target points create crossing paths and bent trajectories that take many steps to use. Transport-based pairings cut crossings by minimizing distance, but distance alone may fight the model's own preferred paths.
Their method pairs points by both distance and how well a pairing fits the model's predicted directions. To save time, only the lowest-error fraction of pairings is kept for training. Tests showed clear gains in generation quality and speed when only a few steps are allowed.
Catalog record
| CREATOR | Yexiong Lin, Yu Yao, Tongliang Liu |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 41 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2505-23346
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Beyond Optimal Transport: Model-Aligned Coupling for Flow Matching Yexiong Lin, Yu Yao, Tongliang Liu Abstract Flow Matching (FM) is an effective framework for training a model to learn a vector field that transports samples from a source distribution to a target distribution. To train the model, early FM methods use random couplings, which often result in crossing paths and lead the model to learn non-straight trajectories that require many integration steps to generate high-quality samples. To address this, recent methods adopt Optimal Transport (OT) to construct couplings by minimizing geometric distances, which helps reduce path crossings. However, we observe that such geometry-based couplings do not necessarily align with the model’s preferred trajectories, making it difficult to learn the vector field induced by these couplings, which prevents the model from learning straight trajectories. Motivated by this, we propose Model-Aligned Coupling (MAC), an effective method that matches training couplings based not only on geometric distance but also on alignment with the model’s preferred transport directions based on its prediction error. To avoid the time-costly match process, MAC proposes to select the top- k k fraction of couplings with the lowest error for training. Extensive experiments show that MAC significantly improves generation quality and efficiency in few-step settings compared to existing methods. Project page: yexionglin.github.io/mac 1 Introduction Continuous-time generative models have obtained increasing attention for their ability to construct flexible mappings between simple source distributions and complex data distributions. Flow Matching (FM) [ 10 , 11 ] is a representative method in this paradigm. It trains a model to learn a time-dependent vector field, which is a function that specifies the direction and magnitude of movement for each point in space and time. By integrating this vector field over time, FM transforms samples from a source distribution (e.g., Gaussian noise) into samples from the target data distribution. This formulation provides a principled framework for modeling continuous transformation paths between distributions, without requiring stochastic simulation or likelihood-based training. A key step in FM is the construction of training data via couplings between source and target samples. These couplings define trajectories and provide supervision to guide the learning of the vector field that transports samples along the trajectories. However, the vanilla FM method constructs these couplings by randomly matching samples from the two distributions. Though it is simple, random couplings often result in chaotic transport directions and frequent path crossings. This leads the model to learn non-straight trajectories, which in turn require many integration steps during inference to generate high-quality samples [ 11 ] . (a) Random Coupling (b) Optimal Transport Coupling (c) Model-Aligned Coupling (MAC) (d) Flow Matching (1 step) (e) OT-FM (1 step) (f) Shortcut Model (1 step) (g) MAC (1 step) Figure 1: SubFig. (a–c) illustrate coupling strategies between a source distribution and a target distribution, defined respectively as a Gaussian mixture with four components and a Gaussian mixture with two components. Both follow the form p(x)=\sum_{i=1}^{K}\pi_{i}\mathcal{N}(x\mid\mu_{i},I) . We compare Random Coupling, Optimal Transport (OT), and our proposed Model-Aligned Coupling (MAC). While OT improves over Random by reducing path crossings, it may still induce local ambiguity (e.g., multiple directions at t=0 ). In contrast, MAC selects couplings that better align with the model’s learned vector field. SubFig. (c-d) show one-step generated samples for different models. MAC yields significantly better sample quality with fewer integration steps compared to standard FM, OT-FM, and Shortcut models. To improve the quality of training supervision, recent works adopt Optimal Transport (OT) to construct couplings that minimize geometric distances (e.g., L_{2} distance) between samples [ 24 , 14 ] . By explicitly reducing the total transport cost, OT-based couplings mitigate path crossings and offer more structured transport directions compared to random coupling. However, we found that relying solely on geometric distance to match couplings is not sufficient. This strategy ignores the model’s capacity, resulting in matched couplings that do not align with the learned vector field. As a result, the model cannot learn effectively, further degrading the quality of one-step generation. Fig. 1 (a–b) illustrates this issue on a synthetic task that transports samples from a Gaussian mixture with four components to a Gaussian mixture with two components. While random couplings (Fig. 1a ) produce many cross paths, OT-based couplings (Fig. 1b ) yield more organized directions. Nonetheless, OT can still introduce ambiguity. For example, in Fig. 1b , at time t=0 , the same source Gaussian is coupled to multiple target Gaussians. This results in inconsistent transport directions originating from the same spatial location. Since the model predicts similar velocity vectors at nearby points in space and time, it cannot simultaneously capture sharply differing directions. As a result, the model regresses toward an average of the conflicting directions after training, leading to non-straight and curved trajectories. These trajectories deviate from the ideal linear transport paths, making the generative process less efficient. As a consequence, the model requires more integration steps during inference to reach the target distribution. This inefficiency is reflected in the degraded quality of one-step generation, as shown in Fig. 1e . Motivated by the above insight, we propose Model-Aligned Coupling (MAC), a strategy that considers not only geometric distance but also the model’s capacity to learn the induced transport directions. To realize the idea, for each mini-batch, we can evaluate all possible source-target coupling and select the matching that minimizes the prediction error, as defined by the model’s current flow field. Since the model’s vector field evolves over time and prediction errors shift throughout training, MAC dynamically updates the selected couplings at each iteration to reflect the model’s current vector field. By aligning couplings with the model’s evolving representation, MAC promotes straighter trajectories and ultimately improves generative efficiency, particularly in one-step generation. As illustrated in Fig. 1c , MAC selects couplings that form more coherent and learnable transport paths. When these couplings are used to train a Flow Matching model, they lead to significantly improved one-step generation quality. Fig. 1 (d–g) compare one-step samples generated using different coupling strategies. Specifically, Fig. 1d shows results from training FM with random couplings, while Fig. 1e uses OT-based couplings. Fig. 1f presents outputs from the Shortcut model [ 3 ] , a state-of-the-art method for one-step generation. In contrast, Fig. 1g shows samples produced by FM trained with our proposed Model-Aligned Coupling (MAC). Across all cases, MAC yields higher-quality samples, producing more accurate and consistent generations in a single step. Notably, it outperforms both the OT-based method and the Shortcut model, demonstrating its effectiveness in bridging the gap between coupling supervision and model learnability. The key contributions of this paper are summarized as follows: • We introduce Model-Aligned Coupling (MAC), a simple yet effective strategy for constructing supervision couplings in Flow Matching. Unlike prior methods that rely solely on geometric distance (e.g., Optimal Transport), MAC selects couplings based on their learnability, measured by the model’s prediction error, to better align with the model’s current vector field. • To ensure computational efficiency, MAC avoids the costly computation of all pairwise errors and global assignment, typically requiring O(n^{2}\log n) to O(n^{3}) time depending on the matching algorithm (e.g., the Hungarian method), by adopting a lightweight sample-selection procedure. At each training step, it randomly samples candidate source-target couplings,