Efficient Transferable Optimal Transport via Min-Sliced Transport Plans

A new framework for optimal transport could lower the computational cost of matching data distributions across computer vision tasks. Researchers propose a method that allows a transport plan learned on one pair of distributions to transfer efficiently to new, unseen pairs, reducing the need for repeated computation [1][2]. Optimal Transport (OT) is a mathematical framework used to find correspondences between probability distributions, with applications in shape analysis, image generation, and multimodal tasks [1][2]. Its adoption has been limited by high computational demands. Slice-based methods address this by reducing the problem to one-dimensional OT, which has a closed-form solution, but they have not previously addressed whether a learned projection can be reused when the underlying data changes [2]. The work, led by Xinran Liu and submitted to arXiv, introduces the min-Sliced Transport Plan (min-STP) framework [1]. The central question is whether an optimized slicer—a one-dimensional projection trained on a specific distribution pair—can generate effective transport plans for new pairs under distributional shift [2]. The authors provide a theoretical result showing that optimized slicers remain close under slight perturbations of the data distributions, which enables efficient transfer across related tasks [1][2]. To further improve scalability, the paper introduces a minibatch formulation of min-STP and provides statistical guarantees on its accuracy [2]. The empirical results demonstrate that the transferable min-STP achieves strong one-shot matching performance and facilitates amortized training for point cloud alignment and flow-based generative modeling [1][2]. Flow-based generative models and diffusion models represent a class of techniques where data is generated by learning to reverse a noising process, and they are widely used in computer vision for tasks such as image generation and denoising [3]. The paper was first submitted on November 24, 2025, with a file size of 12,138 KB, and was revised most recently on May 26, 2026, with a file size of 15,558 KB [1]. The research addresses a practical bottleneck in settings with evolving data or repeated OT computations across closely related distributions [2].