Abstract
The Optimal Transport (OT) problem investigates a transport map that connects two distributions while minimizing a given cost function. Finding such a transport map has diverse applications in machine learning, such as generative modeling and image-to-image translation. In this paper, we introduce a scalable and simulation-free approach for solving the Entropic Unbalanced Optimal Transport (EUOT) problem. We derive the dynamical form of this EUOT problem, which is a generalization of the Schr"odinger bridges (SB) problem. Based on this, we derive dual formulation and optimality conditions of the EUOT problem from the stochastic optimal control interpretation. By leveraging these properties, we propose a simulation-free algorithm to solve EUOT, called Simulation-free EUOT (SF-EUOT). While existing SB models require expensive simulation costs during training and evaluation, our model achieves simulation-free training and one-step generation by utilizing the reciprocal property. Our model demonstrates significantly improved scalability in generative modeling and image-to-image translation tasks compared to previous SB methods.
1. Introduction
Optimal Transport (OT) provides a powerful mathematical framework for transforming one probability distribution into another, with wide-ranging applications in generative modeling and domain adaptation. However, standard OT faces two primary challenges: computational scalability in high dimensions and the high simulation costs associated with dynamic formulations like Schrödinger Bridges (SB). Traditional SB methods typically require simulating Stochastic Differential Equations (SDEs) during training and inference, leading to high Number of Function Evaluations (NFEs).
We present SF-EUOT (Simulation-free Entropic Unbalanced Optimal Transport), a novel algorithm designed to overcome these limitations. By leveraging the "reciprocal property" of Schrödinger Bridges, SF-EUOT enables simulation-free training and efficient one-step generation, achieving state-of-the-art performance on complex image datasets without the need for pre-trained models.
2. Methodology
2.1 Entropic Unbalanced Optimal Transport (EUOT)
Unlike standard OT, which requires strict marginal matching, EUOT utilizes $f$-divergences to allow for "soft matching." This unbalanced approach provides robustness against outliers and class imbalances, which is particularly beneficial when modeling the complex manifolds of natural images.
2.2 The Reciprocal Property and SF-EUOT
The core innovation of our approach is the utilization of the reciprocal property. In a Schrödinger Bridge, the path between two endpoints $x$ and $y$ can be analytically characterized as a Gaussian bridge. This allows the model to sample intermediate points $x_t$ directly from the transport path without iterative SDE or ODE solvers.
We propose a bilevel optimization framework where a generator network $T_\theta$ learns the static transport map and a value function $v_\phi$ learns the time-dependent cost. This formulation results in a simulation-free training process that is significantly faster than previous SB-based methods.
3. Experiments and Results
3.1 Generative Modeling on CIFAR-10
We evaluated SF-EUOT on the CIFAR-10 dataset to test its generative capabilities. As shown in Table 1, SF-EUOT achieves an FID of 3.02 with only 1 NFE. This performance is comparable to state-of-the-art methods like IPF (which requires pre-training and 200 NFEs) and significantly outperforms other SB methods that fail to converge from scratch (FID > 100).
3.2 Image-to-Image Translation
SF-EUOT was also tested on unpaired image-to-image translation tasks using the CelebA (Male $\to$ Female) and Wild $\to$ Cat datasets. In the Male $\to$ Female task, SF-EUOT achieved an FID of 8.44, outperforming established baselines such as CycleGAN (FID 12.94) and DiscoGAN.
4. Discussion and Conclusion
4.1 Key Insights
- Unbalanced Transport as a Regularizer: The relaxation of marginal constraints in EUOT acts as an optimization benefit, helping the model converge more effectively on high-dimensional image data.
- One-Step Generation Efficiency: SF-EUOT enables one-step generation (NFE=1), a significant improvement over diffusion models that typically require 50 to 2000 steps. Furthermore, the simulation-free training reduced training time to approximately 4 days on 2 RTX 3090Ti GPUs, compared to over 10 days for competing EOT algorithms.
4.2 Limitations and Future Work
While SF-EUOT demonstrates high efficiency, the PINN-style loss functions used can be sensitive to optimization hyperparameters. Future research will explore more stable loss formulations and scale the architecture to high-resolution datasets like CelebA-HQ (256x256).
In conclusion, SF-EUOT provides a scalable and efficient solution for high-dimensional transport problems, bridging the gap between the theoretical elegance of Schrödinger Bridges and the practical demands of modern generative modeling.
Summary updated