Abstract
Schrödinger Bridge (SB) is an entropy-regularized optimal transport problem that has received increasing attention in deep generative modeling for its mathematical flexibility compared to the Score-based Generative Model (SGM). However, it remains unclear whether the optimization principle of SB relates to the modern training of deep generative models, which often rely on constructing log-likelihood objectives. This raises questions on the suitability of SB models as a principled alternative for generative applications. In this work, we present a novel computational framework for likelihood training of SB models grounded on Forward-Backward Stochastic Differential Equations Theory—a mathematical methodology that appeared in stochastic optimal control that transforms the optimality condition of SB into a set of SDEs. Crucially, these SDEs can be used to construct the likelihood objectives for SB that, surprisingly, generalizes the ones for SGM as special cases. This leads to a new optimization principle that inherits the same SB optimality yet without losing applications of modern generative training techniques, and we show that the resulting training algorithm achieves comparable results on generating realistic images on MNIST, CelebA, and CIFAR10.
Introduction: Beyond Fixed Diffusion
Generative modeling has seen a massive surge in popularity, with Score-based Generative Models (SGMs) and Diffusion Models leading the charge. While these models produce stunning results, they often operate as a "one-way street": a fixed diffusion process turns data into noise, and the model learns to reverse it. This introduces two main rigidities:
- Fixed Diffusion: The process of diffusing data into noise (usually a simple linear drift) must be pre-specified. A poorly designed process can lead to training instability.
- Slow Sampling: To ensure the final distribution is approximately Gaussian, these models often require thousands of steps to generate a single image.
The Schrödinger Bridge (SB) problem offers a more flexible alternative. It views generative modeling as an entropy-regularized optimal transport problem—finding the most efficient way to transform one distribution into another over a finite time horizon. Unlike SGMs, SB models can learn the optimal diffusion path between data and noise.
The SB-FBSDE Framework
The core innovation of this work is the application of Forward-Backward SDE (FBSDE) theory to the SB problem. This methodology transforms the optimality conditions of SB, typically described by coupled Partial Differential Equations (PDEs), into a set of coupled SDEs.
By solving these SDEs, the authors derive an exact log-likelihood expression for the SB model. This framework shows that the standard SGM objective is a special case of the SB objective where the forward diffusion is fixed. The proposed training schemes include:
- Joint Training: Optimizing both forward (data-to-noise) and backward (noise-to-data) policies simultaneously.
- Alternate Training: A memory-efficient version resembling the classical Iterative Proportional Fitting (IPF) algorithm but utilizing a new divergence-based likelihood objective.
Key Insights and Results
The SB-FBSDE framework treats SB as a "learnable" SGM. Because the model learns the forward path, it can find a shorter, more efficient route to the noise distribution. This results in significant performance gains:
- Sampling Speedup: SB-FBSDE achieved a ~80% reduction in sampling time compared to standard SGMs while maintaining comparable image quality, generating images in 100-200 steps rather than 1000+.
- Quantitative Performance: On CIFAR10, the model achieved a competitive FID score of 3.01 and a negative log-likelihood (NLL) of 2.96 bits/dim, outperforming other optimal transport-based baselines.
- Scalability: The model successfully generated realistic, high-fidelity images across MNIST, CelebA, and CIFAR10, proving the framework scales effectively to high-dimensional data.
Limitations and Future Work
Despite its advantages, the framework requires maintaining two separate neural networks, leading to roughly 2 to 2.5 times more memory usage during training compared to standard SGMs. Additionally, the practical implementation relies on numerical SDE solvers, making the use of a Langevin corrector vital to mitigate discretization errors. Future research may explore more complex, non-Gaussian prior distributions and further analyze the convergence of alternate training schemes.
Code availability: https://github.com/ghliu/SB-FBSDE
Summary updated