Bidirectional Diffusion Bridge Models

• DOI: 10.48550/arxiv.2502.09655
• Publication Date: 2025-02-11
• Authors: Duc Kieu; Kien Do; Toan-Van Nguyen; Dang H. Nguyen; Thin Nguyen
• Keywords: Bridge (graph theory)

Diffusion bridges have shown potential in paired image-to-image (I2I) translation tasks. However, existing methods are limited by their unidirectional nature, requiring separate models for forward and reverse translations. This not only doubles the computational cost but also restricts practicality. In this work, we introduce Bidirectional Bridge Model (BDBM), a scalable approach that facilitates bidirectional between two coupled distributions using single network. BDBM leverages Chapman-Kolmogorov Equation bridges, enabling it to model data distribution shifts across timesteps both backward directions exploiting interchangeability of initial target within framework. Notably, when marginal given endpoints is Gaussian, BDBM's transition kernels possess analytical forms, allowing efficient learning with We demonstrate connection bridge methods, such as Doob's h-transform variational approaches, highlight its advantages. Extensive experiments on high-resolution I2I tasks enables minimal additional outperforms state-of-the-art models. Our source code available at [https://github.com/kvmduc/BDBM||https://github.com/kvmduc/BDBM].