DBMSolver: Fast Diffusion Bridge Sampling for High-Quality Image-to-Image Translation

Sankarshana Venugopal, Mohammad Mostafavi, Jonghyun Choi

Keywords: Diffusion Bridge Models, Diffusion Models, Image-to-Image Translation

TL;DR: We propose a novel, training-free solver for Diffusion Bridge Models, making Image-to-Image Translation better and faster.

Abstract:

Diffusion-based approaches for image-to-image (I2I) translation have garnered significant attention due to their ability to generate high-fidelity images and scalability to large-scale datasets.
However, state-of-the-art Diffusion Bridge Models (DBMs), which utilize diffusion bridges to interpolate between two images $\mathbf{x}_0$ and $\mathbf{x}_T$, are severely hampered by their slow sampling process, often requiring dozens to hundreds of function evaluations.
To address this computational burden, we introduce DBMSolver, a novel, training-free sampler specifically designed for DBMs.
DBMSolver leverages the inherent semi-linear structure of the underlying diffusion equations in DBMs and employs advanced exponential integrators to accelerate the sampling process. This approach not only reduces the number of evaluations but also enhances image quality for I2I Translation tasks.
Our experiments demonstrate that DBMSolver outperforms prior methods across multiple datasets and resolutions, significantly improving visual quality and reducing computational overhead.
DBMSolver improves the scalability of diffusion-based I2I Translation by bridging the gap between theoretical elegance and real-world applicability.

Primary Area: generative models

Submission Number: 1924