DBMSolver: A Training-free Diffusion Bridge Sampler for High-Quality Image-to-Image Translation
Researchers from Seoul National University have introduced DBMSolver, a novel training-free sampler designed to enhance the efficiency and quality of diffusion-based image-to-image (I2I) translation. While current state-of-the-art Diffusion Bridge Models (DBMs) offer high-fidelity generation, they are hindered by slow sampling speeds that often require dozens of function evaluations (NFEs). DBMSolver addresses this bottleneck by exploiting the semi-linear structure of the underlying Stochastic Differential Equations (SDE) and Ordinary Differential Equations (ODE) through exponential integrators. This approach yields highly efficient first- and second-order solutions, reducing NFEs by up to five times while simultaneously improving image quality. Experimental results demonstrate a significant 53% drop in Fréchet Inception Distance (FID) on the DIODE dataset at 20 NFEs compared to baseline methods. The model shows superior performance across various tasks, including inpainting, stylization, and semantics-to-image translation, establishing new state-of-the-art tradeoffs between efficiency and quality. This advancement enables greater real-world applicability for high-resolution image processing. The source code for DBMSolver has been made publicly available to support further research and development in computer vision.
Wire timeline
DBMSolver: A Training-free Diffusion Bridge Sampler for High-Quality Image-to-Image Translation
Researchers from Seoul National University have introduced DBMSolver, a novel training-free sampler designed to enhance the efficiency and quality of diffusion-based image-to-image (I2I) translation. While current state-of-the-art Diffusion Bridge Models (DBMs) offer high-fidelity generation, they are hindered by slow sampling speeds that often require dozens of function evaluations (NFEs). DBMSolver addresses this bottleneck by exploiting the semi-linear structure of the underlying Stochastic Differential Equations (SDE) and Ordinary Differential Equations (ODE) through exponential integrators. This approach yields highly efficient first- and second-order solutions, reducing NFEs by up to five times while simultaneously improving image quality. Experimental results demonstrate a significant 53% drop in Fréchet Inception Distance (FID) on the DIODE dataset at 20 NFEs compared to baseline methods. The model shows superior performance across various tasks, including inpainting, stylization, and semantics-to-image translation, establishing new state-of-the-art tradeoffs between efficiency and quality. This advancement enables greater real-world applicability for high-resolution image processing. The source code for DBMSolver has been made publicly available to support further research and development in computer vision.
cs.AI updates on arXiv.org