DBMSolver: Training-Free Diffusion Bridge Sampler Speeds Up I2I Translation

Researchers introduce DBMSolver, a training-free sampler that accelerates diffusion-based image-to-image (I2I) translation by exploiting the semi-linear structure of Diffusion Bridge Models (DBMs). By using exponential integrators for 1st- and 2nd-order solutions, DBMSolver reduces the number of function evaluations (NFEs) by up to 5x while improving output quality—for example, FID drops 53% on DIODE at 20 NFEs compared to a 2nd-order baseline. The method achieves state-of-the-art efficiency-quality tradeoffs on tasks including inpainting, stylization, and semantics-to-image generation at resolutions up to 256x256. The code is publicly available on GitHub.