SDDBMs: A New Framework for Soft Denoising Diffusion Bridge Models

A new study on arXiv (2608.08594) introduces Soft Denoising Diffusion Bridge Models (SDDBMs), which take a more flexible approach to managing diffusion bridges and their endpoint limits. Traditional diffusion bridge models rely on Doob's h-transform to facilitate stochastic transitions between different endpoint distributions, proving effective in image translation and restoration. However, many such models require strict endpoint conditions, which can lead to issues like terminal-boundary singularities—where the terminal distribution becomes overly specific and problematic. SDDBMs, on the other hand, propose a more adaptable terminal distribution that doesn't force an exact endpoint, allowing for a Gaussian distribution that helps avoid these singularities. This framework could greatly impact machine learning and computer vision, particularly in image tasks.