Efficient Generative Modeling beyond Memoryless Diffusion via Adjoint Schr"odinger Bridge Matching
Adjoint Schrodinger Bridge Matching (ASBM) is a generative framework that recovers optimal high-dimensional trajectories by decoupling the process into two stages. It first constructs an optimal coupling using a data-to-prior forward dynamic and then optimizes the backward dynamics with a simple matching loss supervised by this coupling. This results in stable, fast optimization and straighter trajectories than standard diffusion models.