New Diffusion Model Achieves Simulation-Free Training and Finite-Time Generation

Generative diffusion models rely heavily on the choice of a reference diffusion process, which connects the empirical data distribution to a prior distribution. Traditionally, developers have had to choose between models that allow for simulation-free training and those that enable finite-time generation, often sacrificing one for the other. This trade-off has been a significant hurdle in optimizing the efficiency and practicality of diffusion models. Researchers have introduced a new framework that overcomes this limitation, allowing for both simulation-free training and finite-time generation simultaneously. The core innovation lies in prescribing tractable time-dependent conditional distributions and then constructing a reference process that naturally realizes these as its marginals. This framework offers a fresh perspective on diffusion model training, demonstrating that score matching, a common technique, is not a fundamental requirement but rather a natural outcome of reversing the proposed reference process. Furthermore, the framework shows that conditional flow matching arises as a small-noise limit of this new approach, suggesting a unifying view for various generative modeling techniques.