MIT researchers have developed a new framework called distribution matching distillation (DMD) that simplifies the process of generating images using diffusion models. The traditional multi-step process is condensed into a single step, allowing for faster generation without compromising image quality. The DMD framework combines regression and distribution matching losses while leveraging two diffusion models to guide the training process and improve stability.

The new approach achieves comparable results to more complex models, with potential applications in content creation, drug discovery, and 3D modeling. Further advancements in teacher models could enhance the quality of the generated images. The study will be presented at the Conference on Computer Vision and Pattern Recognition in June.