AI image generators typically use diffusion to create realistic images, starting with a blurry picture and refining it through multiple steps. MIT researchers have developed a new approach called “distribution matching distillation” that streamlines this process down to a single step, significantly reducing the time needed to generate high-quality images. While other models have attempted similar acceleration methods with mixed results, MIT’s approach seems to strike a balance between speed and image quality. Stable Diffusion Turbo, another model that accelerates image generation, also shows promising results in a fraction of the time. Developers like those at MIT who share their AI technology openly are commended for their efforts, as other groups tend to keep their advancements closely guarded.
https://hothardware.com/news/mit-dmd-image-diffusion