Brief

Researchers have developed a new method called Coarse-to-Fine Compositional Diffusion (CoFi) to improve the generation of long-horizon outputs from diffusion models. CoFi separates the process into two stages: first, it forms a global structure by aligning local plans, and then it refines this structure with local details. This approach enhances both global coherence and local sample quality across various applications like robotic planning and video generation, while also reducing the number of required denoiser evaluations. AI