Diffusion models make great images but are slow because they fix noise step by step many times.
LitePT is a new AI backbone for 3D point clouds that uses convolutions in early layers and attention in later layers to be both fast and accurate.