Why Your Diffusion Model Is Slow at Inference (And It's Not the UNet)

TL;DR: Most inference bottlenecks in diffusion pipelines are not in the UNet denoising loop. They are...