@pearl @gildilinie iirc the gist of diffusion models is (very incorrect but still kinda correct if i remember math4datascience classes):
you have like a set size canvas of noise, then you somethings that kinda act like like something that overlays that noise, imagine a prompt "bird with tophat" you'd have a bird filter thing that denoises the image into being bird like, and then a tophat denoise somewhere and probably also a "woodland forest background" denoise thing