Diffusion Models Learn Semantically Meaningful but Not Fully Factorized Representations

Diffusion models learn semantically meaningful representations through three distinct learning phases, but the representations are not fully factorized even under imbalanced datasets.