Diffusion models reveal the hierarchical and combinatorial nature of data.The backward diffusion process exhibits a phase transition at a threshold time.Reconstruction of high-level features drops suddenly after the transition.Generative models can effectively model combinatorial data properties.