How compositional generalization and creativity improve as diffusion models are trained

Open in new window