We characterize the distribution of the generated samples for a solvable model of diffusion model in high dimensions, parametrized by a two-layer autoencoder.