When diffusion models learn cinematography