Image composition with pre-trained diffusion models