Cross-domain Compositing with Pretrained Diffusion Models