Our method learns to control the layout of images generated by large pre-trained models without any training or fine-tuning.