Adding Conditional Control to Text-to-Image Diffusion Models