Localizing Object-level Shape Variations with Text-to-Image Diffusion Models