This paper addresses the problem of customization of an image at an object-level, where the aim is to enable users to customize images containing objects unseen during training without explicit supervision. Our approach utilizes a self-supervised method coupled with a semantic embedding space to facilitate object-level manipulation without the need for paired data. Through extensive experiments on various datasets, we demonstrate the effectiveness of our method in accurately customizing images with unseen objects while preserving visual coherence. Our results showcase the potential of targeted image manipulation in diverse applications such as content creation and image editing. In conclusion, this work presents a promising avenue for enabling users to seamlessly customize images containing unseen objects, opening up new possibilities for creative expression and content manipulation.