BK-Editer: Body-Keeping Text-Conditioned Real Image Editing

With the firestorm of generative macromodelling, text-conditional image editing is a recently emerged and highly useful task with an unlimited future. Although a lot of research progress has been made, most of the methods still fail to achieve editing under body-shape preservation, i.e., they cannot generate results that conform to the semantics of the editing prompt while preserving the body-shape of the original image subject. To address this great challenge, we propose BK-Editer, a method that achieves satisfactory body-shape preservation and accomplishes editing under body shape preservation, which solves two major problems: 1) the edited result matches the corresponding editing prompt, and 2) the edited subject’s body shape is largely the same as the original subject’s body shape. In addition, our method does not require time-consuming training on a large-scale dataset and is a self-supervised method.