Editing capability decreases ineivitably in previous refinement methods, (e.g., PTI, HFGI, and SAM). In this work, we explore the idea of “divide and conquer” to address this problem. We combine two mainstream refinement mechanisms (i.e., weight ande feature modulation) and achieve extroadinary inversion and editing results.

We analyse the resources of “Fidelity, Perception, and Editability” in inversion task and point out that the keypoint is disalignment between inverse latent codes and synthetic distribution. We then propose a simple but efficient and uniform solution in both optimization-based and encoder-based methods.