An Independent Semantic and Fused Latent Model for Local Face Editing

There is a strong correlation among semantic attributes in face editing models, editing one attribute may unintentionally alter other semantic attributes or affect unrelated regions. To enhance the user editing experience and achieve higher precision in facial detail editing, this paper proposes a face editing model based on semantic separation and feature fusion in the image domain, termed independent semantic and fused latent (ISFL). Firstly, facial semantics are disentangled using image masks and organized into a hierarchical tree structure. Next, ISFL enables both local separation and global fusion of image semantics, allowing users to independently edit the structure and appearance of specific semantic attributes through masks. Additionally, two methods, encoder-based and optimization-based, are employed to refine the details in generated images. Experimental results on CelebAMask-HQ dataset demonstrate that ISFL can produce more realistic and detail-rich images.