<p>Image feature manipulation has long been a key focus of deep generative models, such as GANs. Although existing studies often use latent space to modify features, some challenges still exist. On the one hand, due to the influence of the training dataset, feature editing methods based on linear transformation often suffer from feature entangling and may cause identity information loss when editing multiple dimensions of latent code. On the other hand, feature editing approaches based on nonlinear transformations are confronted with challenges such as low training efficiency, limited modification scope, and poor interpretability. To address these issues, we propose a novel approach that builds upon linear transformation-based feature editing. Based on the geometric properties of high-dimensional space, we perform high-dimensional rotation on existing feature direction vectors (FDV) in latent space to achieve feature disentangling. Furthermore, to prevent identity loss caused by multi-dimensional editing of latent codes, we introduce a method to extract attribute-sensitive dimensions (ASD). This method can modify several dimensions of latent code in a targeted manner to achieve feature editing while retaining original identity information of the image as much as possible. These demonstrate the effectiveness of our approach and preserve the structure of generative model.</p>