Nonlinear hierarchical editing: A powerful framework for face editing

Hierarchical Generative Adversarial Networks (GANs) have achieved considerable success in generating images, yet the task of editing these images in an interpretable, prominent, and disentangled manner remains a challenging issue. In this study, we introduce a novel hierarchical editing methodology that leverages nonlinear editing paths within GAN models. Nonlinear editing paths are identified in the GAN's latent space in an unsupervised manner, and attribute evaluators are employed to automatically discern the semantics associated with these paths. Subsequently, a layer -by -layer scoring technique is utilized to pinpoint the most pertinent layer for the editing path. The latent code navigates a nonlinear path reflective of a specific semantic, with modifications confined to layers most germane to the identified semantic. This hierarchical editing strategy results in significant, disentangled, and commutative editing outcomes. Compared to the current state-of-theart, our approach reduces side effect error by 20% to 39% in attribute disentanglement and commutativity error by 30% to 60% in continuous editing.