Face attribute editing method based on dual-branch mapping architecture

Abstract

The rising demand for facial attribute editing has highlighted the limitations of existing text-driven generation methods. Although these methods can enhance feature diversity, they typically depend on large-scale annotated datasets, with varying text guidance necessitating separate optimization processes. Additionally, pre-trained Vision-Language Models (VLMs) have limitations in extracting text information and tuning hyperparameters. To tackle these challenges, this paper introduces a facial attribute editing method utilizing Dual-Branch Mapping (DBM). The core concept of this method is to achieve feature fusion between the source and target images in vector space. This method calculates the vector difference between the source and target images in the Contrastive Language-Image Pre-training (CLIP) space using both global and local branches. The difference is then fused into the Style Generative Adversarial Network (StyleGAN) vector of the source image, enabling precise control over attribute changes. During the inference stage, predictions and adjustments are made based on the correlation matrix in the StyleGAN space, following the change direction indicated by the CLIP text space. Experimental results indicate that this method enables text-free guided facial attribute editing, effectively overcoming the limitations of parameter tuning and prior data. Furthermore, it supports image generation under multiple text conditions without requiring additional training, significantly enhancing editing flexibility and naturalness.