Face Attribute Editing Algorithm Based on Multimodal Attention

• 1、Institute of Artificial Intelligence, Beijing University of Posts and Telecommunications, Beijing 100876

Abstract：To address the conflict between global semantic control and local precise adjustment in facial attribute editing tasks, this paper proposes a facial attribute editing algorithm based on a multimodal attention mechanism. By introducing gradient-weighted class activation mapping (Grad-CAM) to dynamically associate textual semantics with generator feature maps, adaptive localization of editing regions is achieved. A multi-objective loss function is designed to jointly optimize image-text semantic alignment loss, identity preservation loss, and local attention constraint loss, balancing the contradiction between global semantic control and local detail adjustment. Experiments demonstrate that the proposed method enables precise adjustment of local attributes while maintaining unchanged non-target regions. On the Yellow race face dataset, the algorithm outperforms baseline models in terms of image-text semantic similarity, identity retention rate, and learned perceptual image patch similarity (LPIPS), validating its effectiveness and robustness in fine-grained facial feature editing.

Keywords： Artificial Intelligence； Facial Attribute Editing； Multimodal； Attention