Facial attribute editing via a Balanced Simple Attention Generative Adversarial Network

doi:10.1016/j.eswa.2025.127245

Facial attribute editing via a Balanced Simple Attention Generative Adversarial Network

2025-06-01

0

PRE

AI

DOI：10.1016/j.eswa.2025.127245

Original

Original request for help

Save

Abstract

En 中文

By adjusting specific attributes, facial attribute editing seeks to alter facial images. The generative adversarial network (GAN) is one of the most successful approaches to fulfill this task. But existing GANs face challenges in maintaining the integrity of other features in the original facial images. Additionally, they tend to underperform in terms of diversified editing capabilities and detail preservation. To address these issues, we propose a Balanced Simple attention Generative Adversarial Network (BSGAN), which enhances the quality and stability of the generated images by designing a new generator incorporating simple attention modules, a balanced learning rate technique, blur filters, and a dynamic padding method. At the same time, we introduce the perceptual loss to ensure detail retention and avoid over-adjustment. Experimental results demonstrate that BSGAN outperforms existing methods in style manipulation and image quality, providing a robust and flexible solution for facial attribute editing. Our codes and results are released on the project website https://gitee.com/renfh/bsgan.

Keywords:

Facial attribute editing

Generative Adversarial Network

Balanced Simple Attention GAN

Perceptual loss

Facial attribute editing via a Balanced Simple Attention Generative Adversarial Network

Abstract

Journal

Organization