High-Fidelity StyleGAN Inversion and Attribute Editing for Low Light Images

Ling Wu,Jun Zhang

DOI: https://doi.org/10.1109/eebda60612.2024.10486027

2024-01-01

Abstract:Recently, many studies have been explored the latent space of Generative Adversarial Networks (GANs), especially the W-space of StyleGAN. Real-world image manipulation has made great progress, and most of them follow the “invert first, edit later” methodology. These methods use well light images as input, but it is challenging for low light image captured in extremely weak illumination conditions. Therefore, applying the existing StyleGAN inversion techniques under such conditions leads to serious image distortion. In this study, an embedding module is proposed for image inversion and editing of dark faces in the real world. The module can be embedded in popular StyleGAN inversion techniques to enhance low-quality dark face images and provide high-quality input for subsequent inversion editing. It is called the SE module. The structure of the SE module is designed by combining the U-N et structure with the CBAM attention module. This has two benefits. First, it facilitates capturing complete information from the images. Second, the incorporation of the attention mechanism enables the network model to focus more on the crucial features of the images. This generalized embedding module effectively alleviates the shortcomings of dark face image inversion, achieve high-quality inversion of dark images. The experiments demonstrate the superiority of the module.