Transformer-based high-fidelity StyleGAN inversion for face image editing
Chutian Yang, Xiping He, Qixian Kuang, Ling Huang, Lingling Tao · 2023
Recently, there have been many methods utilizing pre-trained GAN generators for image editing. To apply these methods to the editing of real images, it is necessary to invert real images into the latent space before proceeding with the editing process.Existing GAN inversion methods have difficulty providing both high-fidelity inversion and high editability for face images. The most common inversion method based on the W+ space has enough ability to describe image details, but sacrifices editability. To address this problem, this paper proposes a Transformer-based StyleGAN inversion model to achieve high-quality editing of real face images. This paper uses the code of the StyleGAN2 mapping network as the initial query vector and the image features extracted by a CNN encoder as the key and value. Multiple updates to the query vector are made using a multi-head cross-attention module, and the final output query vector is used as the latent code of the inverted image and fed into the generator. The experimental results prove that our method can provide high-fidelity inversion for face images, and the edited images can effectively preserve identity information while achieving precise control over attributes.