High quality GAN encoded using Transformer
Pengsen Zhao, Qiner Wu, Xiaoqiang Jin · 2023
StyleGAN inversion plays a crucial role in enabling the use of pre-trained StyleGAN for real image editing tasks. The goal of StyleGAN inversion is to find the exact latent encoding of a given real image within the latent space of StyleGAN. Existing optimization-based methods can produce high-quality results, but optimization often requires a significant amount of time. On the other hand, encoder-based methods are usually faster but suffer from lower result quality. In this paper, to improve the quality of results from encoder-based methods, we propose the following two enhancements: 1) employing a Transformer to encode image features, and 2) integrating the encoding of fine-grained details. By combining these design choices, we present an effective approach that not only achieves high-quality inversion results with encoder-based StyleGAN inversion but also effectively mitigates issues related to latent vector entanglement.