Collaborative Encoder for Accurate Inversion of Real Face Image

Abstract

In recent times, there has been notable progress in the effectiveness of Generative Adversarial Networks (GANs) for synthesizing images. Consequently, numerous studies have started utilizing GANs for image editing purposes. To enable editing of real images, it is crucial to embed a real image into the latent space of GANs. This involves obtaining the latent code of the real image and subsequently modifying the image by altering the latent code. However, accurately reconstructing the real image using the obtained latent code remains a challenge. This paper introduces a novel inversion scheme that achieves high accuracy. In contrast to conventional approaches that employ a single encoder for image inversion, our method utilizes collaborative encoders to accomplish the inversion task. Specifically, two encoders are employed to invert distinct regions in the image, namely the face and background regions. By distributing the inversion task between these encoders, the burden on a single encoder is reduced. Furthermore, to optimize efficiency, this paper adopts a lightweight network structure, resulting in faster inference speed. Experimental results demonstrate that our proposed method significantly enhances visual quality and improves the speed of inference. By leveraging collaborative encoders and a lightweight network structure, we achieve notable improvements in image inversion, thus enabling more effective image editing capabilities.