StyleEditorGAN: Transformer-Based Image Inversion and Realistic Facial Editing

Recent advances in StyleGAN research have shown significant progress in semantic disentanglement. While controlled editing of facial images has been achieved to some extent, attaining highly controllable attribute editing remains a challenging task. Existing GAN inversion methods require a trade-off between reconstruction fidelity and editing quality. To further investigate facial attribute editing, dual-space GANs have been employed to better disentangle style and content representations. However, these methods still struggle to achieve precise, highly controllable editing results for complex attributes. In response, this paper introduces a novel model, StyleEditorGAN, designed to enable more accurate and intricate attribute control. The proposed model presents an innovative dual-space editing and GAN inversion approach, emphasizing the significance of interaction within dual-space GANs, and leverages a Transformer framework to enhance this interaction for more precise and controllable editing. Extensive experiments on the CelebA-HQ and FFHQ datasets demonstrate the superiority of the proposed StyleEditorGAN model in image reconstruction quality, editing capability, and flexible style transfer, underscoring its effectiveness in highly controllable facial editing.