A Subspace-based Method for Facial Image Editing

Fig.2 The framework of the proposed method with two stages, i.e., adaptive weight modulation and slope truncation. The former aims to determine the parameters of subspace. The latter intends to derive an editing interval to perform safe and controllable editing. Facial image generation and editing are conducted with subspace parameters in Stage 2.

Fig.3 Performing real facial image editing with parameters of subspace. For each example, the user sequentially changes the pose, age, glasses, eyes closed, smile, and surprise of the real images. The raw image is embedded in StyleGAN2 latent space using DNI, which results in an inverted image.