Face inpainting based on GAN by facial prediction and fusion as guidance information

E2f-Net: Eyes-to-Face Inpainting Via Stylegan Latent Space

“…Generally, synthetic and natural masks are considered in face inpainting scenarios, which can be classified into two categories called free-(irregular) and fixed-form (regular) masks. In widely used free-form masks [23,24], there are irregular shapes randomly placed on the images (Fig. 2(a)(b)), useful for inpainting irregular scratches.…”

“…A novel attention layer in a coarse-to-fine design was suggested by Liu et al [30] in the same context. Wang et al [24] proposed a two-stage face inpainting method to detect the corrupted regions and then improve inpainting results using a top-down refinement network.…”

“…(a) [24] (b) [23] (c) [25] (d) [21] (e) [22] (f) Further, we employ a large-size mask that covers about 75% of the face image since the goal of this paper is to complete the face based on the region of the eyes, unlike other existing works. Despite using large-size masks, we do not use any guidance information during our training in our work as compared to other related works [6,17,18,19] to reduce training time and increase inference speed.…”