ReDirTrans: Latent-to-Latent Translation for Gaze and Head Redirection

Learning gaze estimation needs ample training data with precise annotations. To tackle this, some methods used image synthesis, but they were limited to changing gazes in small, eye-focused or low-res facial areas. To address this limitation, we introduced ReDirTrans, a network for translating latent vectors, enabling precise gaze and head orientation changes. By focusing only on desired attributes and employing subtraction-addition operations, it avoids interfering with other features in the image. Pairing ReDirTrans with a pretrained e4e-StyleGAN forms ReDirTrans-GAN, allowing accurate gaze redirection in high-res full-face images ($1024\times1024$) while preserving identity, expression, and hairstyle. Additionally, we enhanced gaze estimation by using redirected images as augmented dataset samples.