Reference-based Face Editing with Spherical Harmonic Illumination and Geometry Improvement
Keqian Chen, Yunwen Zhu, Tingyan Gu, Yonghua Zhu · 2022
Face editing is a topic in computer vision. To accurately meet the demands of users, many researchers focus on the reference-based face editing task which aims to manipulate an original image from reference attributes. However, existing methods tend to have illumination information mistakes and geometry structure missing issues. In this paper, we present an illumination and geometry-improved framework. The framework consists of two vital parts: the illumination adjustment module and the geometry improvement part. The illumination adjustment module is introduced to predict spherical harmonic lighting code concatenating with the original style code to combine more illumination details. And the geometry improvement part serves as an extra disentangled geometry feature provider and a combiner to achieve a fine-grained blended image. Extensive results demonstrate that our method has significant improvements in illumination and geometry preservation.