“…With the above evolution of 3D hair representations, the quality of recovered shape has been improved significantly. As populating pairs of 3D hair and real images is challenging [46], existing learning-based methods [26,30,37,40,46] are just trained on synthetic data before applying on real portraits. However, the domain gap between rendered images (from synthetic hair models) and real images has a great and negative impact on the quality of reconstructed results.…”