Synthetic data has frequently been used for a range of computer vision tasks, such as object identification, scene comprehension, eye tracking, hand tracking, and complete body analysis. However, the development of full-face synthetics for face-related machine learning has been substantially hindered by the difficulty of modeling the human skull. Although realistic digital humans have been produced for films and video games, each character typically requires much artistic time. Because of this, the synthesis of facial training data in literature has been accompanied by simplifications or a focus on specific facial features, like the area around the eyes or the hockey mask.

Due to the disparity in distributions between actual and artificial face data, generalization is difficult due to the domain gap. Because of this, it is believed that synthetic data cannot wholly replace real data for jobs that must be performed in the field. Domain adaptation and adversarial domain training, where models are urged to ignore domain differences, have been the essential methods to close this domain gap.

Procedural sampling can generate and render innovative 3D faces at random without human assistance. The technology does this by overcoming a significant scaling limitation in the methods used by the Visual Effects (VFX) industry to synthesize lifelike individuals. By creating synthetic looks with unparalleled realism, Wood et al. aimed to directly address the issue by minimizing the domain gap at the source. Their approach procedurally blends a parametric 3D face model with a vast collection of excellently crafted elements from artists, such as textures, hair, and apparel.