Practical Parametric Synthesis of Realistic Pseudo-Random Face Shapes

There is a growing demand for populating virtual worlds with large numbers of realistic-looking characters. Besides hand-crafted characters like the main protagonists in video games, the virtual worlds may also need massive numbers of secondary characters. Manual authoring of their features is not usually practical. For parametric models of human faces, a naive approach randomizes all the parameters of the human face to generate a random one. However, the uniform or hand-crafted distribution of the shape authoring parameters is unlikely to represent value ranges and correlations present naturally in human faces. The paper proposes a simple automated method for generating realistic-looking head shapes via learned mapping between latent space like the FaceNet embedding and the explicit parametric space used by the character modeling tools. Our approach is simple, robust, and can efficiently generate a large variety of head shapes with a predictable dissimilarity.