This Thesis introduces several methods and a dataset, aiming to improve the capturing and inference of human faces, in a form that can be photorealistically rendered by computer applications. The proposed methods present advances in facial reflectance acquisition, from capturing to modeling and inference. In the first part of the Thesis, we present various improvements on a state-of-the-art facial capturing system, such as a Light Stage. More specifically, we work on an existing diffuse-specular separation method that does not require light polarization and further optimize it while making it applicable to multi-view capture. We also use this improved setup to capture the largest dataset of high-resolution facial reflectance to date, which we have opened to the research community. In the second part, we present two novel facial capturing setups, and an appropriate facial reflectance separation method. They are the first such systems consisting solely of off-the-shelf devices, while being modular and portable, and thus enabling capturing vast datasets of facial reflectance, at a reduced cost. Supporting static and dynamic facial capture, these systems allow the capture of even larger datasets of facial reflectance that can fuel deep-learning research and create race, gender, and age-balanced datasets. In the third part of this Thesis, we introduce the first method for high-resolution facial reflectance inference from a single arbitrary facial image. Our method AvatarMe and its extension, AvatarMe++, use a combination of 3D Morphable Models for facial shape reconstruction, adversarial image-translation models and super-resolution models, to transform reconstructed facial textures to high-resolution facial reflectance. We show that such a method is capable of retrieving UV maps of spatially varying facial reflectance that is render-ready in typical rendering applications without requiring capturing setups. Finally, we introduce the first deep learning-based 3D Morphable Model of facial reflectance, called FitMe. FitMe proposes a style-based Generative Adversarial Network, which is multi-modal and branched in both the discriminator and the generator, and is capable of concurrently generating high-quality facial diffuse albedo, specular albedo and normals. Moreover, by using an accurate diffuse-specular differentiable renderer, FitMe can retrieve state-of-the-art facial shape and reflectance with an iterative optimization.