We propose to guide anatomical or ‘patient-generic’ (PG) liver modelling by means of 14 anatomical surface point correspondences. We have tested a statistical model built from 71 patient meshes and a generic kernel-based model built from the mean mesh. The latter obtains lower global and local surface reconstruction errors, below 6 mm, while internal structure errors are 11 mm on average. For Mini-Invasive Liver Surgery (MILS) guidance through Augmented Reality (AR), the PG model is registered to an intraoperative 2D image through an adaptation of a patient-specific (PS) 3D/2D registration neural network framework. While this application is slightly less accurate regarding tumour Target Registration Error (TRE) than existing PS methods in in-vivo surgery data, it requires a single training for all patients. It thus eases clinical application and paves the way to anatomical AR through generic anatomical features from the PG model only.