A pipeline for reconstructing faces from partial scans

The Adaptive Neural Facial Deformer tackles a practical 3D vision problem: a scan often captures only part of a face, with noise, occlusion, and an unknown pose. A useful output needs to be more than a static mesh. It should preserve identity while allowing expression and pose to be edited.

The pipeline starts with a point cloud, estimates facial parameters with a learned solver, then uses non-rigid ICP to refine the alignment. The resulting representation can be exported as an editable OpenUSD rig. That final step matters because the reconstruction can be inspected and changed in a downstream 3D workflow.

The project separates the learned estimate from geometric refinement so that their contributions can be evaluated. The evaluation notes compare learned, classical, and hybrid approaches; the repository documents the implementation and its limitations.