photometric_optimization

Why we included this project

Photometric fitting of a textured FLAME head model to a single in-the-wild photo is the core of this repository, and it covers the whole analysis-by-synthesis loop rather than just one piece. The code optimizes FLAME's shape, pose, and expression parameters along with spherical-harmonics lighting and a texture offset, using landmark and photometric losses restricted to the skin region so occlusion does not wreck the fit. It also documents how to build a PCA texture space from roughly 1,500 FFHQ images, including a texture-completion step with an inpainting network to fill occluded areas. That second part matters because FLAME itself has no appearance space, so this doubles as a recipe for constructing one from scratch. The single-image demo runs once PyTorch3D is set up, though you will need to register on the FLAME site to download the model and texture assets.