BlendFields: Few-Shot Example-Driven Facial Modeling

Ours Rendered training expressions Novel expression Figure 1 . Teaser -Given five multi-view frames of different expressions, our approach generates a model capable of capturing the fine-grained details of a novel expression beyond the resolution of the underlying face model [15] (top right corner). This is achieved by blending the radiance fields computed for individual expressions, where the blending coefficients are modulated accordingly to local volumetric changes. These volumetric changes are measured as the difference in the tetrahedral volume of a mesh that deforms with the expression ( increase, decrease, and no change in volume). Such an approach allows BlendFields to render sharp, expression-dependent details of the face without increasing the resolution of the mesh (bottom right corner).