Reconstructing Personalized Semantic Facial NeRF Models from Monocular Video

We present a novel semantic model for human head defined with neural radiance field. The 3D-consistent consist of set disentangled and interpretable bases, can be driven by low-dimensional expression coefficients. Thanks to the powerful representation ability field, constructed represent complex facial attributes including hair, wearings, which not represented traditional mesh blendshape. To construct personalized model, we propose define bases as several multi-level voxel fields. With short monocular RGB video input, our method subject's NeRF only ten twenty minutes, render photorealistic image in tens miliseconds given coefficient view direction. this representation, apply it many tasks like retargeting editing. Experimental results demonstrate its strong training/inference speed. Demo videos released code are provided project page: https://ustc3dv.github.io/NeRFBlendShape/