High-fidelity Facial Avatar Reconstruction from Monocular Video with Generative Priors

Figure 1 . Visualizations of 3DMM and audio-driven face reenactment of our proposed method and NerFACE [11] and DFRF [41] . The leftmost column is the ground truth image. For each method, the left plot is the rendered image with the same view of the ground truth image, and the two right plots are novel view syntheses. By utilizing the high-quality 3D-aware generative prior, our method significantly boosts the performance of face reenactment and novel view synthesis. We highlight some areas with red rectangles for better comparisons.