GPAvatar: High-fidelity Head Avatars by Learning Efficient Gaussian Projections

The bubble size indicates memory usage. Quality, Speed and Memory Usage Figure 1. Left and Middle: A monocular portrait video of an individual is used to reconstruct a 4D avatar, facilitating novel-view synthesis and facial reenactment. Right: Our method is compared with state-of-the-art approaches in terms of visual quality, rendering speed and memory usage, showcasing superior rendering quality coupled with computational efficiency and minimal memory usage.