#mpi_infmpgde #vcaimpi_infmpgde
In the paper presented at SIGGRAPH ASIA 2023, “**AvatarStudio: Text-driven Editing of 3D Dynamic Human Head Avatars**,” Christian Theobalt and colleagues introduce AvatarStudio, a pioneering technology enabling the text-driven editing of the appearance of dynamic full head avatars, enhancing the application in fields like extended reality and media production. This method, based on existing work in capturing dynamic human head performances using a neural radiance field (NeRF), incorporates a text-to-image diffusion model to edit this representation. It introduces an innovative optimization strategy that integrates multiple keyframes from different camera viewpoints and timestamps into a single diffusion model. This personalized diffusion model subsequently edits the dynamic NeRF using view-and-time-aware Score Distillation Sampling (VT-SDS) following a model-based guidance approach. The method ensures that the edits are genuine, personalized, and maintain 3D and time consistency, as validated by user studies and visual and numerical evaluations, thereby outperforming existing approaches.