FaceParts: Unsupervised 3D Facial Editing via Gaussian Splatting
FaceParts serves as a framework designed for the unsupervised segmentation and editing of Gaussian Splatting avatars, allowing for accurate facial modifications and the swapping of features between avatars without the need for manual adjustments. In contrast to traditional 2D or mesh-based techniques, this method functions directly within the Gaussian domain, breaking down avatars into semantically meaningful components like beards, eyebrows, eyes, and mustaches. It employs a combination of feature disentanglement, density-based clustering, and FLAME-anchored part transfer. Testing on the NeRSemble dataset, which includes 11 subjects, shows effective feature isolation, with transferred segments successfully adapting to various poses and expressions. This technique holds potential in fields such as entertainment, digital avatars, and virtual reality.
Key facts
- FaceParts is a framework for unsupervised segmentation and editing of Gaussian Splatting avatars.
- It operates directly in the Gaussian domain without supervision.
- The method integrates feature disentanglement, density-based clustering, and FLAME-anchored part transfer.
- Enables precise editing and cross-avatar part swapping.
- Experiments on the NeRSemble dataset with 11 subjects.
- Robust isolation of features such as beards, eyebrows, eyes, and mustaches.
- Transferred segments adapt to pose and expression.
- Applications in entertainment, virtual reality, and digital avatars.
Entities
—