Generating Animatable 3D Face Avatars from a Single Image

FaceCraft4D is a framework that creates a high-quality animatable 4D avatar from a single image. It targets the limitations of existing methods that either depend on multiple multiview data or struggle with shape accuracy and identity consistency.

The pipeline combines three priors. First, 3D-GAN inversion obtains an initial coarse shape, and then an image diffusion model enhances multiview texture using depth-guided warping signals, improving consistency across viewpoints.

In the expression animation stage, a video prior and viewpoint-synchronized driving signals are used so that the same motion is maintained across multiple viewpoints. Consistent-Inconsistent training is also introduced to effectively handle data inconsistencies that arise during 4D reconstruction, and experiments show better quality and viewpoint/expression consistency than existing methods.