• Author(s) : Chong Bao, Yinda Zhang, Yuan Li, Xiyu Zhang, Bangbang Yang, Hujun Bao, Marc Pollefeys, Guofeng Zhang, Zhaopeng Cui

The digital avatar creation sector, especially in the area of animatable head avatars, has seen remarkable progress with the introduction of diverse volumetric representation techniques. However, a significant hurdle remains: the absence of a standardized method for performing complex editing tasks, like 3D head avatar modifications, across these varied representations. To bridge this gap, researchers have unveiled an innovative solution tailored for editing avatars created through different 3D Morphable Model (3DMM)-powered volumetric methods.

Central to this novel approach is the creation of an expression-aware generative model. This model is adept at transforming edits made on a 2D image into consistent 3D modifications, effectively bridging the gap between two-dimensional and three-dimensional editing. The success of this method is attributed to a combination of advanced techniques. It starts with an expression-dependent modification distillation scheme that draws on the vast knowledge pool of head avatar models to guide the editing process. The approach is further enriched with 2D facial texture editing tools and the application of implicit latent space guidance, ensuring the model’s effective convergence. Additionally, a segmentation-based loss reweighting strategy is introduced, focusing on achieving meticulous texture inversion for more detailed and precise edits.

The efficacy of this method has been rigorously tested through extensive experiments, which have confirmed its capability to deliver high-quality and consistent outcomes. Impressively, this method has shown exceptional adaptability, maintaining its effectiveness across a variety of expressions and viewpoints. This achievement represents a significant leap forward in the realm of 3D head avatar editing, offering a versatile and unified solution to a longstanding challenge in the industry.