EHA3D: Expressive Head Avatar via Disentangled Latent Code

Current NeRF-based head avatars typically use either explicit mesh representations based on head templates or implicit expressive coefficients derived from head templates as driving signals. Although these approaches have achieved promising results, they often depend on robust tracking systems, complex head template representations, long optimization times, and suffer from limited generalization abilities. In this work, we propose a novel method for driving expressive neural head avatar using disentangled latent code to address these issues. Our approach involves two key steps. We perform an EG3D inversion, where we disentangle the expression code from the base latent code, utilizing a supervised method based on contrastive learning and optical flow field supervision. To address differences in latent encoding distributions across different identity domains, we introduce the residual emotional enhancement in the subsequent stage to enhance similarity between driving and target frame expressions. With these designs, we successfully overcome tracking and complex head template issues. Moreover, the encoder-decoder structure of our pipeline demonstrates strong generalization performance, allowing for real-time and expressive reenactment across individuals of differing identities after training. Our method has achieved favorable visual results in qualitative experiments and demonstrates performance comparable to other state-of-the-art methods in quantitative metrics.