ABSTRACT High‐fidelity 3D facial retargeting requires transferring a source expression to a target face while preserving identity and geometric plausibility. Most existing methods rely on explicit parametric face coefficients or carefully paired cross‐identity data. However, explicit parameterization limits expressiveness and fine geometric detail, while reliable paired cross‐identity supervision is expensive and often fundamentally unavailable, since the same expression is rarely synchronized with matching geometry and intensity across different identities. We propose GeoReFace, a geometrically consistent framework for cross‐identity 3D facial retargeting that avoids explicit coefficient manipulation and instead recombines identity and expression in a learned mesh latent space. Our key component, dual‐domain adaptive normalization (DDAN), standardizes and re‐modulates latent features to inject source expression dynamics into target identity features. Combined with a dual‐domain retargeting objective and a geometry‐aware consistency constraint, DDAN enables identity‐preserving expression transfer without paired supervision. GeoReFace further uses spectral mesh convolutions with residual learning and a Transformer‐based fusion module to model vertex‐level geometric dependencies. Experiments on registered and raw‐scan datasets show competitive retargeting quality, strong geometric consistency, and improved reconstruction accuracy over prior mesh autoencoders.
Feng et al. (2026) studied this question.