Facial quality assessment of digital humans: A dual-branch framework integrating morphological harmony and expressive coordination

With the rapid advancement of metaverse technologies, digital humans (DH), as core interactive entities in virtual-physical integrated ecosystems, face unique challenges in their quality assessment frameworks. Existing research predominantly focuses on quantifying natural image distortions but fails to address DH-specific issues such as facial morphological disharmony and expression incoherence. To bridge this gap, we propose a dual-branch quality assessment framework for digital humans: (1) Leveraging medical aesthetic priors, we construct structural features based on facial aesthetic subunits and model temporal dependencies using gated recurrent units, combined with a loss-averse pooling strategy to capture transient severe distortions. (2) We quantify expression coordination through multi-dimensional Action Unit (AU) topology graphs, proposing triple-edge definitions and regressing dynamic distortion levels via graph convolutional networks. Experiments on the multiple THQA datasets demonstrate that our framework significantly outperforms conventional methods in subjective mean opinion score consistency, with the dynamic branch playing a dominant role in performance optimization. This work establishes a quantifiable evaluation standard for DH modeling refinement and real-time rendering.