CT: A Controllable Transformer for Multi-Task TCM Facial Inspection

Traditional Chinese Medicine (TCM) facial inspection is a non-invasive, holistic approach to health assessment, but its reliance on practitioners' subjective judgment limits standardization and scalability. Applying artificial intelligence (AI) to this domain faces unique challenges: the spatial heterogeneity of diagnostic regions, complex inter-task correlations, and the difficulty of integrating domain priors. To address these challenges, we propose CT (Control Transformer), a hierarchical multi-task vision transformer that (1) injects task-specific spatial priors via a ControlNet-based spatial control module using "Mingtang" partition masks; (2) employs a Mixture-of-Experts (MoE) backbone to encourage functional specialization and flexible knowledge sharing; and (3) introduces a knowledge-graph-infused Feature- wise Linear Modulation (FiLM) module that converts TCM syndrome element theory into a task correlation graph to drive theory-guided, task-adaptive feature modulation. Evaluated on a large, self-constructed TCM Facial Diagnosis (TCM-FD) dataset, CT substantially outperforms strong baselines. Ablation studies and qualitative analyses further validate the effectiveness and interpretability of each component, demonstrating a practical paradigm for AI-assisted TCM diagnostics.