Abstract
Abstract In the Fashion and Apparel (F&A) industry, traditional sizing systems do not align with anthropometric diversity, resulting in frequent misfit. To meet the growing demand for customized apparels fit, deep learning and attention mechanisms are increasingly applied. Existing systems are based on Artificial Weak Intelligence (AWI), which employs complex feature learning, rigid sizing system (S/M/L) and fail to adopt individual style preferences that limit generalization. This study introduces an Advanced Narrow AI approach, selecting key features that uphold model performance while promoting scalability and transparency. The proposed Deep feed forward Neural Network (DNN) utilizes mesh-derived anthropometric measurements to enable size-aware learning and multi-output style-oriented garment prediction with improved accuracy and generalization. This bi-dimensional architecture promotes customized tailoring and enhancing computational efficiency beyond AWI’s style-adaptive sizing capabilities from standard wear to bespoke tailoring. In parallel, a Transformer TabNet (TF-TabNet) operates in a single stream for interpretability and LLM-aligned processing, but obtain less precise refinements for restricted feature network. The proposed Deep feedforward Neural Network achieved (98–99%) R² scores for three distinct shirt variants, with MSEs of (0.14–0.26) inches, MAEs of (0.28–0.41) inches, and 100% size-aware classification performance across all metrics. Experimental results achieves optimum accuracy and training stability enabling task-aligned formulation. The multi-output regression stream is guided by the size-aware stream, which helps focus and generalizes. On the other hand, TF-TabNet underperforms because its sequential feature masking assumes selective relevance, which suppress jointly informative anthropometric features that are consistently important for garment length prediction. The results shows that the decomposed modeling and targeted feature selection can improve performance and help achieve task specific learning, maximizing computational efficiency, interpretability and alignment with Responsible AI goals.