<p>The preservation of traditional costume cultural heritage faces challenges such as the fragility of physical artifacts and the interruption of knowledge transmission. Current computational methods primarily focus on surface-level semantic understanding, such as the instance-level style classification and extraction which often relies on manual annotation, and it also overlooks the fine-grained meaningful elements within the instance. To address these challenges, this paper take modern Chinese traditional costumes as example and proposes a Disentangled Orthogonal Generative Adversarial Network (termed as DO-GAN) to decompose costume image into content and style for fine-grained style extraction and regeneration. By introducing orthogonal constraints, it minimizes stylistic codebook redundancy in an unsupervised way, ensuring that each codeword corresponds to a unique, culturally significant style primitive (e.g., specific motifs like intertwined branches or wavy water patterns). Experimental results show that the DO-GAN achieves superior semantic distinctness in style representation, enabling accurate and interpretable digital archiving of discrete heritage elements. While marginally trailing Vanilla VQGAN in pixel-level metrics, it outperforms all compared methods in perceptual authenticity and maintains strong structural consistency. Notably, our method supports effective style transfer without extra transfer training, realizing stable and authentic migration of traditional motifs (e.g., intertwined branch pattern, wavy water pattern) onto contemporary costumes, which bridges static heritage archiving with dynamic creative application. Meanwhile, the extracted style codebooks can serve as the “genes" of costumes to build a database for broader and more durable analysis.</p>