LangFabric: A Language-Annotated Multimodal Fabric Dataset for Fabric Defect Detection

Fabric defect detection plays a critical role in intelligent textile manufacturing, where reliable identification of surface anomalies is essential for maintaining product quality and production efficiency. Although numerous fabric defect datasets have been proposed, existing resources primarily focus on visual modalities and lack structured semantic descriptions of defects, limiting their applicability to emerging vision–language learning paradigms. To address this limitation, we introduce LangFabric, the first language-annotated multimodal fabric dataset designed for fabric defect detection in real textile production environments. LangFabric consists of 6,384 paired samples, each comprising a high-resolution fabric image and a corresponding structured textual description. The dataset covers 16 distinct fabric texture categories and 12 defect types, including both structural and non-structural anomalies. In addition to pixel-level semantic segmentation masks for precise defect localization, each image is annotated with fine-grained textual prompts describing defect category, structural attributes, spatial location, and relative size. The data are collected inline from an industrial air-jet loom under realistic manufacturing conditions, ensuring practical variability in texture patterns and illumination. By systematically integrating visual and linguistic modalities, LangFabric provides a comprehensive benchmark for multimodal fabric defect analysis, supporting supervised detection, anomaly detection, one-class learning, and vision–language alignment research. We anticipate that LangFabric will facilitate the development of more robust, interpretable, and semantically-aware industrial inspection systems.