A human-in-the-loop framework for generative batik design: integrating computer vision, cultural constraints, and IPR metadata

Abstract

Generative artificial intelligence (AI) can rapidly produce motifs and designs at scale, but its application to Indonesian batik requires consideration of cultural meaning and artisans’ intellectual property rights (IPR). This study presents a hybrid, human-in-the-loop AI-native framework that integrates computer vision, artisan-provided cultural and IPR permission metadata, constraint-based filtering, and external multimodal generative models. A proof-of-concept evaluation was conducted using 447 images representing nine Lumbini batik motif categories, with 357 images (80%) used for training and 90 images (20%) for validation. ResNet-50 achieved 100% accuracy, precision, recall, and F1-score on the validation dataset. The framework subsequently classified motif inputs into Generation Allowed, Generation Prohibited, and Uncertain Pending Manual Verification based on a 0.70 confidence threshold and permission metadata. Generated designs for permitted motifs were further examined through human-in-the-loop cultural validation by Lumbini batik artisans and garment mockup assessment. The findings demonstrate the feasibility of integrating artisan-defined cultural constraints into a generative batik design workflow. However, the results are limited to the Lumbini motif collection and validation dataset, as no separate held-out test set or external validation across other batik collections was conducted.