Artificial intelligence in fabric design: a critical review of technological advancements and socio-creative implications (2019–2024)
摘要
Artificial Intelligence is revolutionizing the textile and fashion sectors. Machine learning and deep learning are fundamentally changing the fabric design process, which has always relied on intuitive yet laborious hand-crafted steps, such as surface patterns, material properties, and structural designs. This critical review systematically evaluates the state of the art in the application of Artificial Intelligence (AI) to fabric design from 2019 to 2024, through a comprehensive analysis of 65 peer-reviewed studies, and combines these key developments with a theoretical framework for human-AI co-creation. We critically compare the relative performance of several AI techniques, such as Generative Adversarial Networks (GAN), Variational Autoencoders (VAE), and diffusion models for aesthetics generation, as well as deep learning for material property prediction and structural design, using standardized performance metrics, including FID scores, accuracy rates, and computational efficiency measures. Among other components, this paper examines the changing dynamics of designer roles, skill transformations, and other social issues in the implementation and discussion of creative ownership, intellectual property, and algorithmic bias in automated systems. The co-creation of humans and AI reveals that AI tools boost human creativity by enabling space exploration design, predicting material behavior, enabling mass customization, and automating repetitive processes. Despite significant improvements, data quality challenges remain endemic, as do barriers to model interpretability and workflow integration. Our quantitative analysis of performance measures in the literature suggests that diffusion models achieve markedly superior aesthetic quality compared to GANs, with median FID scores of 23.4 (Range: 18.9–35.1) and 45.2 (Range: 32.4–58.0) for diffusion models and GANs, respectively. This comparison is based on studies that use similar texture/pattern datasets (e.g., DTD, Fabric-1000) and should be interpreted as relative performance trends, not as a universal means, due to differences in underlying model architectures. The accuracy of deep neural networks is higher (R2 > 0.85) in material property prediction tasks than that of GANs. These capacities can lead to reported development cost reductions (e.g., 40–60% in selected case studies), increased collection success rates (25–40% in predictive systems), and reduced over production (30–50% in personalized design models) (Júnior et al.