CASE STUDY IN GENERATIVE ADVERSARIAL NETWORKS FOR TEXTILE PATTERNS GENERATION

This study focuses on the implementation and evaluation of generative models for the generation of textile designs using Generative Adversarial Networks (GANs). The approach involved developing both unconditional and conditional versions of Wasserstein GANs (WGANs) and Wasserstein GANs with Gradient Penalty (WGAN-GP), as well as adaptations for higher resolution outputs. A diverse dataset of 13,000 textile patterns was compiled, and the models were trained on this data, with architectures designed to optimize image generation in terms of both resolution and feature learning. The training process was analyzed using loss stability assessments, visual evaluation, and accuracy metrics. Results showed that WGAN-GP models demonstrated greater loss stabilization but lower overall accuracy since the discriminator learned faster, while conditional models showed improvement in image fidelity but with some divergence issues during training. Additionally, efforts to upscale output resolution to 256x256 pixels were largely unsuccessful, with significant loss oscillations and poor constructed generated samples. This study concludes with recommendations for further refinement of the model architectures and training strategies to improve the generation of high-quality, high-resolution textile designs.