Paper Title
GENERATIVE MODELING FOR AFRICAN AND INDIAN FASHION: ADAPTING AND EVALUATING STYLEGAN AND STABLE DIFFUSION APPROACHES

Abstract
In fashion applications, models are predominantly trained on Western datasets, which limits their ability to generalize to non-Western domains. The scarcity of non-Western datasets further constrains model adaptation and evaluation. This paper proposes a systematic framework for adapting and analyzing generative models for textile pattern and apparel generation in non-Western and mixed-domain settings. We present two experimental studies. First, StyleGAN2-ADA and Stable Diffusion are trained on a curated synthetic dataset of 2,000 high-resolution African wax textile images, and distributional alignment is evaluated using Fréchet Inception Distance (FID). Second, Stable Diffusion is fine-tuned using Low-Rank Adaptation (LoRA) on a dataset of 2,000 Indian saree images. We apply prompt embedding interpolation and perform a structured parameter sweep across LoRA scale, blend ratio, prompt scaling, classifier-free guidance, and inference steps to analyze their impact on generative behavior and output quality. Evaluating diffusion-based outputs remains challenging, as existing metrics capture only partial aspects of visual fidelity and semantic alignment. We therefore combine CLIP-T, CLIP-I, DINO, and Intra-LPIPS to characterize performance and output diversity across parameter regimes. Our results provide a reproducible methodology for generative fashion model adaptation and expose limitations in current evaluation frameworks for diffusion models.