Artificial Intelligence has profoundly transformed the fashion industry, reshaping design, production, and consumer engagement. In particular, AI-driven techniques have revolutionized fashion recommendation systems and image generation, enhancing both creativity and personalization. This thesis explores two key methodologies within these domains. First, we introduce a GAN Based Fashion Outfit Generator for which we train a StyleGAN2-ADA model on the Lookbook dataset to generate realistic fashion images, achieving competitive performance. Using GANSpace, we identify and analyze latent space directions in an unsupervised manner, eliminating the need for labelled attributes and improving interpretability. The final model is deployed on Hugging Face Spaces with an interactive Gradio UI, enabling real-time exploration of generative fashion design.Second, we develop a LightGBM-based Fashion recommender system by integrating collaborative filtering and gradient boosting (LightGBM) to enhance personalization. Leveraging text processing, dimensionality reduction, and similarity-based techniques, we extract meaningful insights from the H&M Fashion and Kaggle Fashion Product datasets. We train two models: User-User Collaborative Filtering (UUCF) and LightGBM, evaluating their performance using Mean Average Precision @ 12 (MAP@12). Our results demonstrate that boosting techniques consistently outperform filtering-based methods, underscoring their effectiveness in improving recommendation accuracy. Our findings highlight the transformative potential of generative modelling for fashion synthesis and machine learning techniques for data-driven, personalized recommendations, paving the way for more intelligent AI applications in fashion retail.
We introduce a StyleGAN2-ADA model which is trained on a subset of the Lookbook dataset, consisting of 8,726 clothing images and achieved matching SOTA perfomance levels in most metrics. We then used the GANSpace method to identify the most significant directions in the latent space in an unsupervised manner while eliminating the need for an attribute classifier. These directions were then analysed and labeled based on their inferred representations. The final model was deployed on Hugging Face Spaces using a simple Gradio UI.
We develop a personalized fashion recommender system by integrating collaborative filtering and gradient boosting techniques. Our models are trained using the H&M Fashion Dataset and Kaggle’s Fashion Product datasets. To extract meaningful insights from product metadata and customer transactions, we employ text processing (TF-IDF), truncated SVD for dimensionality reduction, cosine similarity, and collaborative filtering. The performance of LightGBM and User-User Collaborative Filtering (UUCF) is evaluated using the Mean Average Precision @ 12 (MAP@12) metric. Our findings demonstrate that LightGBM consistently outperforms filtering-based approaches, highlighting its effectiveness in improving recommendation accuracy.
