PROMETHEUS-VTON: Precise Rendering Of Mixed Environments and Textiles for High-fidelity Ensemble Unification System in Virtual Try-On
Virtual try-on (VTON) technology has the potential to revolutionize online shopping experiences, but existing approaches face challenges in achieving photorealism and adapting to diverse clothing styles. This paper introduces PROMETHEUS-VTON, an enhanced virtual try-on system that builds upon the IDM-VTON framework to address these limitations. Our key contribution is the fine-tuning of the UNet2DConditionModel architecture to improve performance on complex garments and poses. We curated a dataset of 60,000 high-resolution images, including traditional Pakistani clothing, to address the lack of diversity in existing datasets. Through comprehensive experiments, we demonstrate significant improvements over state-of-the-art methods, achieving a 15.7% reduction in LPIPS score, a 2.5% in-crease in SSIM, and a 4.4% improvement in CLIP Image Similarity score compared to our baseline model. PROMETHEUS-VTON shows particular strength in handling non-Western garments and complex poses, representing a significant step towards making virtual try-on technology more robust and versatile.
