International Journal of Scientific Research and Engineering Development

The rapid flourishing of fashion e-commerce has greatly reshaped the dynamics of retail in the world today as fashion consumers are now able to shop a broad variety of apparel items in the comfort of their homes. Nonetheless, this change has brought a serious dilemma of high rates of product returns, which are mainly occasioned by the inability of customers to be able to visualize perfectly the appearance of garments on their own bodies before they make a purchase. Conventional visualization techniques of products involve images captured by professionals using models under controlled conditions, which are not reflective of the heterogeneity of the actual users of the products in terms of body shape, posture and environmental consideration. The current paper introduces a full-stack Virtual Try-On (VTON) technology, which combines both the Segment Anything Model (SAM) of Meta and IDM-vtolan diffusion-based synthesis to produce extremely realistic visualization of garments. The suggested system uses the zero-shot segmentation ability of SAM to create more accurate human body masks under a variety of input settings, without using human parsing models specific to a domain. The IDM-VTON diffusion model then takes these masks and generates conditional image synthesis on top of them using iterative denoising, with the model ensuring that images align garments correctly, retain high-fidelity texture, and adapt to lighting situations. The proposed diffusion-based pipeline has enhanced visual coherence and realism, as compared to traditional GANbased VTON systems, which tend to have blurring, geometric distortion, and limb occlusion artifacts in them. The system is deployed in a scalable web application with 3 tiers using React.js, Node.js, MongoDB Atlas, Supabase and HuggingFace APIs and can be deployed on consumer grade devices without the need to have dedicated GPU hardware. Experimental testing on the various combinations of portraits and garments indicates that the artifact rate can be less than 5% and the average inference latency is 1025 seconds, which is within reasonable range in commercial e-commerce practice. These findings confirm the usefulness of the suggested system in terms of the confidence of the users, the low rate of returns, and overall online shopping experience.

Shifa Bilal Tamboli, Simeen Phiroj Mulani, Arman Tajuddin Shiakh,"A Diffusion-Model-Based Virtual Try-On System Using Segment Anything for Realistic Garment Synthesis in E-Commerce" International Journal of Scientific Research and Engineering Development, V9(2): Page(2263-2269) Mar-Apr 2026. ISSN: 2581-7175. www.ijsred.com. Published by Scientific and Academic Research Publishing.