A virtual clothes try on application, to aid shopping.

It's a project we developed during a Gen-AI hackathon organised by Paras Madan and Nas.io, where we developed an e-commerce wesite, which has an virtual try on feature. That allows user to upload their own photos and try out the clothes for an accurate depiction. It's mostly targeted for B2B, where this feature can be implemented by smaller retailers, which increase their user interaction and brings a wow factor. The main aim for this was to bridge the gap between techology and non-tech people.

• GithubGithub (might be private)
• ArticleFeatured on Paras Madan's medium article
• StackNext.js 14 / JavaScript / Python / PyTorch / CSS
• VideoDemo video
• Huggingface spaceSpace (might be private, if we are developing this further)

Ideation

This project was built during the Gen-AI hackathon organised by Paras Madan and NAS.io. The main task was to build something revolving around the theme "For India from people of India", where we planned to make an application that can be used by smaller retailers to showcase their clothing line ups without having to spend big chunks of money on professional photoshoot's. Along with this we build our own recommendation model. The whole process is called Inpainting with Image Guidance.

Building

We decided to use Next.js to build the e-commerce website for the demo purposes. For the main stable difussion model we used Automatic1111, as the base model and trained it on our own dataset. We also used PyTorch for the training and the model was deployed on huggingface spaces, using the pro subscription. We used gradio for the interface. We built the recommendation model with the combination of colaborative filtering and content based filtering. We used ControlNet for the pose preservation and CodeFormer for refining the final images. The user is given 3 options upper body, lower body and dresses to choose from.

Learning

How stable diffusion works and how to incorporate open source models for our own applications, along with this dealing with problems like image segmentation, hallucinations while getting the resultant images, and working on the scalability of the product.

Results