Treffer: Towards practical AI: deploying virtual try-on models as an interactive application

This project aims to bridge the gap between complex artificial intelligence workflows and end users by developing a user-friendly virtual try-on system. We leverage ComfyUI, a node-based system designed for building and managing artificial intelligence workflows, to run IDM-VTON inference to create virtual try-on images. Unlike previous GAN-based try-on methods that face difficulties in preserving garment fidelity, IDM-VTON is a diffusion-based approach that encodes both the high-level semantics and the low-level features of the garment image to improve accuracy. The novelty of this project lies in the creation of a reproducible process to convert a technical, node-based workflow into a more user-friendly application that reduces the barrier of entry so as to increase the outreach of such technologies. This is demonstrated by our application which uses IDM-VTON, a state-of-the-art try-on model, as our inference model. In this project, we set up a FastAPI server to act as a backend-for-frontend (BFF) service between ComfyUI and our frontend application to transform the data into the required formats between the frontend and ComfyUI. This layer of abstraction allows us to create a user-friendly workflow without modifying the underlying logic in the ComfyUI server. Despite challenges such as the lack of built-in parameter validation in ComfyUI and workflow specificity, this work provides an approach to develop future applications for other workflows, thereby increasing outreach of such technologies. ; Bachelor's degree