This is demo repo for our BMVC 2022 paper:
Fill in Fabrics: Body-Aware Self-Supervised Inpainting for Image-Based Virtual Try-On
Training details available in the official repo fifa-tryon.
Figure 1. Final try-on outputs of our method with other recent try-on methods.
Interactive app build using Gradio
You can use the model as a simple UI made with gradio. See gradio_app for details on how to run the app. This app currently works on a local machine with a GPU. Can be hosted on a GPU server.
Note: I attempted to do a CPU implementation first. Got running the try-on and pose estimator models on CPU. The issue is when getting the parsing results (i.e multi-class segmentation) using this for the person image. The pre-trained model uses In-Place Activated BatchNorm for memory-optimized training. The implementation of In-Place Activated BatchNorm is currently only for GPUs.
This inference codebase is modified from https://github.com/levindabhi/ACGPN to run custom models. The human parser and segmentation models are from https://github.com/hasibzunair/Self-Correction-Human-Parsing-for-ACGPN and https://github.com/hasibzunair/U-2-Net.